The Field Company

Field Notes: A Week Testing Data Collection Apps in the Cederberg

Sat, 05 Dec 2026 00:00:00 GMT

import { Image } from "astro:assets"; import BlogImage from "../../components/blog/BlogImage.astro"; import FieldNotes2 from "../../assets/blog/field-notes-cederberg/1465905_nur_andi_ravsanjani_gusma.jpg"; import FieldNotes1 from "../../assets/blog/field-notes-cederberg/10828186_nikita_igonkin.jpg"; import cederbergRidge from "../../assets/blog/field-notes-cederberg/5322829_magda_ehlers.jpg"; import cederbergDawn from "../../assets/blog/field-notes-cederberg/5322828_magda_ehlers.jpg";

Dawn broke over the sandstone cliffs at 05:47. I was already up, hunched over a gas stove brewing coffee in the thin mountain cold. On the camping table in front of me: four Android phones, each running a different data collection app. FieldLog, CyberTracker, ODK Collect, and SMART Mobile. The phone running SMART had 10% battery and no signal. This was going to be interesting.

The plan was simple enough. Spend a week in the Cederberg Wilderness Area — 71,000 hectares of sandstone, fynbos, and UNESCO World Heritage landscape about 250km north of Cape Town — and test these four apps against each other in real field conditions. Not simulated conditions. Not "airplane mode on a bench in the Company Gardens." Real conditions: no signal in the valleys, patchy 3G on the ridges, 38°C heat on the exposed slopes, dust in every charging port, and a troop of chacma baboons that took a professional interest in our breakfast supplies.

I want to be clear about what this is and what it is not. This is not a product review. It is not a spec sheet comparison. It is a field report — an honest account of what happened when we took four tools into a place that was specifically chosen to break them. We build FieldLog. We have skin in this game. But you need to know what broke, what held up, and what we changed because of this trip. If we cannot be honest about where our own tool fell short, we have no business building tools for people whose data actually matters.

The Setup

Four identical Tecno Spark Go devices — 1GB RAM, 8GB storage, Android Go edition. These are the phones people actually carry in the field, not the Pixel 8 Pro I use at my desk. Each phone had exactly one app installed: FieldLog, CyberTracker, ODK Collect, or SMART Mobile. Each was configured with a matching observation form: species, count, age class, sex, habitat type, GPS coordinates, photo. The same form across all four apps. The same transect routes. The same observer. The only variable was the software.

We stayed at Sanddrif, the holiday resort on the farm Dwarsrivier, home of Cederberg Private Cellar — the highest-altitude winery in South Africa. Sanddrif sits at the base of the Wolfberg, the massive sandstone massif whose cracked summit and freestanding arch define the Cederberg skyline. The campground has cold showers, unreliable electricity, and a dam cold enough to stop your heart. It also has a wine tasting room 400 metres from your tent. This last fact became increasingly relevant as the week progressed.

Day One: Arrival and First Transect

We arrived mid-morning, the N7 giving way to 17km of gravel road that rattled the Hilux's suspension and coated everything in a fine red dust. The Cederberg in November is a study in contrasts: the fynbos is in bloom — proteas like pink fireworks, ericas in white and yellow, restios waving in the wind — but the heat is already building. By 11:00 it was 31°C and climbing.

First task: configure the observation forms. CyberTracker required the desktop Windows application to build the sequence — a hierarchical menu of icons that the field worker taps through. The interface looks like Windows 98 and the learning curve is steep, but the icon paradigm is genuinely brilliant for non-literate users. I built the form in about 40 minutes, swearing at the dropdown menus approximately 17 times.

ODK Collect uses XLSForm — you build the form in Excel, upload it to an ODK Central server, then download it to the phone. Conceptually clean. In practice, our ODK Central instance was running on a laptop in a tent with no internet, so the phone downloaded the form over a local Wi-Fi hotspot running off a power bank that was slowly dying. It worked, but "cloud-first form deployment" feels absurd when the cloud is a 2015 ThinkPad under a groundsheet.

SMART Mobile was the heaviest setup. Full desktop application required — you build a patrol configuration with spatial data, patrol routes, observation categories, and enforcement modules. The configuration process took two hours and generated a 47MB data package that had to be transferred to the phone via USB. This is a law enforcement tool pretending to be a data collection tool, and the setup weight reflects that. If you are managing a protected area with formal ranger patrols, the weight is justified. If you are doing ecological monitoring, it feels like bringing a fire engine to collect a water sample.

FieldLog: built the form in four minutes on my phone. Tapped "New Expedition," described the observation protocol in plain English to the AI form builder, reviewed the generated fields, adjusted two things, deployed. I am aware this sounds like marketing. It is also what happened. The other three apps were still being configured when I had already logged my first test observation. This is not a flex — it is an acknowledgment that setup speed matters when your field team is standing around in the sun waiting for you to get the tools working.

We ran the first transect at 14:00, heading up the trail toward the Maltese Cross — a 5-hour round trip through classic Cederberg terrain. Sandstone steps, fynbos corridors, the occasional klipspringer materializing on a boulder and then vanishing before you can raise your binoculars. I logged observations on all four phones simultaneously: same species, same coordinates, same time. Each observation required me to stop, unlock the phone, navigate to the form, enter the data, and continue walking. This sounds trivial. After the 14th observation, it was not.

CyberTracker was the fastest in the field — genuinely fast. The icon grid means you tap through a sequence without reading. Tap the baboon icon, tap adult, tap male, tap healthy, confirm. Three seconds. The trade-off is flexibility: if the species you are logging is not in the pre-built icon sequence, you cannot log it without rebuilding the form on a Windows desktop. CyberTracker's icon interface worked brilliantly until I had to log 14 observations of the same dassie species in three minutes. The repetition was fast but mind-numbing — no batch entry, no "same as last" shortcut, no count increment. Fourteen identical tap sequences, one after another, while the rest of the group walked ahead.

ODK Collect was precise but slow. The form-based interface is clear and the validation is excellent — it will not let you submit an incomplete record. But every field requires a deliberate tap or swipe. Dropdowns for species with 40+ options are tedious. After 20 observations my thumb was sore and my pace had dropped to roughly half the group's walking speed. ODK is a form engine, not a field tool, and you feel the difference in every interaction.

SMART Mobile was the slowest by a significant margin. The observation form required navigating through nested patrol menus. Each observation had mandatory fields that made sense for law enforcement — patrol sector, threat level, action taken — but were irrelevant for ecological monitoring. I found myself entering "N/A" into required fields just to submit a dassie sighting. SMART is not bad software. It is just not designed for this job.

FieldLog was fast but I noticed something I had not noticed in testing: on this particular phone, with the screen brightness cranked to max to compete with direct sunlight, the GPS lock took noticeably longer than CyberTracker's. CyberTracker uses a lightweight GPS call that grabs the last known location and updates it in the background. FieldLog was requesting a fresh fix on every observation. This was correct behaviour — fresh fixes are more accurate — but in practice it meant a 2-3 second delay per observation that CyberTracker did not have. I made a note. This would become important.

We returned to camp at dusk, sunburned and dusty. A pair of Verreaux's eagles had been riding the thermals above the Wolfberg Cracks, and a Cape leopard had left tracks in the sand near the trailhead — fresh enough that I looked over my shoulder more than once on the walk back. Total observations across all four devices: 37 each. Total battery remaining: FieldLog 62%, CyberTracker 71%, ODK 45%, SMART 28%. The SMART phone had been running GPS continuously for patrol tracking, which explained the drain. But the ODK phone's drain was worrying — the form rendering engine appeared to be keeping the CPU awake.

Magda Ehlers on Pexels`} /> Magda Ehlers on Pexels`} />

Day Two: The Valley With No Signal

We drove east at 06:00, heading into the Breekkrans River valley — deeper into the range, lower, hemmed in by sandstone walls on both sides. Cell signal: zero. Not "one bar if you stand on the roof of the Hilux." Zero. The phones displayed "No Service" with a finality that was almost comforting. This was exactly what we came for.

The transect followed the river course, mostly dry, through thick riverine fynbos. Cape white-eyes worked the bushes. A boomslang crossed the path and disappeared into a rock crevice. We logged 52 observations per device over six hours, all offline.

Offline behaviour is where these apps diverge most dramatically.

CyberTracker was completely unfazed. It has been offline-first since 1996 — before offline-first was a term. Records saved to the local database without any indication anything was missing. The sync process, however, requires a Windows desktop. You connect the phone via USB, open the CyberTracker desktop application, and import the data. This is the 1996 architecture showing through. In the field, you cannot confirm your data has been backed up anywhere. You cannot sync to a server. You trust the phone and you hope. For a tool whose entire brand is "works in the bush," the sync story is surprisingly fragile.

ODK Collect queued submissions locally and showed a pending count. This is good. What was less good: the app could not display previously-submitted observations while offline. The "View Sent" list was empty because the data lives on the server, not the device. If you needed to check whether you had already logged a particular dassie at a particular coordinate — a common field scenario — you could not. ODK treats the phone as a submission terminal, not a data repository. This is an architectural choice with real consequences.

SMART Mobile behaved identically offline as online, which is to say it logged patrol observations to its local database. The patrol tracking continued — GPS breadcrumbs stored locally. The problem was battery. Continuous GPS tracking plus no ability to sync and offload data meant the phone's storage and processor were running full-tilt with nowhere to send anything. By 13:00, with 41 observations logged, the SMART phone was at 9% battery. I attached a power bank. The phone got hot. Not warm — hot. Hot enough that I stopped logging observations on it because I was worried about battery swelling.

FieldLog was in its element — offline-first is our entire architecture. Observations saved to SQLite instantly. The full observation history was browsable, searchable, and filterable while offline. Sync count showed pending uploads. The only issue: the sync pending badge was not prominent enough. I knew what the small number in the corner meant because I built it. A first-time user would not. More on this later.

We got back to camp at 16:00. Still no signal. The data from all four phones was sitting on the devices, unreachable by anyone except the person holding the phone. In a team context — which is most field contexts — this is a problem. If I got hit by a falling rock tomorrow (not impossible in the Cederberg), the data on my phone would be recoverable only by someone who had physical access to the device and knew how to extract it. For CyberTracker and SMART, that requires desktop software. For ODK, the data was not even viewable without the server. For FieldLog, the data was viewable on the device but still trapped there. Offline is a capability. Offline without team visibility is a liability.

Day Three: The Heat

The temperature hit 38°C by 10:30. The kind of heat where your sweat evaporates before it can cool you, where the sandstone radiates back at you like an oven, where you stop wanting to touch anything — including a glass screen.

We ran a transect on the exposed eastern slope of the Wolfberg. No shade. No breeze. Just rock, fynbos, and glare. This was the day the phones started to break — not the software, the hardware. Or rather, the interaction between the two.

Screen Glare

At max brightness, not a single phone was readable in direct sunlight. Not one. The glossy screen on the Tecno Spark Go is a mirror in these conditions. You hold the phone at an angle, shade it with your hand, squint, curse, and eventually turn your back to the sun and hunch over the screen like you are guarding a small fire. This posture is unsustainable for a 12km transect. By observation 30, I was stopping less to log data and more because logging data hurt.

CyberTracker's icon interface was the most resilient here. Large icons, high contrast, minimal text. You could identify and tap the baboon icon through the glare when you could not read the word "baboon." This is an underrated advantage of visual interfaces for field work, and one we are now incorporating into FieldLog's rapid-logging mode.

ODK Collect's text-heavy forms were nearly unusable. Tiny font. Dropdown arrows invisible in the glare. I mis-tapped species selections twice and had to go back and edit records — edits that would require server approval later, which meant connectivity, which meant waiting.

Battery Drain

Heat kills batteries. At 38°C ambient, with the phone in direct sunlight and the screen at full brightness, the internal temperature of all four devices exceeded 45°C. Android's thermal throttling kicked in — the processor slows down, GPS refresh rate drops, screen brightness dims automatically. Everything gets worse.

The SMART phone shut down at 13:30. Dead. The ODK phone hit 5% at 14:00 and I stopped using it to preserve the records. The FieldLog phone was at 18%. The CyberTracker phone was at 34% — CyberTracker's minimal UI and efficient GPS handling meant it was doing less work per observation, and it showed in the battery numbers.

I sat on a rock, drank warm water from a bottle that had been baking in my pack, and made a note:{" "} Battery optimisation is not a feature. It is a safety issue. If a ranger's phone dies halfway through a patrol, they are not just missing data — they are missing their communication device. In an emergency, that matters.

Data Entry Fatigue

By observation 50, I was making mistakes. Wrong age class. Wrong habitat type. Missed the photo entirely. The cognitive load of navigating a complex form while dehydrated, sun-struck, and physically exhausted is not something you can simulate in an office. The form that felt "comprehensive" during setup felt "hostile" by 14:00.

This is the insight that has stuck with me most:{" "} form design is not a UX problem. It is a physiology problem. {" "} A field worker at kilometre 10 of a 12km transect does not have the same fine motor control, visual acuity, or cognitive bandwidth as a UX designer at a standing desk. If your form requires precision, it will fail in the field. If it requires reading, it will fail in the sun. If it requires more than two taps, it will fail by mid-afternoon.

Nikita Igonkin on Pexels`} />

Day Four: The Summit and the Partial Sync

We hiked the Wolfberg Cracks route — a strenuous climb through a narrow rock fissure that opens onto the summit plateau, where the Wolfberg Arch stands like a sandstone gateway to nothing. The altitude here is roughly 1,600 metres. And crucially: there is signal. Weak, flickering, one-bar 3G. But signal.

This was the sync test. All four phones had accumulated two days of offline observations. Let us see what happens when we try to push that data through a connection that can generously be described as "theoretical."

CyberTracker

Could not sync. CyberTracker does not sync over cellular. USB to Windows desktop only. This is architectural, and for a tool launched in 1996 it made sense at the time. In 2026, it is disqualifying. If you cannot push data off the device in the field, you are carrying a single point of failure in your pocket.

ODK Collect

Attempted to sync 89 pending submissions. The connection was 3G with roughly 80KB/s upload. Each submission included a photo — not large by modern standards, roughly 2MB each, but 89 of them is 178MB. At 80KB/s, that is approximately 37 minutes of continuous upload. The connection was not continuous. It dropped six times. ODK handles this well — the sync is checkpoint-resumable, each submission is an independent upload, and failed submissions are retried individually. After 53 minutes of standing on a rock holding the phone above my head, 84 of 89 submissions had uploaded. The remaining five were still queued, the photos too large to push through the intermittent connection. I made a note: ODK needs photo compression before upload.

SMART Mobile

The SMART phone was dead from Day Three. No sync possible.

FieldLog

Synced 84 records in under 2 minutes. This sounds like marketing. Here is why it worked: FieldLog syncs deltas — only changed rows move across the wire. Photos are compressed to 1920px on the long edge before storage, reducing average photo size from roughly 6MB to roughly 400KB. The sync protocol is chunked into batches of 20 records, each batch an atomic checkpoint. The connection dropped twice mid-sync. Both times it resumed from the last checkpoint without re-uploading completed batches.

This is what we built FieldLog for. And in this moment, standing on a sandstone plateau with one bar of 3G, watching data that had been trapped on a device for two days finally leave it, I felt the kind of satisfaction that makes you forget the sunburn.

But I also noticed what was missing:{" "} partial record sync. FieldLog syncs whole records or nothing. If an observation has a photo that fails to upload, the entire observation stays on the device. In ODK, the text data uploads and the photo retries separately. In a conservation context — where the text data (species, count, coordinates) is often more urgent than the photo — ODK's approach is better. I made another note.

We ate lunch on the summit, watching Verreaux's eagles — the same pair, I think — ride the updrafts along the cliff edge. The view stretched to the Tankwa Karoo in the east, a haze of heat and distance. A dassie watched us from a rock three metres away, utterly unafraid. I logged it on all four phones, just to see. CyberTracker: 3 seconds. ODK: 11 seconds. FieldLog: 4 seconds. SMART: still dead.

Nur Andi Ravsanjani Gusma on Pexels`} />

Day Five: The Rain

A cold front pushed through overnight. The temperature dropped from 35°C to 14°C in six hours. Rain began at 04:00 — steady, cold, the kind of rain that finds every gap in your tent fly. By 06:00 the campground was a series of puddles connected by mud. The baboons had retreated to the cliffs. Smart animals.

We ran a short transect anyway. Because in the field, you do not get to skip days because the weather is bad. The monitoring schedule does not care about your comfort.

Touchscreens in Rain

Capacitive touchscreens do not work when wet. Raindrops register as touches. The screen flickers between apps, zooms in and out, opens menus you did not want. You wipe the screen with your sleeve, which is also wet, which makes it worse. You wipe it with a dry cloth from a ziplock bag and get approximately 15 seconds of usable screen time before the rain reclaims it.

CyberTracker's icon interface was again the most resilient. Large touch targets meant fewer mis-taps. The sequential navigation meant you could develop a rhythm — wipe, tap-tap-tap, confirm, pocket the phone — that minimized screen exposure. The phone got wet but the data got logged.

FieldLog was adequate. The rapid-logging mode worked once I found a rhythm, but I had to change the observation flow to use only the largest buttons and avoid dropdowns entirely. Dropdowns in rain are a special kind of frustration — the raindrop hits the list item above or below the one you want, and you cycle through wrong selections until you want to throw the phone into a river.

ODK Collect was almost unusable. The form was too dense, the touch targets too small, the validation too strict. I abandoned the ODK phone after 40 minutes and logged the rest of the transect on the other three devices alone. That is data loss. Not because of a software bug, but because the interaction model{" "} failed under environmental conditions that are routine for field work.

Waterproofness

None of the Tecno Spark Go phones are waterproof. I kept them in ziplock bags with the corners cut for the charging cable when needed, which is the universal field solution and also terrible. The charging ports on three of the four phones collected moisture. One phone — the ODK phone, ironically — stopped charging entirely around midday. I had to dry the port with a corner of shirt and wait an hour before it would accept a charge. If this had happened on Day One with no backup phone, the test would have been over.

Day Six: What Worked and What Didn't

We spent the morning at camp, drying gear and reviewing the data. The sun had returned, the baboons had returned, and the cold shower felt less like punishment and more like absolution. Over coffee — real coffee this time, not the instant sludge of the previous five days — I wrote down everything.

What Worked

Offline-first architecture. Across all four apps, the basic promise held: you can log observations without signal. This should be table stakes, and it is. But the quality of the offline experience varied enormously. An app that lets you log offline but not view your data offline is only half-offline. CyberTracker and FieldLog got this right. ODK did not. SMART was dead.

Icon-based interfaces. CyberTracker's icon grid was the single most field-appropriate design pattern we tested. Large, high- contrast, language-independent, usable in glare and rain. We are rebuilding FieldLog's rapid-logging mode to use this pattern. Not because we want to copy CyberTracker — because CyberTracker got this right in 1996 and nobody else has caught up.

Delta sync with checkpoints. FieldLog's sync was fast, resilient, and efficient. The checkpoint model — batch of 20, atomic commit, resume on failure — is the right architecture for patchy connections. ODK's individual submission model was also resilient but slower due to photo uploads. The lesson: sync architecture matters more than any other technical decision in offline-first software.

Photo compression. FieldLog compressing photos to 1920px before storage saved the sync test. Full-resolution photos are unnecessary for species identification and habitat documentation. ODK's lack of compression made sync painful and will cause problems for anyone on a data-limited plan in the developing world.

What Didn't

Text-heavy forms in the field. ODK Collect is a form engine for surveyors, not a data collection tool for field biologists. The difference is not academic — it is the difference between data that gets logged and data that does not. Small text, dense layouts, and long dropdowns fail under sun, rain, and fatigue.

Desktop-dependent sync. CyberTracker requiring a Windows desktop for data extraction is a dealbreaker in 2026. I understand the history — this architecture was built before smartphones existed — but the failure to modernize the sync path is holding back what is otherwise an excellent field tool.

Heavy setup processes. SMART's configuration complexity is a barrier. Two hours of desktop setup for a phone app that runs out of battery in four hours is a bad ratio. If the tool requires more time to configure than it can operate in the field, the tool is broken.

All-or-nothing record sync. FieldLog's requirement that photos upload with the observation text means a failed photo upload blocks the text data. In conservation, the species and coordinates are often more time-sensitive than the photo. ODK's approach — text first, photos async — is better. We are changing this.

Battery management. The SMART phone dying mid-morning on Day Three was a hardware failure compounded by software design. Continuous GPS tracking on a $50 phone with no battery optimisation is a predictable disaster. Any app that runs GPS continuously must have aggressive power management or it will not survive a full field day.

Sync state visibility. I almost lost data on Day Two because I did not realise the FieldLog phone had not synced. The sync pending badge was too subtle. A ranger in my position — tired, distracted, thinking about the leopard tracks he saw that morning — would absolutely miss it. Making sync state impossible to ignore is not annoying UX. It is honest software.

Day Seven: Departure and What We Changed

We packed camp at dawn. The baboons watched from the cliffs, no doubt planning their next breakfast raid. The Hilux was a layer cake of dust, damp gear, and dead phones. We drove the gravel road back to the N7, stopping at Cederberg Wines to buy two bottles of their Bukettraube — partly for the research value, mostly because we had earned it.

Here's what this week taught us, and what we changed in FieldLog because of it.

1. Icon-Based Rapid Logging Mode

We are building a configurable icon grid for rapid species logging. Admin defines the icons during expedition setup. The field worker taps icons, not text. Three taps to a complete observation: species, count, submit. This is not a replacement for the full form — it is a fast-path for the 80% of observations that are routine. The full form remains for the 20% that need detail. CyberTracker proved this works. We are taking the lesson, not the interface.

2. Partial Record Sync

Observation text data now syncs independently of photos. If a photo fails to upload — timeout, network drop, storage limit — the text record (species, count, coordinates, timestamp) uploads anyway. The photo retries in the background. This means critical data reaches the server even on marginal connections. The photo follows when conditions allow.

3. Sync State That Cannot Be Ignored

The sync pending indicator is now a persistent banner at the top of every screen: "14 observations pending sync — Last synced: 2 hours ago." When sync is pending and the user tries to close the app or log out, a dialog warns: "You have unsynced observations. If you continue, this data may be lost." Bright red. Not subtle. Not optional.

4. GPS Mode Selection

FieldLog now offers two GPS modes: "Fast" (last known location, updates in background — faster observation logging, slightly lower accuracy) and "Precise" (fresh fix on every observation — the current behaviour). The user chooses based on their monitoring protocol. For rapid species counts on a known transect, Fast mode saves 2-3 seconds per observation and significant battery. For precise mapping of rare or endangered species, Precise mode provides the accuracy.

5. Battery-Aware Logging

The app now monitors battery level and temperature. At 20% battery, it reduces GPS polling frequency. At 15%, it dims the screen slightly. At 10%, it enters a "preservation mode" — GPS only on manual request, photo capture disabled, sync deferred. The user can override any of these settings, but the default is survival. A dead phone collects no data.

6. Offline Team Visibility (Roadmap)

This one we could not fix in a week. When the team is offline, each phone is an island. Nobody can see what anyone else has logged. In a coordinated survey, this leads to duplicated effort and missed coverage. Phase 2 of FieldLog includes LoRa mesh networking for offline team sync — phones talking to each other directly, sharing observations without internet. This trip made it clear that Phase 2 is not a luxury. It is essential.

What This Means

Here is what I believe after a week in the Cederberg with four phones and a notebook.

The best field data collection app in the world is not the one with the most features. It is not the one with the best AI. It is not the one with the prettiest interface or the most integrations or the biggest marketing budget.

It is the one that works when your hands are cold. When the sun is blinding. When the rain is falling. When the battery is at 14%. When there is no signal. When you are tired. When you are at kilometre 10 and the only thing you want to do is stop walking. The app that still lets you log the observation under those conditions — that is the one that collects the data that matters.

Everything else is noise.

CyberTracker understood this in 1996. It built an interface for San trackers who could not read, using icons and tap sequences, and it still works — offline, unbreakable, fast. But it never modernized the sync path, and it trapped its users behind a Windows desktop in a mobile world.

ODK built a rigorous form engine with excellent data validation and a huge community, but it optimized for survey methodology, not field biology. Its forms are too heavy, its offline data viewing is too limited, and its photo handling is too naive for the conditions in which it is deployed.

SMART built a law enforcement platform and called it a data collection tool. It is excellent at what it was designed for — ranger patrol management — and awkward at everything else. It is the right tool for a specific job, and almost no other job.

FieldLog is the youngest of the four. We have the advantage of learning from 30 years of other people's mistakes, and the disadvantage of not having 30 years of our own field hardening. This trip showed us exactly where we are soft. The sync state visibility. The GPS latency. The all-or-nothing record upload. These are not bugs we found in code review. They are failures we experienced in the field, standing on a rock in the rain, watching a progress bar that should not exist.

We build FieldLog because we believe field data collection should be fast, reliable, and free. But belief is not enough. You have to go to the places where the tools will be used — the valleys with no signal, the slopes with no shade, the campsites with baboons in the breakfast — and you have to use them until they break. Then you fix them. Then you do it again.

This is not a product launch. It is a field report. The bugs are real. The fixes are in progress. The commitment is unchanged: build tools for the people who do the work, in the places where the work happens, under the conditions those places impose.

And if you are building field software yourself: go to the Cederberg. Or wherever your users are. Leave the emulator behind. Bring the cheap phone. Stand in the sun. Get rained on. Watch the battery die. Your software will tell you the truth about itself. You just have to be willing to listen.

Field Log is a field-first mobile platform built by The Field Company. Offline-first. Team sync. Structured forms and rapid logging. Free to start, your data stays yours.{" "} Get started at fieldlog.thefieldco.com . For more on building conservation technology that respects the people doing the fieldwork, read{" "} What We Learned Building Offline-First Software for the Field .

Field testing conducted in the Cederberg Wilderness Area, Western Cape, South Africa. Permits: CapeNature. Campsite: Sanddrif/Dwarsrivier. Apps tested: FieldLog, CyberTracker, ODK Collect, SMART Mobile. This is a field report, not a controlled study. Conditions were real. Every failure described was reproduced at least once. Every fix described was shipped. We are builders, not marketers.

The Field Co

Open-Source Conservation Technology

How to Contribute to Open Source Conservation Technology — No PhD Required

Wed, 28 Oct 2026 00:00:00 GMT

import { Image } from "astro:assets"; import BlogImage from "../../components/blog/BlogImage.astro"; import HowToContr2 from "../../assets/blog/how-to-contribute/14553713_bibek_ghosh.jpg"; import HowToContr1 from "../../assets/blog/how-to-contribute/10725897_muhammed_ensar.jpg"; import codingCollab from "../../assets/blog/how-to-contribute/374559_digital_buggu.jpg"; import developersCollab from "../../assets/blog/how-to-contribute/4816921_myburgh_roux.jpg";

You do not need a PhD in ecology. You do not need to live near a rainforest. If you can code, design, write, translate, or test software, there is a conservation project that needs you — today. Not after a training programme. Not after a career change. Now.

The open source conservation technology ecosystem is larger, more welcoming, and more desperate for contributors than most developers realise. Projects tracking endangered species, detecting poachers via AI, mapping deforestation in near-real-time, and giving rangers offline data collection tools that work on $50 phones — all of these are open source. All of them need help. And almost none of them require you to understand ecology before you can contribute.

This is a guide to exactly where to start. Not a think piece. A directory of projects, skills, and contribution pathways that are open right now. If you have ever wanted your skills to matter for the planet, here is where you begin.

Why Open Source Matters in Conservation

Conservation runs on five-year grant cycles. An NGO gets funded by a foundation, buys software licences, trains their staff, and integrates the tool into their workflow. The grant ends. The vendor raises prices or pivots to a different market. The tool becomes abandonware. The rangers who depend on it are left with unsupported software that gradually stops working on newer devices. This is not hypothetical. It has happened to field teams on every continent.

Open source breaks this cycle. When the code is public, the tool outlives its original funding. When CyberTracker's original developer moved on, the community kept it running for 28 years. When SMART's founders needed to guarantee longevity, they built an 8-organization partnership and merged operations with EarthRanger under the SERCA alliance. These projects survive because no single entity has to carry them alone. If you build something useful, and the code is open, it does not die when you move on.

There is a second reason: auditability. Conservation data increasingly ends up in court — poaching prosecutions, land rights disputes, environmental impact lawsuits. A closed-source app is a black box to a judge. You cannot prove the software did not alter the GPS coordinates. You cannot verify the timestamp chain. With open source tools, you can. The code is public. The data pipeline is inspectable. The chain of custody is verifiable from the moment a ranger taps "record" to the moment a prosecutor presents it as evidence. When species protection laws depend on data integrity, open source is not a preference — it is a requirement.

The third reason is community ownership. Conservation data collected with public funds — government grants, university budgets, NGO donations — should be held by public infrastructure. An open source tool that anyone can deploy, audit, and fork means the data stays with the community that collected it. No vendor lock-in. No proprietary formats. No "your subscription has expired, export your data in the next 30 days." This is why ODK, used by the WHO for disease surveillance across 1.4 billion people, is open source. This is why EarthRanger, monitoring 900+ conservation sites, is free. This is why we built FieldLog under a permissive licence. The tools that protect what remains must belong to everyone.

Myburgh Roux on Pexels`} /> Digital Buggu on Pexels`} />

What Skills Are Needed

Conservation technology is not one thing. It is a stack of interdependent problems — data collection, AI processing, hardware deployment, user interface, documentation, translation — and each layer needs different skills. Here is what is needed right now.

Skill	What You Would Do	Example Project	Time
Developer	Write Python, JS, Kotlin, or React Native. Fix bugs, build features, improve offline sync. Most projects have "good first issue" tags.	ODK Collect (Kotlin), iNaturalist Seek (TypeScript), MegaDetector (Python), FieldLog (React Native/Python)	An evening to ongoing
Designer	Design icon-based interfaces for low-literacy users. Improve accessibility for bright sunlight, wet hands, old devices. Create low-bandwidth UI patterns.	FieldLog (icon sets for African wildlife), iNaturalist Seek (accessibility), ODK Collect (Android UI)	A weekend to ongoing
Writer	Write user guides, technical docs, field manuals. Translate developer-speak into something a ranger can use at dawn with no signal.	ODK Docs (Sphinx, 6 "help wanted" issues), SMART manuals, MegaDetector docs	An hour to a sprint
Translator	Translate UI strings, forms, and documentation. Swahili, Portuguese, French, Zulu, Xhosa, Afrikaans are all urgently needed by conservation projects operating across Africa.	ODK (multi-language), iNaturalist (global app), FieldLog (local African languages), SMART (English/Spanish/French)	30 minutes per session
Field Tester	Install an app on a $50 phone, go outside, report what breaks. Beta test in real conditions — dirt, rain, bad signal, dead battery.	FieldLog pilot programme (Southern Africa), EarthRanger beta, ODK release testing	One expedition to ongoing
Data Annotator	Label camera trap images. Classify bioacoustics recordings. Verify AI-generated species IDs. Every labelled image trains a model that saves a researcher weeks of manual review.	Snapshot Safari (Zooniverse), European Camera Trap Project (Zooniverse), iNaturalist (identify), eBird checklists	10 minutes whenever
Hardware Tinkerer	Build open-source sensors, modify camera traps, contribute to Arduino and Raspberry Pi conservation projects. Improve waterproofing, battery life, assembly instructions.	SPARROW (solar edge AI device, full BOM), AudioMoth ($70 acoustic sensor), Arribada Initiative (turtle tags, GPS loggers)	A weekend build to ongoing

If you are unsure which path fits, start with documentation or translation. These are the highest-leverage, lowest-barrier contributions in open source, and conservation projects are critically under-resourced in both. A translated UI can unlock a tool for millions of people. A well-written setup guide prevents days of frustration for a field team that has no IT support.

Where to Start — Specific Projects

Here are the projects with open issues, contribution guides, and community channels that are welcoming contributors right now. Each one has a direct path from "I want to help" to "I shipped something."

Microsoft MegaDetector / PyTorch-Wildlife

GitHub: microsoft/MegaDetector — Camera-trap animal detection AI used by 80+ conservation organisations.
Discord: discord.gg/TeEVxzaYtm
Contribute: Microsoft Biodiversity Contribution Guidelines
Needs: Python developers, ML engineers, documentation writers.
Start: Browse the PyTorch-Wildlife project board, pick a task from the "Ready" column, comment to claim it.

ODK (Open Data Kit)

GitHub: github.com/getodk — Offline data collection used by 2M+ people. 250M submissions annually.
Forum: forum.getodk.org — 17,000 members.
Contribute: docs.getodk.org/contributing
Open issues tagged "help wanted": 5 on Central, 10 on Collect, 6 on Docs.
Needs: Kotlin (Android), JavaScript (Node.js/Vue.js), TypeScript, Python, technical writers.

iNaturalist / Seek

GitHub: github.com/inaturalist — 300M+ observations, 400K active users, the world's largest biodiversity social network.
"Good first issue" tagged: Seek React Native (TypeScript).
"Help wanted" tagged: iNaturalist API (JavaScript).
Needs: TypeScript, React Native, JavaScript, Ruby on Rails, Java, Swift, Objective-C, translators, designers.

EarthRanger / SERCA

Website: earthranger.com — Real-time protected area monitoring. 900+ sites, 23K+ animals tracked. Now part of the SMART-EarthRanger Conservation Alliance (SERCA).
Community: community.earthranger.com
Beta programme: earthranger.com/beta
Needs: Beta testers, field testers, integration developers, documentation writers.

Zooniverse

GitHub: github.com/zooniverse — The world's largest citizen science platform. 374 repos, 70+ active projects.
Active conservation projects: European Camera Trap Project, Cameras for Conservation, Snapshot Safari.
Code: JavaScript, Ruby, Python, React Native. No-code: Classify images directly on the platform.

SMART Conservation Software

Website: smartconservationtools.org — Patrol monitoring at 1,100+ sites in 95 countries.
Partners: WWF, WCS, Panthera, Frankfurt Zoological Society, ZSL, Re:wild, North Carolina Zoo.
Community forum: SMART Community Forum
Needs: Field testers, trainers, documentation translators (English → Spanish, French, Swahili, Portuguese).

SPARROW (Microsoft)

GitHub: microsoft/SPARROW — Solar-powered AI edge device. Full open-source BOM (bill of materials), assembly guide at aka.ms/sparrowassembly.
Needs: Hardware engineers, embedded systems developers, field deployers, Python developers, Docker tinkerers.

The Field Company / FieldLog

We are building a better CyberTracker — offline-first, AI-powered, free forever. Data collection for field researchers, designed for $50 phones and zero signal.
GitHub: github.com/thefieldcompany
Needs: React Native developers, Python/FastAPI developers, designers (icon sets, accessible mobile UI), translators (Zulu, Xhosa, Afrikaans, Swahili), field testers in Southern Africa. See the Join Us section below for specifics.

Muhammed Ensar on Pexels`} />

No-Code Ways to Help

You do not need to write code to contribute to conservation technology. Here is what you can do from your phone or laptop, right now, with zero setup.

iNaturalist — Photograph and Identify Species

Download the app, photograph a plant or animal, upload it. The community and AI will help identify it. Every verified observation trains species classification models and feeds the GBIF global biodiversity database. If you are a subject expert, spend 10 minutes a day verifying other people's observations. The AI needs ground truth data to improve. You provide it.

eBird — Submit Bird Checklists

Every time you notice birds — a walk in the park, your back garden, anywhere — record what you saw in eBird. Millions of these casual observations power the models that track bird populations, migration patterns, and climate-driven range shifts globally. No expertise required. Start with the birds you can identify and learn more over time.

Zooniverse — Classify Camera Trap Images

Open European Camera Trap Project and start identifying animals. Or Snapshot Safari — African wildlife from camera traps across the continent. Or Savanna Spy: Sound — verify AI-classified bird calls from Texas savanna. Each classification is a training data point. Do a hundred, and you have meaningfully improved an AI model that researchers deploy at scale.

Translate — Unlock Tools for Millions

Most conservation tools are built in English, then need translation into Swahili, Portuguese, French, and dozens of local languages before field teams can use them. This is one of the highest-impact, lowest- barrier contributions in open source. ODK, iNaturalist, and FieldLog all need translators. You do not need to be a developer. You need to be fluent in a language used in a conservation area and willing to spend 30 minutes a week translating strings.

Share What You Find

Open source conservation projects suffer from a visibility gap. Most developers have never heard of MegaDetector or ODK. If you try a project and it works, tell people. Write a blog post. Post on LinkedIn. Send it to your team. The single biggest barrier to contribution is that people do not know these projects exist.

Bibek ghosh on Pexels`} />

The People Who Do This

You do not need a conservation background to make conservation impact. Here are three profiles of people who contributed without one.

The Translator Who Unlocked East Africa

A university student in Nairobi, fluent in Swahili, English, and Kikuyu, started contributing Swahili translations to ODK Collect. Within six months, their translations were deployed to thousands of community health workers and conservation rangers across Tanzania, Kenya, and Uganda. They never wrote a line of code. They simply made an existing tool accessible to people who could not read English. If you speak a language used in conservation areas, your translation work has a direct line to impact that most code contributions will never match.

The Retired Teacher Who Trained AI

A retired teacher in the American Midwest spent 30 minutes a day identifying animals in Snapshot Safari camera trap images on Zooniverse. Over two years, their 20,000+ classifications became training data for species detection models deployed across African national parks. They were not a biologist. They were not a coder. They looked at photos of animals and clicked the right button. Consistent, careful work, repeated over time, directly improved AI that saves researchers months of manual image review.

The UX Designer Who Fixed Accessibility for 400,000 People

A designer tired of building corporate dashboards found iNaturalist Seek's "good first issue" tag. They contributed accessibility improvements — larger touch targets, better colour contrast, clearer icon labels — that made the app usable for children, elderly users, and people with vision impairments. Their design work is now in the pockets of over 400,000 citizen scientists worldwide. They did not need to understand ecology, machine learning, or mobile development. They needed to understand users.

Join Us

We are The Field Company. We are building FieldLog — an open source, offline-first data collection platform for conservation. It works on $50 phones. It works without signal. It is free forever.

We need help. Specifically:

Developers: Python (FastAPI), React Native, offline sync architecture. Our stack is SQLite on device, PostgreSQL on server, delta-based sync protocol.{" "} github.com/thefieldcompany

Designers: Icon sets for African wildlife (mammals, birds, reptiles, insects, plants). Accessible mobile UI patterns for bright sunlight, gloved hands, low-literacy users. Low-bandwidth sync indicators that are impossible to miss.

Field Testers: Conservation researchers and rangers in Southern Africa. We need people running real expeditions with FieldLog and telling us everything that breaks. You are the most important contributor we have.

Translators: Zulu, Xhosa, Afrikaans, Swahili. These languages are spoken by millions of people in conservation areas. Our app is in English. This is a barrier we need to remove.

Writers: User guides. Developer docs. Tutorial videos. Field manuals. If you can write clearly about software, we have an endless supply of things that need explaining.

The code is open source. The roadmap is public. The community is growing. If you want your skills to matter for the planet, here is exactly where to start.

GitHub:{" "} github.com/thefieldcompany
Join the Discord (link coming — or email us through the site).

Community Hubs

These are the places where conservation technologists find each other, share work, and coordinate across projects.

Hub	What It Is	Link
WILDLABS	Largest conservation tech community. Forums, events, job board, research.	wildlabs.net
Microsoft Biodiversity Discord	MegaDetector, PyTorch-Wildlife, SPARROW community. Ask questions, share results, contribute.	discord.gg/TeEVxzaYtm
ODK Forum	17,000 members. All things ODK — Collect, Central, forms, deployments.	forum.getodk.org
EarthRanger Community	EarthRanger users and developers. Training, beta programme, support.	community.earthranger.com
Zooniverse	70+ citizen science projects. Classify, contribute, or build your own project.	zooniverse.org/projects
Conservation X Labs	Innovation hub, competitions, funding for conservation tech.	conservationxlabs.com
SMART Community Forum	SMART users and practitioners. Deployment support, training.	smartconservationtools.org → Community

The planet is under pressure. Species are disappearing. The people doing the work — rangers, researchers, community members — need tools that work, and they need them to be free and open. If you can write code, design interfaces, translate strings, test apps, or label images, you are already qualified to help. Pick a project from this post. Open a GitHub issue. Join a Discord. Make your first contribution. The tools that protect what remains need you.

Projects referenced:{" "} FieldLog, CyberTracker, ODK, SMART, EarthRanger, MegaDetector, iNaturalist, eBird. Community hubs:{" "} The Field Company Discord, WILDLABS, ODK Forum. "Good first issue" tags accurate as of June 2026. Always check each project's contribution guide before submitting.

The Field Co

Open-Source Conservation Technology

The Amazon Is Approaching a Savanna Tipping Point

Fri, 11 Sep 2026 00:00:00 GMT

import { Image } from "astro:assets"; import BlogImage from "../../components/blog/BlogImage.astro"; import AmazonTipp2 from "../../assets/blog/amazon-tipping/12027849_alexandre_p._junior.jpg"; import AmazonTipp1 from "../../assets/blog/amazon-tipping/12027846_alexandre_p._junior.jpg"; import amazonAerial from "../../assets/blog/amazon-tipping/35313820_pok_rie.jpg"; import amazonRiverAerial from "../../assets/blog/amazon-tipping/37786898_pok_rie.jpg";

A single large tree in the Amazon releases a thousand litres of water into the atmosphere every day. Multiply by 390 billion trees across 5.5 million square kilometres and you get a hydrological engine of planetary scale. The Amazon does not merely receive rain. It generates roughly half its own — pumping moisture skyward through evapotranspiration and recycling it five or six times as the air mass moves from the Atlantic coast toward the Andes.

This engine feeds agriculture as far south as Argentina. It stabilizes the climate of an entire continent. It stores carbon equivalent to 15–20 years of current global CO₂ emissions. It holds more than 10% of all known species on Earth.

It is also, by the best available evidence, beginning to fail.

What Is the Amazon Tipping Point?

In 2007, a team led by Gilvan Sampaio and Carlos Nobre ran simulations asking how much of the Amazon could be cleared before the rainfall-recycling system broke down. Their answer: 40% deforestation would cause a sharp, basin-wide drop in rainfall. In 2018, the late Thomas Lovejoy and Carlos Nobre lowered the estimate. They argued 20–25% deforestation — combined with rising temperatures and an accelerating fire regime — could be enough.

The mechanism is straightforward. Trees transpire water. That water becomes rainfall downwind. Remove enough trees and the rainfall declines. Less rain means more drought stress. Stressed trees die, or burn, which removes more trees — reducing transpiration further. The system crosses a feedback threshold where the forest can no longer sustain the humidity it requires to be a forest.

The neighbouring Cerrado — a vast savanna-forest mosaic south of the Amazon — offers a preview of what that looks like. The Cerrado was never rainforest, but it demonstrates that under the same broad climate conditions, two very different ecosystems can exist as stable states. Cut the Amazon and it may not grow back as Amazon. It may grow back as Cerrado. Or as degraded grassland.

Threshold	Value	Source
Deforestation tipping point	20–25%	Lovejoy & Nobre (2018)
Sharp rainfall collapse	40%	Sampaio & Nobre (2007)
Global warming threshold (dieback)	3.5°C (range 2.0–6.0°C)	Armstrong McKay et al. (2022)
Carbon released — partial dieback	30 Gt C	Armstrong McKay et al. (2022)
Carbon released — total dieback	75 Gt C	Armstrong McKay et al. (2022)
Total carbon stored in Amazon	150–200 Gt C	Flores et al. (2024)

Pok Rie on Pexels`} />

Where Are We Now?

The Amazon has lost roughly 17–20%{" "} of its original forest cover. Another 6%{" "} is highly degraded. Taken together, about a quarter of the biome is compromised.

Year	Deforestation (Brazilian Amazon, km²)	Context
2004	27,772	All-time peak
2012	4,571	Historic low — PPCDAm enforcement working
2019	9,762	First Bolsonaro year — enforcement dismantled
2021	13,038	Bolsonaro-era peak — highest since 2006
2023	9,001	Lula — enforcement resumed, decline begins
2024	6,288	Continued decline — but fires surged
2025	5,264	Lowest since 2012

The chainsaw trend under Lula is encouraging. The problem is that deforestation is no longer the only driver. In 2024, fires — not chainsaws — caused 66% of Brazil's forest loss, a more than sixfold increase from 2023.{" "} 6.7 million hectares of tropical primary forest were lost globally in 2024 — nearly double 2023 — and Brazil accounted for 42% of that total. The Amazon experienced its highest tree cover loss since 2016.

The 2024 Drought

The Amazon's rivers are its highways. In 2024, those highways ran dry. The Rio Negro at Manaus fell to its lowest level in 121 years of measurement. The Solimões, Madeira, and Amazon rivers all hit record lows. More than{" "} 420,000 children were affected across Brazil, Colombia, and Peru — schools and health clinics cut off, clean water inaccessible, food supplies stranded. Over 120 Amazon river dolphins died in Tefé Lake in 2023 when water temperatures exceeded 39°C; the 2024 drought compounded the crisis.

This was not an anomaly. It was the fifth "once-in-a-century" drought in 20 years: 2005, 2010, 2015–16, 2023, 2024. The World Weather Attribution group found climate change is the primary driver. The droughts are not natural variability. They are the new climatology.

Pok Rie on Pexels`} /> Alexandre P. Junior on Pexels`} />

The Fire Feedback Loop

For most of its history, the Amazon did not burn. The forest was too wet. Fire was rare and when it occurred, it stayed small.

The mechanism that changed this is simple enough to describe in four sentences:

Deforestation and logging open the canopy. Sunlight reaches the forest floor and dries it. Fires set on adjacent pasture or cropland — intended to clear brush or renew grazing — breach the forest edge. Large, slow-growing trees with thin bark and no fire adaptation die. Their deaths open the canopy further, drying the understory more, making the next fire worse.

The cycle accelerates. A study by Paulo Brando and colleagues (2014, PNAS) found that a single drought-fire event could increase tree mortality by five times the baseline rate. Repeated fires transform forest structure. Grasses and pioneer species move in — and these burn far more readily than closed rainforest. The fire regime itself can "tip" from localized burns to uncontrollable mega-fires.

Alexandre P. Junior on Pexels`} />

What Happens When It Flips

1. Regional Rainfall Collapse

A deforested Amazon produces 20–30% less rainfall across the basin. The eastern and southern Amazon — already the driest regions — are hit hardest. These are precisely the agricultural frontier zones where soy and cattle production are concentrated.

The "flying rivers" — atmospheric moisture streams that carry Amazon rainfall south — weaken. This affects the Pantanal (the world's largest tropical wetland), the La Plata river basin, and agricultural regions as far south as Uruguay and northern Argentina. Brazil is the world's largest exporter of soy, beef, coffee, sugar, and orange juice. All of these depend on rain that the Amazon generates. Deforestation for agriculture destroys the rainfall that agriculture depends on. The World Bank calls this "agro-suicide." Their 2023 report estimated economic losses from deforestation could be{" "} ~7× the value of all commodities produced on the cleared land.

2. The Carbon Bomb

Cross the tipping point and the Amazon flips from net carbon sink to net carbon source. This is not a projection — in parts of the southeastern Amazon, it has already happened.

Region	Carbon status (2010–2018)	Source
Southeastern Amazon	Net carbon source	Gatti et al. (2021), Nature
Northeastern Amazon	Roughly neutral	Gatti et al. (2021)
Southwestern Amazon	Small net sink	Gatti et al. (2021)
Western Amazon	Net carbon sink	Gatti et al. (2021)

The Amazon stores 150–200 billion tonnes of carbon. A full dieback would release the equivalent of 15–20 years of current global CO₂ emissions on top of everything else humanity is already emitting. It would, in practical terms, make the Paris Agreement targets unachievable.

3. The Biodiversity Cost

The Amazon holds more than 10% of all known species on Earth: 15,000+ tree species, 2.5 million insect species, 2,200 fish species, 1,294 bird species, 427 mammals. A single hectare in the central Amazon can contain more than 300 tree species — more than exist in all of Britain.

A 2019 analysis found that{" "} up to 50% of Amazon tree species{" "} could become threatened by 2050 through the combined effects of deforestation, degradation, and climate change. Species loss is irreversible on any timescale that matters to human civilization. Evolution operates in millions of years. We are operating in decades.

Is It Inevitable?

Four things are true simultaneously — and together they define the arena in which this plays out.

1. Deforestation is dropping.

The Lula government has reversed much of the Bolsonaro-era surge. 2025 deforestation was roughly one-third{" "} the 2021 peak. Enforcement agencies are functioning again. IBAMA levied{" "} R$365 million (US$64 million) in fines on cattle companies — including JBS, the world's largest meat packer — for involvement in illegal deforestation in October 2024. The 2006 Soy Moratorium proved supply-chain interventions work: only 1% of new soy expansion occurred at the expense of forest in the following eight years. Brazil has done this before. The 2004–2012 decline in deforestation reduced annual loss by more than 80%.

2. Indigenous territories work.

The Amazon contains 3,344 formally acknowledged Indigenous territories. Deforestation rates inside them are 2–3× lower than outside. In Peru, legal land titling for Indigenous communities reduced deforestation by 75%. Indigenous land rights are the single most effective and cost-efficient forest protection mechanism available — and they cost a fraction of technological or enforcement-only approaches.

3. But climate change is accelerating independent of chainsaws.

Even if deforestation drops to zero — and no credible scenario has it doing so — the warming and drying already locked in from global emissions will continue to stress the forest. A 2025 study in{" "} Nature Communications found that deforestation-induced drying lowers the Amazon's climate threshold: the more forest we clear now, the lower the global warming at which dieback becomes inevitable. Two stressors are interacting in the worst possible direction.

4. The resilience is already draining.

The 2022 Boulton, Lenton & Boers study — using satellite-derived vegetation data from 1991 to 2016 — found that{" "} more than three-quarters of the Amazon rainforest has been losing resilience since the early 2000s. This loss is greatest in drier regions and in areas closer to human activity. The 2024 Flores et al. analysis found that{" "} 10% to 47% of the Amazon could be exposed to compounding disturbances by 2050 that trigger ecosystem transitions.

The honest assessment: zero deforestation is achievable. Brazil has demonstrated that before. But preventing the Amazon's transition in a warming world requires both zero deforestation and{" "} global emissions reductions consistent with 1.5°C. Neither condition is currently being met.

The Amazon tipping point is not a single switch that flips on a Tuesday. It is a cascade of regional transitions — the southeastern forest already emitting more carbon than it absorbs; the eastern margins fraying into grassland; the fire regime accelerating beyond human control; the rainfall pump weakening year by year. At some point in this cascade, the word "rainforest" no longer describes the system. What replaces it will store less carbon, support fewer species, and generate less rain. South American agriculture will adjust to a drier continent. The global carbon budget will tighten further. And the planet will have lost one of its primary life-support systems — not to a meteorite or a supervolcano, but to a combination of cattle, soy, fire, and heat.

The numbers do not doom-scroll. They simply sit there, unambiguous. The forest is at roughly 17–20% deforested. The threshold sits at 20–25%. The margin is anywhere from zero to five percentage points.

This post is part of a broader series on the planetary systems that keep Earth habitable. For the full picture — all nine planetary boundaries, their current status, and what happens when they interact — read{" "} Planetary Boundaries Explained{" "} and{" "} A Letter to Humanity.

Sources

Key research: Nobre, C.A. & Borma, L.S. (2009) "Tipping points for the Amazon forest" Current Opinion in Environmental Sustainability. Lovejoy, T.E. & Nobre, C. (2018) "Amazon Tipping Point"{" "} Science Advances. Boulton, C.A., Lenton, T.M. & Boers, N. (2022) "Pronounced loss of Amazon rainforest resilience since the early 2000s" Nature Climate Change. Gatti, L.V. et al. (2021) "Amazonia as a carbon source linked to deforestation and climate change"{" "} Nature. Berenguer, E. et al. (2021) "Drivers and ecological impacts of deforestation and forest degradation" in Nobre et al. (eds.){" "} Amazon Assessment Report 2021. Deforestation data: INPE/PRODES (Brazilian National Institute for Space Research). Fire data: Brando, P.M. et al. (2014){" "} PNAS. See also Planetary Boundaries Explained and A Letter to Humanity.

The Field Co

Open-Source Conservation Technology

How Much Does a Camera Trap Survey Cost?

Fri, 14 Aug 2026 00:00:00 GMT

import { Image } from "astro:assets"; import BlogImage from "../../components/blog/BlogImage.astro"; import cameraTrapField from "../../assets/blog/camera-trap-cost/11749416_matheus_bertelli.jpg"; import trapEquipment from "../../assets/blog/camera-trap-cost/16662872_veysel_boz.jpg";

You are planning a camera trap survey. You have read the TEAM protocol. You know you need Browning or Reconyx units, SD cards, Python locks. But what does this actually cost — end to end?

The honest answer is that most published papers do not tell you. They mention “60 camera stations” but omit the $3,000 in batteries, the 800 hours of image classification, and the three cameras the elephants destroyed.

This is the budget breakdown we wish someone had handed us before we learned all of this the expensive way.

$500 — The Pilot Survey

You have enough for a handful of cameras and some SD cards. You are probably a student, or testing an idea before writing a grant.

What You Buy

2× Alpha Cam Dual Lens No Glow ($120/ea) or used Browning: $240
4× SanDisk 32GB SD cards ($20/ea): $80
2× Master Lock Python cables ($25/ea): $50
16× Energizer Ultimate Lithium AA ($2.50/ea): $40
Camo tape + desiccant packs: $30
Total hardware: ~$440

What You Get

Presence/absence data for medium-large mammals at 2 sites. Enough for a proof-of-concept or a BSc thesis chapter. You will spend 40–80 hours classifying images yourself.

What You Cannot Do

Occupancy modeling. Density estimation. Anything publishable beyond “we detected X species.” No cellular — you see the data when you retrieve the cards.

The Real Cost Nobody Mentions

Your time. If you value your labor at $20/hour, you will spend $800–1,600 classifying images. The “free” labor is the subsidy that makes this tier work.

$5,000 — The MSc Survey

You have a small grant or departmental funding. You need publishable results.

What You Buy

12× Browning Strike Force Pro DCL ($160/ea): $1,920
12× Python cable locks ($25/ea): $300
12× security cases ($40/ea): $480
24× SanDisk 32GB SD cards ($20/ea): $480
Rechargeable battery system — 12 cameras × 8 AA + 2 sets + smart chargers: $800
1× Garmin GPS unit: $250
Shipping/import buffer: $300
Equipment subtotal: $4,530

Field Costs

Deployment (4 days, 2 techs at $50/day): $400
Retrieval visits ×2 (6 days, 2 techs): $600
Vehicle hire (10 days at $100/day): $1,000
Fuel: $150
Field subtotal: $2,150

Data Processing

MegaDetector AI filtering: Free
Human classification of ~80,000 filtered images at 250/hr = 320 hours
If done by student (free): $0
If paid annotator at $15/hr: $4,800

Grand total: $6,680–11,480

What You Get

12-station grid. Occupancy modeling for 5–10 common species. A solid MSc thesis. With cellular on 2–3 cameras (add $500), you get near-real-time monitoring of key locations.

What You Cannot Do

Density estimation for wide-ranging carnivores. Rare species analysis (insufficient detections). Community-level inference.

Reality Check

A real 2025 MSc survey in Kenya by the authors ran 15 cameras for 60 days. Equipment cost was $4,200. Field costs (vehicle + per diem) were $2,800. Image processing was 400 volunteer hours. Published a decent occupancy paper. Total cash outlay: $7,000.

$50,000 — The Professional Survey

Grant-funded research, NGO monitoring program, or well-resourced PhD. This is where the science gets serious.

What You Buy

50× Browning Defender Pro Scout Max HD ($200/ea): $10,000
10× Reconyx HyperFire 4K for key stations ($400/ea): $4,000
60× Python cables + keyed-alike padlocks + security boxes ($80/set): $4,800
10× cellular upgrade (Spartan GoCam 2M + data plans, 1 year): $3,000
120× SD cards (64GB, industrial-grade): $3,000
Full rechargeable battery system (50 cameras, 3 sets each, smart chargers): $4,000
3× Garmin GPS + field tablets: $1,200
Spares and loss buffer (15%): $4,500
Equipment subtotal: $34,500

Field Costs (6-Month Survey)

Vehicle (4×4 rental, 6 months): $12,000
Fuel: $3,000
3 field technicians (6 months at $800–1,200/month): $18,000
Permits, insurance, admin: $3,000
Field subtotal: $36,000

Data Costs

MegaDetector + cloud VM processing: $100
Contract image annotators (800 hours at $20/hr): $16,000
AWS S3 storage (1 year): $500
Wildlife Insights (free for academic): $0
Data subtotal: $16,600

Grand total: ~$87,100

What You Get

60-station grid covering 500–1,000 km². Robust occupancy and community analysis. Real-time data pipeline on key stations. Publishable in Conservation Biology, Journal of Applied Ecology, etc. Data that an NGO can actually use for management decisions.

What You Cannot Do

National-scale inference. Multi-year population trends (need repeat surveys). Comprehensive carnivore density (need specialized designs).

Where to Cut

Use all non-cellular cameras and visit monthly instead. Saves $3,000 in cellular costs but adds field labor. Use student/volunteer annotators via Zooniverse. Saves $16,000 in annotation costs.

Matheus Bertelli on Pexels`} />

$500,000+ — The National Monitoring Program

Multi-year, multi-site, national-scale biodiversity monitoring. Think TEAM Network, Snapshot Safari scale. Funded by multilaterals (GEF, World Bank), bilateral aid (USAID, EU), or national governments.

What You Buy

300 cameras (institutional bulk pricing, mix of Browning/Reconyx/Spartan): $60,000
60 cellular units + multi-year data contracts: $25,000
Full security, mounting, battery infrastructure: $45,000
2× dedicated 4×4 vehicles (purchased): $70,000
6× full-time field staff (annual): $72,000
2× data managers/analysts (annual): $50,000
Server infrastructure + cloud + backup (3 years): $40,000
Custom software/API development: $30,000
Consumables (batteries, cards, desiccant — annual): $25,000
Equipment replacement fund (annual): $20,000
Training workshops, capacity building: $25,000
Permits, admin, institutional overhead (20%): $80,000
First year total: ~$542,000

Annual recurring (years 2–5): ~$200,000–250,000 (labor + consumables + replacement + data infrastructure)

What You Get

National biodiversity baseline. Multi-year population trends for 30+ species. Real-time dashboard for protected area managers. Feeds into CBD Kunming-Montreal GBF Target 4 reporting. Capacity building for 10–20 local researchers. Multiple PhD theses and 20–30 papers.

Comparison Table

	$500	$5,000	$50,000	$500K+
Cameras	2–4	12–15	50–60	200–300+
Grid size	<5 km²	50–100 km²	500–1,000 km²	National
Duration	30 days	60–90 days	6–12 months	3–5 years
Occupancy modeling	No	Common species	15–25 species	30+ species
Density estimation	No	No	Limited	Full
Real-time data	No	Optional	Yes (key stations)	Full pipeline
Staff	You	You + 1–2 techs	3–5 staff	6–10+ staff
Labor (field + data)	40–80 hrs	200–400 hrs	1,500–2,500 hrs	15,000+ hrs/yr
Publishable output	BSc thesis	MSc thesis, 1 paper	PhD, 3–5 papers	Multiple PhDs, 20+ papers

Veysel Boz on Pexels`} />

The Hidden Costs People Forget

After talking to dozens of field researchers and running our own surveys, these are the budget items that show up uninvited.

1. Batteries Will Bankrupt You If You Pick the Wrong Camera

Two cameras at the same retail price can have wildly different battery consumption. The Browning Defender Pro Scout Max HD draws 0.14 mW at rest. The Tactacam Reveal Pro 3.0 draws 2.74 mW. That is a 19× difference in standby power.

Over 3 years, the Browning costs $144 in lithium batteries. The Tactacam costs $648. Multiply by 50 cameras, and you are looking at $7,200 vs. $32,400 — a $25,000 difference from one purchasing decision.

What to do: Check resting current draw before buying. TrailCamPro publishes these numbers. Prioritize low-draw models for remote deployments where battery changes are expensive.

2. Image Processing Is a Full-Time Job

A 60-camera, 60-day survey generates approximately 200,000 images. At 250 images per hour (realistic sustained rate), that is 800 hours — 20 weeks of full-time work.

If your grant budget has $0 for image classification, you are the image classifier. If you are the PI, you just lost 4 months.

What to do: Run MegaDetector (free) first. It will remove 70–95% of images as blanks/false triggers. Budget for annotation labor — even $5,000 for a contract annotator saves you 4 months of your life.

3. Equipment Loss Is Not an Edge Case — It Is Expected

Meek et al. (2019) surveyed 153 camera trap researchers. Annual equipment loss in tropical developing countries: 5–15%. Some projects reported 30%+.

This is not just theft. Elephants crush cameras. Floods submerge them. Fungus grows inside lenses. Ants colonize housings. Baboons chew off external antennae.

In tropical conditions, assume a 3-year replacement cycle for all cameras. Budget 15% of your equipment value annually for loss/damage.

4. Shipping Lithium Batteries Is a Regulatory Headache

International shipping of lithium batteries triggers IATA dangerous goods regulations. Carriers may refuse shipments. Customs in destination countries may hold packages for weeks. Import duties on electronics in some African countries reach 25%.

For a $20,000 equipment shipment, budget $1,000–3,000 for shipping and $2,000–5,000 for duties and clearance fees.

5. The Spare Camera Problem

You planned 50 stations. You bought 50 cameras. Week 3, three cameras fail (humidity, battery defect, elephant). Now you have 47 stations and a gap in your sampling design. Your occupancy model’s precision drops. You cannot publish with methodological gaps.

Always buy 15–20% more cameras than your target station count. For 50 stations, buy 60 cameras. The spares also serve as loaners during maintenance rotations.

6. Data Has Ongoing Costs

You finish the survey. The grant ends. Your 600GB of images are on AWS S3 at $14/month. In 5 years, when someone asks for your raw data for a meta-analysis, who is paying the $840 in storage costs? Who is maintaining access?

Plan for data archiving in your budget. Dryad charges $120 per data package. Zenodo is free. Wildlife Insights provides free archival for academic users.

So What Should You Actually Budget?

Here is a rule of thumb that holds across scales:

Equipment: 30–40% of total budget (cameras, accessories, batteries, security)
Field labor & transport: 25–35%
Data processing: 15–25%
Administration, permits, overhead: 10–20%
Loss buffer: 5–10% of equipment value

For a $50,000 survey: $15–20K equipment, $12–17K field ops, $8–12K data processing, $5–10K admin.

The ratio shifts toward labor and data at higher budgets, and toward equipment at lower budgets.

The Bottom Line

Camera trapping is the most cost-effective method for multispecies mammal monitoring at scale. But “cost-effective” does not mean “cheap.” A serious survey costs real money, and the things that make it expensive are rarely the cameras themselves.

The good news: costs are dropping. A Browning unit that cost $250 in 2015 is $160 today and performs better. MegaDetector eliminates 80% of classification labor for free. Cloud infrastructure costs decline every year.

The bad news: battery costs, equipment loss, and data processing labor still catch first-time PIs off guard. The spreadsheet looks fine until you are in the field, swapping out 400 AA batteries in the rain, realizing you forgot to budget for the smart charger that prevents your rechargeables from dying after 3 cycles.

Budget for the whole thing. Not just the cameras.

Field Log is a field-first mobile platform built by The Field Company. Offline-first. Team sync. Structured forms and rapid logging. Free to start.{" "} Get started at fieldlog.thefieldco.com. For more on designing monitoring that actually works, see{" "} How to Set Up a Biodiversity Monitoring Protocol.

Hardware prices from TrailCamPro, June 2026. Equipment loss rates: Meek et al. (2019) "Camera trap theft and vandalism" Remote Sensing in Ecology and Conservation. Survey design: TEAM Network protocol, Rovero & Zimmermann (2016). Classifier: MegaDetector (Microsoft AI for Good Lab). Data hosting: Wildlife Insights. All prices USD unless noted. Import duties and regional pricing vary.

The Field Co

Open-Source Conservation Technology

The 32 Companies Driving Half of Global Emissions

Sun, 19 Jul 2026 00:00:00 GMT

import { Image } from "astro:assets"; import BlogImage from "../../components/blog/BlogImage.astro"; import CarbonMajo2 from "../../assets/blog/carbon-majors/19419368_janusz_walczak.jpg"; import CarbonMajo1 from "../../assets/blog/carbon-majors/10353038_elle-photoart.jpg"; import industrialPlant from "../../assets/blog/carbon-majors/9861246_sami__aksu.jpg"; import industrialEmissions from "../../assets/blog/carbon-majors/6675078_chris_leboutillier.jpg"; import { CompaniesChart } from "@components/blog/charts";

Thirty-two companies. That is the number of corporate entities responsible for more than half of the carbon dioxide emitted from the world's fossil fuels and cement in 2024. Not nations. Not industries. Companies. Their names are on stock exchanges, sovereign wealth funds, and quarterly reports. Their emissions have names, too: 34.7 billion tonnes of CO₂ equivalent, a figure that rose 0.8% from the year before.

The concentration is accelerating. Five years ago it took 38 companies to account for half of global fossil CO₂. Now it takes 32. As smaller producers have held steady or declined, the largest emitters have continued to grow — through acquisitions, through expanded production, through what their CEOs call “the demand decade.”

This is not opinion. It is the annual update to the Carbon Majors database, first built by researcher Richard Heede in 2013 and now maintained by the think tank InfluenceMap. It traces production data from 178 of the world’s largest oil, gas, coal, and cement producers back to 1854. The methodology has been scrutinised in courtrooms, cited in international legal opinions, and used to attribute heatwaves and economic losses to specific companies. Every number below is real. Every source is cited.

The Carbon Majors

The Carbon Majors database quantifies something that was largely invisible until Richard Heede published his first results in 2013: the direct link between specific corporate entities and the cumulative CO₂ that is heating the planet. Before Heede’s work, emissions were attributed to countries or sectors. After it, a new question became answerable: Which companies did this?

The database now covers 178 entities — 100 investor-owned companies, 72 state-owned companies, and 6 nation-state producers — tracking 1.44 trillion tonnes of CO₂ equivalent. That represents 70% of all fossil fuel and cement CO₂ emitted since the Industrial Revolution began in 1750. In the most recent year alone, 2024, the database traced 34.7 gigatonnes of CO₂ equivalent to 166 active entities. That is 80% of global fossil fuel and cement CO₂ for the year.

The finding that has made headlines globally: 32 companies produced over half of global fossil CO₂ in 2024. The figure was 36 companies in 2023, 38 five years ago, and consistently above 40 between 2005 and 2013. Emissions are concentrating. The largest producers are getting larger.

Chris Leboutillier on Pexels`} />

The Top 10

Here is a fact that should stop you: the ten largest emitters in 2024 are responsible for 27.6% of all global fossil CO₂. All ten are fully or majority state-owned. Not one investor-owned company appears until position 13.

### 2024 — Top Emitters by Company

Rank	Entity	Emissions (MtCO₂e)	% Global CO₂	Type	Country	Primary
1	Saudi Aramco	1,786	4.28%	State	Saudi Arabia	Oil & Gas
2	Coal India	1,684	3.92%	State	India	Coal
3	CHN Energy	1,679	3.91%	State	China	Coal
4	National Iranian Oil	1,387	3.13%	State	Iran	Oil & Gas
5	Gazprom	1,293	2.76%	State	Russia	Oil & Gas
6	Jinneng Group	1,129	2.63%	State	China	Coal
7	China (Cement)	950	2.46%	Nation	China	Cement
8	Rosneft	763	1.79%	State	Russia	Oil & Gas
9	CNPC	750	1.70%	State	China	Oil & Gas
10	Shandong Energy	750	1.74%	State	China	Coal
* Total emissions include fugitive methane (CO₂e). % of global column uses CO₂ only. Source: InfluenceMap, Carbon Majors 2024 Data Update, January 2026.

Some translations of those numbers:

Saudi Aramco alone emitted 1.79 billion tonnes of CO ₂ equivalent. If Aramco were a country, it would be the world’s fifth-largest carbon polluter — behind China, the United States, India, and Russia. Most of those emissions come from oil that Aramco extracts, exports, and someone else burns. That someone else could be you.

ExxonMobil, the largest investor-owned emitter at position 13 (677 MtCO₂e), would rank as the ninth-largest country polluter — ahead of South Korea. In January 2026, CEO Darren Woods told a White House meeting: “We’re in a depletion business for a product that is in great demand and will be in demand for many, many, many decades to come.” The American Petroleum Institute’s CEO called the next ten years “the Demand Decade.”

Eight Chinese companies appear in the top 20. Six of them are coal producers. Chinese companies alone account for 22.8% of global fossil CO₂ tracked by the database. Asia as a whole: 31.9%, with 58% of Asian companies increasing emissions year-over-year.

Sami Aksu on Pexels`} />

Who Built the Problem

The annual rankings tell one story. The historical rankings tell another — and it is the one that matters for warming. CO₂{" "} accumulates. A tonne emitted in 1965 is still trapping heat today. The Carbon Majors database traces emissions back to 1854, and the cumulative totals reveal which entities built the bulk of the atmospheric carbon that now drives 1.55°C of global heating.

### Cumulative Emissions, 1854–2024 — Top 10

Rank	Entity	Total (MtCO₂e)	% Global CO₂	Type
1	Former Soviet Union (1900–1991)	135,113	6.54%	Nation State
2	China Coal (1945–2004)	104,888	5.10%	Nation State
3	Saudi Aramco	72,457	3.66%	State
4	Chevron	62,503	3.08%	Investor
5	ExxonMobil	57,458	2.79%	Investor
6	Gazprom	53,116	2.33%	State
7	National Iranian Oil	45,826	2.25%	State
8	BP	43,231	2.13%	Investor
9	Shell	41,517	2.02%	Investor
10	Coal India	35,221	1.71%	State

Historically, investor-owned companies account for 24.1% of cumulative global fossil CO₂. State-owned companies account for 31.0%. Nation-state producers account for 14.8%. But the proportions are shifting rapidly: in 2024, state-owned companies were linked to{" "} 53.4% of global fossil CO₂, compared with 23.7% for investor-owned companies. Most state-owned companies increased emissions year-over-year. A majority of investor-owned companies reduced theirs. The centre of gravity is moving east, and it is moving into state hands.

Elle-photoart on Pexels`} />

The Production Gap

There is a concept in climate policy called the production gap — the difference between what governments plan to extract and what the planet can absorb. The 2025 Production Gap Report, published by the UN Environment Programme and leading research institutes, found that ten years after the Paris Agreement, governments collectively plan to produce more than 120% more fossil fuels in 2030 than would be consistent with limiting warming to 1.5°C.

The breakdown is stark:

### Planned Production vs 1.5°C Pathway (2030)

Fuel	Planned vs 1.5°C Limit
Coal	500% above limit
Oil	31% above limit
Gas	92% above limit

At COP30 in Belém, Brazil, in November 2025, an EU-led coalition of more than 80 countries proposed a formal roadmap to transition away from fossil fuels. It was blocked. The opposition came from Saudi Arabia, Russia, China, India, Iran, the United Arab Emirates, Algeria, Iraq, and Qatar. Seventeen of the top 20 Carbon Majors entities in 2024 are controlled by countries that opposed the phaseout plan.

Meanwhile, the American Petroleum Institute’s CEO called this “the Demand Decade.” Chevron has filed to acquire Hess Corp. ExxonMobil completed its acquisition of Pioneer Natural Resources. ConocoPhillips absorbed Marathon Oil. Core Natural Resources was formed in 2025 from the merger of CONSOL Energy and Arch Resources. The biggest emitters are not just maintaining output. They are consolidating.

Janusz Walczak on Pexels`} />

Who Owns These Companies

The line between “state-owned” and “investor-owned” is not academic. It defines who profits, who decides, and who can be held to account.

State-owned entities dominate the 2024 rankings: 16 of the top 20. Their combined footprint in 2024 was 53.4% of global fossil CO₂. These companies answer not to shareholders but to governments — many of which are actively blocking international climate agreements. Saudi Aramco, the single largest emitter, funds the Saudi state budget. Gazprom funds the Russian one. Coal India is the engine of India’s power grid. Their emissions are not accidents of the market. They are deliberate, state-directed production decisions.

Investor-owned companies — ExxonMobil, Chevron, Shell, BP, ConocoPhillips — account for a smaller but still enormous share: 23.7% in 2024. Their shareholders include pension funds, university endowments, index funds owned by ordinary people. InfluenceMap’s parallel LobbyMap database rates the policy engagement of these companies on an A+ to F scale. Chevron scores D− with 33% engagement intensity. ExxonMobil scores D with 39%. Shell scores C−. BP scores C−. None of the assessed top 20 companies scores above C−. These are not passive observers of climate policy. They are active, well-funded opponents of it.

The Legal Tide

Something has shifted in the last five years. The Carbon Majors database, originally a research project by one man compiling production records, is now a central piece of evidence in courtrooms on three continents.

In 2021, a Dutch district court ordered Shell to reduce its global emissions by 45% by 2030, including the emissions from the products it sells — the first time a court had held a corporation to the Paris Agreement’s temperature targets. Shell appealed. In November 2024, the appeals court overturned the specific reduction target, ruling there was no “social standard of care” for a precise percentage. But the court affirmed the principle: companies do have a duty to contribute to combating climate change. The door remains open.

In Germany, a Peruvian farmer named Saúl Luciano Lliuya sued the energy company RWE, arguing that its historical emissions contributed to the melting of a glacier threatening his community. In May 2025, the Higher Regional Court of Hamm dismissed the case — but in doing so, it accepted the scientific evidence that traced specific emissions to RWE through the Carbon Majors database. For the first time, a German court recognised that companies can be held liable for climate-related harm. The case is being appealed further.

In the United States, two states have passed Climate Superfund laws — New York (seeking up to $75 billion) and Vermont (approximately $2.5 billion) — requiring major fossil fuel companies to pay for climate adaptation based on their historical share of emissions. More than a dozen additional states are advancing similar legislation. The bills explicitly reference Richard Heede’s methodology. California’s analysis suggests recoverable damages could reach hundreds of billions of dollars.

In 2024, the Inter-American Court of Human Rights issued an advisory opinion citing Carbon Majors as evidence that a small group of corporate producers are disproportionately responsible for global emissions, concluding that states have a duty to regulate corporate climate harms on human rights grounds. At least 226 new climate cases were filed globally in 2024, and 20% now target corporations — a category where plaintiffs have a higher success rate than against governments.

What This Means

It is easy to read this data as an indictment of 32 specific companies. That is the wrong interpretation. It is not wrong factually — the data is unambiguous — but it misses the structure that produces the data.

These 32 companies exist because demand exists. Saudi Aramco pumps 12 million barrels a day because the world burns 100 million barrels a day. Coal India extracts 700 million tonnes a year because India’s grid needs it, because hundreds of millions of people need electricity, because coal is the cheapest and most available source. The emissions are not a conspiracy. They are a system — an energy system that the entire global economy was built on, that lifted billions out of poverty, and that is now destroying the conditions that made that progress possible.

The Carbon Majors database does something useful: it names names. It quantifies responsibility. It provides the evidence base for legal accountability, for policy targeting, for financial risk assessment. But the most important number in this report is not the ranking. It is the trend. Emissions are rising. Concentration is increasing. The companies with the most to lose from the energy transition are growing fastest and lobbying hardest against it. This is not inertia. It is acceleration in the wrong direction.

Read the full letter: A Letter to Humanity →

Sources:

Carbon Majors 2024 Data Update — InfluenceMap
The Guardian — Half of world's CO\u2082 emissions come from 32 fossil fuel firms
Inside Climate News
Production Gap Report 2025
LSE Grantham Institute — Climate Litigation 2025
Heede, R. (2014) Tracing anthropogenic CO\u2082 to fossil fuel producers.{" "} Climatic Change
Quilcaille et al. (2025) Attributing heatwaves to fossil fuel producers.{" "} Nature
Callahan & Mankin (2025) Economic losses attributable to fossil fuel producers. Nature
Carbon Majors Methodology
InfluenceMap LobbyMap

The Field Co

Open-Source Conservation Technology

AlphaEarth Foundations: Google DeepMind's Universal Model for Mapping the Earth

Sun, 14 Jun 2026 00:00:00 GMT

import BlogImage from "../../components/blog/BlogImage.astro"; import alphaearthCollage from "../../assets/blog/alphaearth-foundations/alphaearth-collage.png"; import alphaearthRgbViz from "../../assets/blog/alphaearth-foundations/alphaearth-rgb-viz.png"; import googleSatelliteEmbeddingSample from "../../assets/blog/alphaearth-foundations/google-satellite-embedding-sample.png"; import earthView1 from "../../assets/blog/alphaearth-foundations/30596223_zelch_csaba.jpg"; import earthView2 from "../../assets/blog/alphaearth-foundations/30596234_zelch_csaba.jpg";

In July 2025, Google DeepMind introduced{" "} AlphaEarth Foundations, a geospatial foundation model designed to turn huge volumes of Earth observation data into a compact, reusable representation of the planet. Google described it as functioning like a “virtual satellite”: not because it replaces real satellites, but because it combines many different satellite and environmental data sources into one consistent digital view of land and shallow coastal waters.

The practical product of this model is the{" "} Google Satellite Embedding dataset in Google Earth Engine. Instead of giving users raw red, green, blue, infrared, radar, terrain, climate, and other bands separately, the dataset gives every 10 m pixel a learned 64-dimensional embedding for each year. That embedding acts like a compact fingerprint of the surface conditions at that place and time.

This matters because much of environmental science is bottlenecked by labels. Researchers often have a few hundred field points, camera-trap locations, crop samples, habitat polygons, or protected-area observations, but not enough labelled data to train a large satellite model from scratch. AlphaEarth Foundations shifts the workflow: Google has already run the expensive feature-learning step, and scientists can use the embeddings directly for classification, regression, similarity search, and change detection.

Source note: This post was prepared from Google DeepMind’s AlphaEarth Foundations announcement, the Google Satellite Embedding V1 Earth Engine Data Catalog page, the AlphaEarth Foundations preprint, Google Earth and Earth Engine technical posts and tutorials, and Earth Engine noncommercial/commercial access guidance. Links are listed throughout and again in the source section.

The Quick Answer

Yes: the “universal model of the Earth” people are talking about is{" "} AlphaEarth Foundations. A better technical description is:{" "} a geospatial foundation model that produces reusable satellite embeddings for the Earth’s surface .

Question	Answer
Who built it?	Google DeepMind, with the resulting dataset released through Google Earth Engine.
When was it announced?	30 July 2025.
What is released for users?	The `GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL`{" "} ImageCollection in Earth Engine, plus a Google Cloud Storage copy.
What does each pixel contain?	A 64-dimensional embedding vector summarising a year of surface conditions at 10 m resolution.
What years are available?	Annual layers from 2017 through 2024 in the Earth Engine Data Catalog. Google states that annual production is intended to continue, subject to input data availability.
What is it useful for?	Land-cover mapping, habitat mapping, crop classification, biomass modelling, change detection, similarity search, wildfire recovery, coastal mapping, and other sparse-label science tasks.
What is it not?	It is not a full climate simulator, not a weather model, not a labelled map by itself, and not a replacement for field validation.

What the Embeddings Look Like

A 64-dimensional embedding cannot be seen directly. To make it visible, Google and other users often assign three embedding dimensions to red, green, and blue. The colours are not natural colour. They are a way of visualising patterns in the learned representation.

Google DeepMind announcement`} /> Google DeepMind`} /> Google for Developers`} />

What Is AlphaEarth Foundations?

AlphaEarth Foundations is a{" "} geospatial embedding field model. That phrase sounds abstract, but the idea is straightforward:

The model learns how places on Earth behave and appear across many different sources of Earth observation data. For a given location and time period, it produces a short numeric vector that captures useful information about that place. This vector can then be used as an input to simpler models and geospatial workflows.

Traditional satellite analysis often starts with a hand-built stack of inputs: Sentinel-2 bands, Landsat bands, NDVI, radar backscatter, elevation, slope, rainfall, seasonal composites, cloud masks, gap filling, and many other transformations. AlphaEarth Foundations tries to compress much of that messy multi-source information into a single learned representation.

Google DeepMind says the model integrates large amounts of Earth observation data into a unified representation that can help researchers analyse food security, deforestation, urban expansion, water resources, and environmental change. The Earth Engine Data Catalog describes the released Satellite Embedding dataset as a global, analysis-ready collection of 10 m geospatial embeddings, where each pixel contains a 64-dimensional vector derived from multiple Earth observation data sources.

Official entry points: Google DeepMind announcement , Earth Engine Data Catalog page , and{" "} AlphaEarth Foundations preprint .

Zelch Csaba on Pexels`} />

What Is a Satellite Embedding?

An embedding is a learned numeric representation. In language models, words or sentences are turned into vectors so that similar meanings sit near each other in vector space. In AlphaEarth Foundations, places on Earth are turned into vectors so that similar surface conditions sit near each other in a geospatial embedding space.

In the Google Satellite Embedding dataset, every 10 m pixel has 64 bands named A00 through A63. These are not normal spectral bands. A07, for example, does not mean “red reflectance” or “moisture”. It is one axis of a learned 64-dimensional coordinate.

Conventional satellite composite	AlphaEarth satellite embedding
Bands correspond to physical measurements such as red, near-infrared, shortwave infrared, radar backscatter, or thermal emission.	Bands are learned dimensions in a 64-dimensional embedding space.
Users often need cloud masking, seasonal compositing, index calculation, radar filtering, and gap filling.	The dataset is already prepared as annual, analysis-ready feature layers.
Good for direct interpretation: NDVI, water indices, burn indices, reflectance values.	Good for machine learning and similarity: clustering, classification, regression, change detection, and vector search.
The user decides which features to engineer.	The foundation model has already learned a compact feature representation from many inputs.

The important warning is that{" "} embedding dimensions are not independently interpretable . You should use all 64 dimensions together for modelling unless you have a strong reason not to. The value of the dataset comes from the geometry of the full embedding space, not from naming each band as if it were a physical measurement.

What Data Went Into the Model?

AlphaEarth Foundations was trained to integrate multiple categories of Earth observation data. The paper and Google materials describe inputs including optical satellite imagery, radar, LiDAR, elevation, climate reanalysis, gravity/mass data, labelled land-cover or land-use products, and text sources.

Input type	Examples mentioned in Google materials	Why it helps
Optical imagery	Sentinel-2, Landsat 8/9	Vegetation, bare ground, water, burned areas, built surfaces, crop cycles.
Radar	Sentinel-1, ALOS PALSAR-2	Cloud-penetrating structure, roughness, moisture-related signals, forest and flood information.
LiDAR / structure	GEDI canopy and surface metrics	Vertical vegetation structure and canopy height information.
Terrain	GLO-30 elevation	Topography, slope-related landscape context, hydrological context.
Climate / environment	ERA5-Land, GRACE	Seasonal and environmental context beyond what a single image can see.
Annotated sources	NLCD, later updates including USDA Cropland Data Layer targets	Helps the representation align with human-recognisable land categories.
Text sources	Geo-temporally located text labels / Wikipedia-style sources described in the paper	Adds semantic context about places and categories.

Google’s technical post says the model was trained on over 3 billion individual image frames sampled from more than 5 million locations globally. The Data Catalog also notes that the currently hosted embeddings were generated with version 2.1 of the model, including improvements over the version evaluated in the preprint.

Why This Matters for Science

AlphaEarth Foundations is important because it changes the starting point for remote-sensing science. Instead of every team building a custom stack of raw satellite features, many teams can start with the same global feature representation and spend more time on field data, validation, and scientific interpretation.

1. It lowers the label barrier

Environmental labels are expensive. A field team may only be able to collect a few hundred verified points for wetlands, mangroves, invasive plants, crop types, burn severity, or habitat condition. The AlphaEarth paper reports strong performance in sparse-label settings, comparing AEF to several designed and learned featurization approaches across land cover, land use, crop, species, evapotranspiration, and surface-emissivity tasks.

2. It makes complex satellite data easier to use

A good remote-sensing workflow normally requires many preprocessing decisions: cloud masking, compositing, seasonal windows, spectral indices, radar speckle handling, topographic correction, and sensor harmonisation. Satellite embeddings do not remove the need for expertise, but they give users a ready-made feature layer that can be plugged into Earth Engine classifiers and reducers.

3. It supports comparison through time

The annual embeddings are designed to be consistent across years. That means a user can compare a pixel’s 2019 vector with its 2024 vector and look for meaningful change in the embedding space. This is useful for urban expansion, agricultural change, reservoir changes, wildfire recovery, deforestation, wetland degradation, and habitat restoration monitoring.

4. It makes similarity search possible at planetary scale

With embeddings, you can ask a new kind of question: “find more places like this.” A conservation team could identify sites environmentally similar to known breeding habitat. A restoration team could find areas similar to successful restoration sites. A renewable-energy team could search for surfaces similar to existing solar farms. A disease ecology team could search for environmental analogues to known risk zones.

5. It makes Earth Engine feel more like modern AI infrastructure

Earth Engine already made planetary-scale remote sensing accessible. AlphaEarth adds a foundation-model feature layer inside that environment. This means users can run modern AI-style workflows — classification, vector search, clustering, regression, change detection — without running the deep model themselves.

Zelch Csaba on Pexels`} />

What It Is Not

The phrase “universal model of the Earth” is exciting, but it can also mislead. AlphaEarth Foundations is powerful, but it is not magic.

Misunderstanding	Better explanation
“It is a complete digital twin of Earth.”	No. It is a learned geospatial representation of terrestrial land surfaces and shallow waters, useful for mapping and monitoring tasks.
“It predicts weather or climate by itself.”	No. It is not a weather model like WeatherNext or a climate simulator. It may be useful as a feature layer in environmental modelling.
“It labels every pixel automatically.”	No. It provides embeddings. You still need labels, rules, clustering, classifiers, regression models, or similarity thresholds.
“The 64 bands are directly interpretable.”	No. The bands are learned vector dimensions. Use the full vector and validate results.
“It removes the need for fieldwork.”	No. Field data is still essential for training, validation, uncertainty checks, and ecological meaning.
“It works equally well everywhere.”	No. Sensor coverage, training data, biome differences, local land-use patterns, clouds, and artefacts can still matter. Always validate locally.

How to Use AlphaEarth in Google Earth Engine

The easiest way to use AlphaEarth Foundations is through the Satellite Embedding ImageCollection in Earth Engine:

```js ee.ImageCollection("GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL") ```

You do not download the model or run inference yourself. Google has already produced the annual embeddings. Your task is to choose the year, region, labels, and analysis method.

Step 1: Open the dataset

Start in the Earth Engine Code Editor and load one year of embeddings for your area of interest.

```js // Example: load the 2024 Google Satellite Embedding layer. var aoi = geometry; // Draw or import your study area. var embeddings = ee.ImageCollection('GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL') .filterDate('2024-01-01', '2025-01-01') .filterBounds(aoi); // The collection is tiled, so mosaic the tiles covering your area. var image2024 = embeddings.mosaic().clip(aoi); // Visualise three embedding dimensions as false colour. Map.centerObject(aoi, 10); Map.addLayer( image2024.select(['A00', 'A16', 'A32']), {min: -0.3, max: 0.3}, 'AlphaEarth embedding RGB, 2024' ); ```

The colours are only a visualisation. Do not interpret the red, green, and blue channels as normal satellite colour.

Step 2: Use all 64 bands for analysis

For classification, regression, clustering, or change detection, select all 64 embedding bands.

```js var bandNames = ee.List.sequence(0, 63).map(function(i) { return ee.String('A').cat(ee.Number(i).format('%02d')); }); var features2024 = image2024.select(bandNames); ```

Step 3: Choose a workflow

Goal	Method	What you need
Map habitat or land-cover classes	Supervised classification	Training points or polygons with class labels.
Estimate a continuous variable	Regression	Field measurements such as biomass, canopy height, yield, soil properties, or water quality.
Find unknown patterns	Unsupervised clustering	A region of interest and a plan to interpret/label clusters afterward.
Find more places like one known place	Similarity search / dot product	One or more reference points or polygons.
Detect change between years	Compare vectors across years	Two annual embedding layers for the same area.

Workflow 1: Supervised Classification

Use this when you have known examples: mangrove vs non-mangrove, intact habitat vs degraded habitat, crop type, invasive vegetation, wetland class, or land-use type.

```js // trainingPoints must contain a property called "class". // Example classes: 0 = non-mangrove, 1 = mangrove. var samples = features2024.sampleRegions({ collection: trainingPoints, properties: ['class'], scale: 10, geometries: true }); var classifier = ee.Classifier.smileRandomForest({numberOfTrees: 100}) .train({ features: samples, classProperty: 'class', inputProperties: bandNames }); var classified = features2024.classify(classifier); Map.addLayer( classified, {min: 0, max: 1, palette: ['gray', 'green']}, 'Classified map' ); ```

For a scientific result, do not stop at a pretty map. Split your labels into training and validation data, calculate accuracy metrics, inspect errors, and verify uncertain areas with field knowledge or high-resolution imagery.

Google’s community tutorial demonstrates a supervised classification workflow for mapping mangroves with Satellite Embeddings.

Workflow 2: Similarity Search

Similarity search is one of the most interesting uses of AlphaEarth. Instead of training a full classifier, you choose a reference location and ask: where else has a similar embedding?

```js // Choose a reference point inside the feature you care about. var referencePoint = ee.Geometry.Point([31.58, -24.99]); // Example only. // Extract the reference vector. var reference = features2024.reduceRegion({ reducer: ee.Reducer.first(), geometry: referencePoint, scale: 10, maxPixels: 1e6 }); // Turn the dictionary of reference values into a constant image. var referenceVector = ee.Image.constant(reference.values(bandNames)) .rename(bandNames); // Because embeddings are unit-length, dot product is cosine similarity. var similarity = features2024 .multiply(referenceVector) .reduce(ee.Reducer.sum()) .rename('similarity'); Map.addLayer( similarity, {min: 0.6, max: 1.0, palette: ['black', 'yellow', 'red']}, 'Similarity to reference point' ); ```

This can be powerful for conservation prospecting: find areas similar to known habitat, known restoration success sites, known invasive-plant patches, known erosion areas, or known human-impact patterns. But similarity is not proof. Use it to generate candidates, then validate them.

Workflow 3: Change Detection

Because the embedding space is designed to be comparable across years, you can compare the same pixel in two annual layers. A high dot product means the embedding stayed similar. A lower dot product suggests change.

```js var image2020 = ee.ImageCollection('GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL') .filterDate('2020-01-01', '2021-01-01') .filterBounds(aoi) .mosaic() .clip(aoi) .select(bandNames); var image2024 = ee.ImageCollection('GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL') .filterDate('2024-01-01', '2025-01-01') .filterBounds(aoi) .mosaic() .clip(aoi) .select(bandNames); var similarity2020To2024 = image2020 .multiply(image2024) .reduce(ee.Reducer.sum()) .rename('cosine_similarity'); var changeScore = ee.Image(1).subtract(similarity2020To2024) .rename('embedding_change'); Map.addLayer( changeScore, {min: 0, max: 0.4, palette: ['white', 'orange', 'red']}, 'Embedding change, 2020 to 2024' ); ```

This is useful as a screening layer. Once you find areas of high change, compare against known events: fires, floods, land clearing, crop rotation, construction, restoration, water-level change, or sensor artefacts.

Workflow 4: Unsupervised Clustering

If you do not have labels, clustering can help reveal natural groupings in the landscape. This is not a substitute for classification. It is an exploration tool.

```js var training = features2024.sample({ region: aoi, scale: 10, numPixels: 5000, seed: 42 }); var clusterer = ee.Clusterer.wekaKMeans(12).train(training); var clusters = features2024.cluster(clusterer); Map.addLayer( clusters.randomVisualizer(), {}, 'Embedding clusters' ); ```

After clustering, inspect each cluster with field notes, high-resolution imagery, existing maps, and expert knowledge. A cluster might represent a habitat type, a crop stage, a soil/terrain pattern, a disturbance pattern, or a mixture of several things.

Scientific and Conservation Use Cases

Field	Possible use	Why embeddings help
Biodiversity and ecology	Habitat mapping, ecological condition mapping, species distribution covariates, protected-area change.	Combines vegetation, terrain, water, seasonality, and land-use context in one feature layer.
Conservation operations	Find areas similar to known breeding, denning, nesting, or migration-support habitat.	Similarity search can identify candidate areas for ranger surveys or restoration planning.
Forestry	Forest type classification, degradation screening, fire recovery, plantation mapping.	Radar, optical, LiDAR, topography, and annual temporal signals can all contribute to the representation.
Agriculture	Crop type mapping, fallow detection, irrigation pattern detection, phenology grouping.	Annual embeddings can capture within-year crop cycles rather than a single date.
Water resources	Reservoir change, wetland dynamics, shallow coastal and inland water monitoring.	Temporal consistency allows comparison between years and across similar water environments.
Urban science	Urban expansion, informal settlement mapping, impervious-surface patterns, infrastructure growth.	Embeddings can include local context, helping distinguish visually similar surfaces in different settings.
Climate risk	Fire scars, drought impacts, land degradation, restoration trajectories.	Embeddings provide a compact way to compare pre- and post-event conditions.

Good Scientific Practice

AlphaEarth Foundations should make workflows faster, but it does not remove the need for scientific discipline. Treat the embeddings as powerful input features, not as truth.

Use proper validation

Always keep an independent validation set. Avoid training and testing on points that are too close together, especially in spatial data, because neighbouring pixels are often similar. Use spatial cross-validation where possible.

Check local bias

A model that performs well globally may still fail locally. Validate in your biome, country, season, sensor conditions, land-use system, and species/habitat context.

Compare against simpler baselines

Do not assume embeddings are always better. Compare against a standard Sentinel-2 composite, Landsat features, NDVI/EVI, radar bands, elevation, or an existing land-cover product. The embedding result should earn trust by outperforming or simplifying alternatives in your use case.

Document training data

Record who labelled the data, when it was collected, how accurate it is, what classes mean, and how ambiguous samples were handled. A strong embedding layer cannot fix weak labels.

Protect sensitive locations

Conservation teams should be careful when mapping rare species, nesting sites, den sites, poaching-risk areas, and sensitive habitats. Publishing exact maps can create risk. Consider aggregation, masking, delayed release, or restricted access.

Limitations and Caveats

Limitation	What to do about it
Black-box representation	Use embeddings as features and validate outputs. Do not over-explain individual bands.
Potential artefacts	Inspect outputs visually and compare against raw imagery, existing products, and field knowledge.
Annual summary may miss short events	For floods, fire timing, crop harvest windows, or short-lived disturbances, combine embeddings with event-specific imagery.
Spatial resolution is 10 m	Small features below pixel scale may be mixed or missed. Use higher-resolution imagery where needed.
Not a labelled product	Train, cluster, or compare embeddings; do not treat the embedding as a finished map.
Coverage constraints	The Data Catalog notes coverage of land and shallow waters, with limitations at the poles due to satellite orbits and instrument coverage.
Access and licensing	Earth Engine remains available at no additional cost for eligible noncommercial research and education, but operational/commercial use may require Google Cloud commercial access.

Earth Engine vs AlphaEarth: The Difference

It is useful to separate three things that are often mentioned together:

Term	What it is	How it connects
Google Earth Engine	A cloud geospatial data and compute platform.	You use it to access, analyse, visualise, and export geospatial datasets.
AlphaEarth Foundations	A Google DeepMind geospatial foundation model.	It produces learned embeddings from many Earth observation inputs.
Google Satellite Embedding dataset	The released annual embedding ImageCollection in Earth Engine.	This is what most scientists and developers will actually use.
Google Earth AI	A broader Google effort around AI-powered geospatial products and workflows.	AlphaEarth and Satellite Embeddings are part of this direction, but not the whole story.

A Simple Student Project Idea

A good student project would be:{" "} map and validate vegetation change around a protected area between 2017 and 2024 using AlphaEarth embeddings and field/desktop validation .

Choose a protected area or reserve boundary.
Load the 2017 and 2024 Satellite Embedding layers.
Compute an embedding change score using vector similarity.
Inspect high-change areas against Sentinel-2 or high-resolution basemap imagery.
Classify changes into likely causes: fire, clearing, crop change, urban expansion, water variation, restoration, or artefact.
Validate a sample of points manually or with field ranger input.
Produce a map, uncertainty notes, and recommendations for where field teams should inspect next.

This kind of project is realistic because it does not require training a deep learning model from scratch. It uses the foundation model’s embeddings as a scientific feature layer and focuses the student’s effort on interpretation and validation.

Key Sources and Further Reading

Wildlife Conservation in Cape Town: A Field Guide to the City's Living Systems

Sun, 14 Jun 2026 00:00:00 GMT

import BlogImage from "../../components/blog/BlogImage.astro"; import penguinColony from "../../assets/blog/cape-town-wildlife-conservation/32754694_dane_damons.jpg"; import fynbosMountains from "../../assets/blog/cape-town-wildlife-conservation/32589602_magda_ehlers.jpg";

Cape Town is not just a city next to nature. It is a city built inside one of the most biologically important landscapes on Earth. The mountains, wetlands, lowland fynbos fragments, rocky shores, kelp forests, beaches, estuaries and urban edges of the metro all sit inside the Cape Floristic Region, a global biodiversity hotspot.

This makes conservation in Cape Town unusual. It is not only about fencing off wilderness somewhere far away. It is about keeping critically endangered habitats alive between roads, housing estates, railway lines, schools, vineyards, beaches, harbours and tourist sites. The work is practical, political and often messy: clearing alien trees, managing fire, protecting penguins from cars and dogs, keeping baboons out of kitchens, rescuing oiled seabirds, helping toads cross roads, reducing rat poison, restoring wetlands and deciding where development must stop.

Source note: This article uses City of Cape Town planning documents, SANParks reports, local NGO materials, and peer-reviewed papers including Rebelo et al. (2011), Holmes et al. (2012), Forsyth & van Wilgen (2008), van Wilgen et al. (2012), Hoffman & O’Riain (2012), Engelbrecht et al. (2017), Serieys et al. (2019), Kyriazis et al. (2024), McInnes et al. (2024), and Sherley et al. (2024). Links are listed in the reference section.

Why Cape Town Is a Conservation Hotspot

The easiest mistake is to think of Cape Town conservation as “Table Mountain plus penguins.” Those are the famous symbols, but the deeper story is the city’s lowland ecosystems. Much of Cape Town’s most threatened biodiversity is not on the mountain. It is on the Cape Flats, the West Coast lowlands, the wetlands, the sand fynbos remnants and the old seasonal vleis now surrounded by urban growth.

Rebelo and colleagues described metropolitan Cape Town as a city inside a biodiversity hotspot, estimating roughly 3,250 plant species within the city, including hundreds of threatened species and several already extinct locally. Their paper is still one of the best scientific starting points for understanding why ordinary-looking open ground in Cape Town can be globally important habitat.

The City’s current planning backbone is the{" "} Cape Town BioNet, a fine-scale biodiversity network that identifies Critical Biodiversity Areas, Ecological Support Areas and other priority sites needed to keep native biodiversity and ecological processes functioning across the municipality. The 2025 Cape Town Biodiversity Spatial Plan formalises that map as a planning tool so biodiversity priorities are visible before development decisions are made.

Conservation layer	What it protects	Why it matters
Table Mountain National Park	Mountain fynbos, Afrotemperate forest pockets, coastline, Cape Point and marine protected areas.	Large continuous habitat, world heritage value, recreation, tourism and fire-adapted fynbos management.
City nature reserves	Lowland fynbos, renosterveld, wetlands, estuaries, coastal dunes and urban biodiversity fragments.	Protects habitats that are often more threatened than the mountain itself.
BioNet / Biodiversity Spatial Plan	The minimum spatial network needed to conserve biodiversity and ecological function.	Turns biodiversity into a land-use planning layer instead of an afterthought.
NGO and citizen-science projects	Penguins, caracals, toads, seabirds, turtles, sharks, baboons, plants and local reserves.	Adds monitoring, rescue, public reporting, education and practical field labour.

1. The City’s Reserve Network and BioNet

Cape Town’s municipal conservation work is centred on the City’s nature reserves, conservation areas and the BioNet planning system. This is where the city tries to hold onto pieces of habitats that have been heavily reduced by agriculture, housing, roads and industry.

Key sites include Blaauwberg Nature Reserve,{" "} Table Bay Nature Reserve,{" "} False Bay Nature Reserve, Rondevlei,{" "} Rietvlei, Tygerberg Nature Reserve,{" "} Edith Stephens Wetland Park,{" "} Rondebosch Common,{" "} Kenilworth Racecourse Conservation Area,{" "} Milnerton Racecourse,{" "} Witzands Aquifer Conservation Area,{" "} Bracken Nature Reserve, Wolfgat and smaller fragments that often carry rare plant populations.

Holmes et al. (2012) asked a blunt question: can Cape Town’s unique biodiversity be saved while development pressure continues? Their answer was conditional. It can only happen if biodiversity planning is embedded early in land-use decisions, if irreplaceable fragments are secured, and if restoration and management continue after land has been set aside.

The point: Cape Town’s conservation problem is not simply “protect more land.” It is “protect the right fragments, keep them connected where possible, and keep managing them forever.” In fynbos and wetland systems, a reserve that is not burned, cleared, restored and monitored can lose its biodiversity even if it remains legally protected.

2. Fynbos, Fire and Invasive Alien Plant Clearing

Fynbos needs fire, but Cape Town makes fire dangerous. The city’s conservation managers have to balance ecological fire cycles with houses, roads, hikers, heritage sites, power lines and dense invasive alien trees. In the Cape Peninsula, fire is both a regeneration process and a public-safety risk.

Forsyth & van Wilgen (2008) analysed the recent fire history of Table Mountain National Park and showed why fire management in an urban national park is so complex. Van Wilgen and colleagues (2012) later described Table Mountain National Park as a case where ecological evidence, public perception and trade-offs collide: fynbos conservation needs alien clearing and appropriate fire, while residents often experience fire and tree removal as threats to recreation, views and safety.

The biggest practical intervention is still{" "} alien plant clearing. Pines, wattles, hakeas, gums and other alien trees can outcompete fynbos, alter water flows, increase fuel loads and make fires burn hotter. SANParks describes alien-fuelled fires as more intense than fynbos fires, sometimes damaging soil and reducing fynbos recovery.

Threat	Conservation response	Why it is difficult
Alien trees	Mechanical clearing, follow-up clearing, Working for Water teams, restoration after removal.	Seedbanks persist, follow-up funding is essential, and some public users like shaded alien forests.
Fire suppression	Ecologically informed burns, firebreaks, post-fire monitoring.	Too little fire can age fynbos; too frequent or too intense fire can damage species recovery.
Urban edge ignition	Fire-risk planning, access control, public education, rapid response.	Arson, accidental ignitions and climate-driven heat/wind events raise risk.

Magda Ehlers on Pexels`} />

3. African Penguins, Seabirds and Coastal Rescue

Cape Town’s most famous wildlife conservation story is probably the African penguin. The Boulders / Simon’s Town colony is globally known, but it is also part of a species-wide crisis. African penguins have suffered major declines linked to prey scarcity, commercial fishing pressure, climate-driven shifts in sardine and anchovy, disease, oiling, predation, disturbance and nesting habitat loss.

The Simon’s Town Penguin Management Area is a partnership space involving the City of Cape Town, SANParks, CapeNature, seabird specialists and volunteers. The 2023 Simon’s Town annual report describes daily monitoring, penguin ranger work, injured birds moved to SANCCOB, nest monitoring, road patrols and public-facing management at an urban colony.

SANCCOB, based in Table View, is one of Cape Town’s most important wildlife rescue institutions. Its work includes rescuing, rehabilitating and releasing ill, injured, abandoned and oiled seabirds, especially African penguins. The conservation value is not only in saving individual birds; it is also in keeping threatened breeding populations from losing recoverable adults, chicks and eggs.

Recent penguin science has become increasingly policy-relevant. McInnes et al. (2024) evaluated fishery no-take zones around African penguin colonies, while Sherley et al. (2024) argued that the African penguin’s population decline supports uplisting the species to Critically Endangered. In 2025, a court settlement created new fishing closures around key colonies including Robben Island, an important local win for penguin conservation.

Dane Damons on Pexels`} />

4. Cape Peninsula Baboons: Coexistence Under Pressure

Chacma baboons on the Cape Peninsula are a classic urban edge problem. They are native, intelligent, social and ecologically valuable — but they can also raid bins, enter houses, damage property, injure pets or people, and become dependent on human food. The conflict is not simply “people versus baboons.” It is also people versus people: residents, conservationists, animal-welfare groups, city managers, SANParks, landowners and tourism operators often disagree about what humane management should look like.

Hoffman & O’Riain (2012) used spatial ecology to understand the extent and severity of human-baboon conflict in the Cape Peninsula. Their work helped show that conflict is spatially predictable: baboon raiding is shaped by access to high-calorie human food, urban edges, troop movement routes and the effectiveness of barriers and monitors.

Current management includes ranger teams, waste control, education, deterrence, spatial planning, fencing debates, and ongoing attempts to reduce attractants. The hard lesson is that baboon conservation is as much a governance problem as a biology problem. A troop can only be kept wild if households, businesses and public spaces stop subsidising raiding behaviour.

Practical conservation lesson: Feeding baboons, leaving bins unsecured, or normalising baboons inside houses is not kindness. It increases habituation and often leads to injury, conflict and lethal outcomes for baboons.

5. Urban Caracals and the Hidden Cost of Rat Poison

Cape Town’s caracals are a rare case of a medium-sized wild cat persisting inside a major city. They move through Table Mountain National Park, the urban edge, vineyards, greenbelts and coastal fragments. The Urban Caracal Project has turned them into one of the best-studied urban carnivore populations in Africa.

The research story is sobering. Serieys et al. (2019) found widespread anticoagulant rodenticide exposure in predators in Cape Town, with caracals among the most exposed species. Later genomic work by Kyriazis et al. (2024) examined the consequences of isolation and gene flow for the Cape Peninsula caracal population, highlighting how urban development can turn a charismatic predator into a small, genetically vulnerable population.

Conservation action here is not a single reserve. It is a city-wide behaviour change: reduce second-generation anticoagulant rat poisons, report roadkill and snares, protect movement corridors, manage vineyards and urban edges responsibly, and keep public enthusiasm connected to evidence rather than myth.

6. Western Leopard Toads: Road Crossings, Wetlands and Citizen Rescue

The western leopard toad is an endangered amphibian strongly associated with Cape Town’s urban wetlands and seasonal breeding movements. The conservation challenge is brutally simple: adults need to cross roads to reach breeding sites, and many are killed by vehicles.

Conservation efforts include volunteer road patrols during breeding season, public reporting, reflective road signs, wetland rehabilitation, habitat protection, school education and local monitoring. The City’s own fact sheet lists urbanisation, habitat loss, roads, pollution, litter and water quality as major threats.

This is one of Cape Town’s best examples of small-scale citizen conservation mattering. A single wet night can be a major breeding movement. A few trained volunteers with torches, buckets, signage and good data can directly reduce mortality while also building public awareness of the wetland system beneath the suburb.

7. Wetlands, Estuaries and the Ramsar Wetland City Story

Cape Town is also a wetland city. Table Bay Nature Reserve, Zandvlei, Rondevlei, Zeekoevlei, Rietvlei, Edith Stephens Wetland Park, Lower Silvermine and the broader False Bay wetland system support birds, amphibians, fish, hippos, invertebrates, plants and flood-buffering functions.

The City’s Biodiversity Management Progress Report notes that Cape Town was accredited as a Ramsar Wetland City in 2022, one of only a few in Africa. That recognition matters because urban wetlands are often treated as wastelands until they flood, burn, smell, clog with litter, or lose the species that made them valuable.

Wetland conservation in Cape Town includes alien plant clearing, water-quality monitoring, litter removal, reedbed management, sewage-spill response, bird counts, amphibian protection, hydrological restoration and education. The work is less glamorous than penguins, but it is central to climate adaptation and urban biodiversity.

8. False Bay, Kelp Forests and Marine Protected Areas

Cape Town’s marine conservation work sits inside one of the richest cold-temperate marine systems in the world. The Table Mountain National Park Marine Protected Area wraps around much of the Cape Peninsula and includes controlled zones and no-take areas. The MPA protects rocky reefs, kelp forests, sandy habitats, seabird foraging areas and species such as African penguins, abalone, linefish, sharks and rays.

SANParks’ 2022 state-of-knowledge report for the Table Mountain National Park MPA summarises the evidence base for managing this system. False Bay is not just a beach and surf area; it is a living seascape shaped by fishing pressure, shark movements, kelp forest dynamics, whale migration, tourism, pollution and climate-linked marine heat events.

Shark Spotters is one of Cape Town’s most internationally interesting conservation programmes because it solves a conflict without killing the animal. Engelbrecht et al. (2017) showed that the Shark Spotters programme can reduce spatial overlap between recreational water users and white sharks in False Bay. It is a practical example of coexistence technology: trained observers, flags, sirens, public communication and beach behaviour change.

9. Sea Turtle Rescue and Ocean Plastic Evidence

Cape Town is not a major turtle nesting city, but it is a major rescue node. Young loggerhead turtles and other species can wash up cold-stunned, injured or weak along the Western Cape coast after storms and currents carry them into colder waters. The{" "} Two Oceans Aquarium Foundation Turtle Conservation Centre {" "} rescues, rehabilitates and releases hatchling and adult turtles, and its Turtle Rescue Network turns public reports into real conservation action.

The programme also produces important evidence about ocean plastic. Many rescued turtles have ingested plastic or are entangled in marine debris. In this way, the turtle hospital is both a rescue facility and a window into what is happening offshore.

10. Restoration, Friends Groups and Citizen Science

Cape Town conservation depends heavily on local people. Friends groups, botanical societies, bird clubs, toad volunteers, iNaturalist users, school groups, neighbourhood clean-up teams and reserve volunteers often provide the eyes, labour and political support that formal agencies cannot supply alone.

The most valuable citizen work is usually unglamorous: removing invasive seedlings before they become trees, recording plant species after fire, reporting caracal roadkill, monitoring frog calls, counting birds, rescuing penguin chicks, collecting litter before it reaches wetlands, and pushing back when irreplaceable fragments are treated as vacant land.

In a city as biologically fragmented as Cape Town, small patches matter. So do small acts repeated for years.

Who Is Doing the Work?

Organisation / programme	Main focus	Useful link
City of Cape Town Biodiversity Management	Municipal reserves, BioNet, wetlands, invasive species, urban biodiversity, reserve management.	City nature reserves
SANParks / Table Mountain National Park	Mountain fynbos, fire, alien clearing, Cape Point, Boulders, marine protected area.	Table Mountain National Park
SANCCOB	Seabird rescue, African penguins, oiled bird response, chick/egg rescue and rehabilitation.	SANCCOB
Shark Spotters	Non-lethal shark safety and white shark coexistence at Cape Town beaches.	Shark Spotters
Urban Caracal Project	Urban carnivore ecology, caracal genetics, roadkill, snares, rodenticide exposure and public reporting.	Urban Caracal Project
Western Leopard Toad conservation groups	Road patrols, breeding-season rescue, wetland protection and public education.	Western Leopard Toad
Two Oceans Aquarium Foundation	Turtle rescue, turtle rehabilitation, plastic pollution evidence, ocean education.	Turtle Conservation Centre
BirdLife South Africa / Cape Bird Club	Penguins, seabird monitoring, bird counts, Important Bird and Biodiversity Areas, advocacy.	BirdLife South Africa

How to Help Without Making Things Worse

Do not feed wildlife. This is especially important for baboons, penguins, seals, gulls and urban mammals.
Stop using high-risk rat poisons. Rodenticides move through the food chain and can kill caracals, owls, mongooses and otters.
Secure bins and compost. Waste is one of the biggest drivers of baboon conflict.
Drive carefully near wetlands during rainy breeding nights. {" "} Western leopard toads and other amphibians are vulnerable on roads.
Volunteer through credible groups. Reserve friends groups, SANCCOB, toad patrols and beach clean-ups are practical entry points.
Use citizen science apps responsibly. Upload biodiversity observations, but obscure locations for sensitive species when needed.
Respect closures after fire. Burnt fynbos needs time to regenerate, and trampling can damage seedlings.
Support planning that protects irreplaceable fragments. {" "} The least scenic patch may hold the rarest plants.

Selected Papers and Reports

Rebelo, A. G., Holmes, P. M., Dorse, C. & Wood, J. (2011).{" "} Impacts of urbanization in a biodiversity hotspot: conservation challenges in Metropolitan Cape Town . South African Journal of Botany.
Holmes, P. M., Rebelo, A. G., Dorse, C. & Wood, J. (2012).{" "} Can Cape Town’s unique biodiversity be saved? Balancing conservation imperatives and development needs . Ecology and Society.
Forsyth, G. G. & van Wilgen, B. W. (2008).{" "} The recent fire history of the Table Mountain National Park and implications for fire management . Koedoe.
van Wilgen, B. W. et al. (2012).{" "} Evidence, perceptions, and trade-offs associated with invasive alien plant control in Table Mountain National Park . Ecology and Society.
Hoffman, T. S. & O’Riain, M. J. (2012).{" "} Monkey Management: Using Spatial Ecology to Understand the Extent and Severity of Human-Baboon Conflict in the Cape Peninsula, South Africa . Ecology and Society.
Engelbrecht, T. et al. (2017).{" "} Shark Spotters: Successfully reducing spatial overlap between white sharks and recreational water users in False Bay, South Africa . PLOS ONE.
Serieys, L. E. K. et al. (2019).{" "} Widespread anticoagulant poison exposure in predators in a rapidly growing South African city . Science of the Total Environment.
Kyriazis, C. C. et al. (2024).{" "} The influence of gene flow on population viability in an isolated urban caracal population . Proceedings of the Royal Society B.
McInnes, A. M. et al. (2024).{" "} Commercial fishery no-take zones for African penguins are not fit for purpose . ICES Journal of Marine Science.
Sherley, R. B. et al. (2024).{" "} The African Penguin Spheniscus demersus should be classified as Critically Endangered . Ostrich.

Further Local Sources

Conservation X Labs and Wildbook

Sun, 14 Jun 2026 00:00:00 GMT

import { Image } from "astro:assets"; import BlogImage from "../../components/blog/BlogImage.astro"; import whaleShark from "../../assets/blog/conservation-x-labs-wildbook/10154784_emma_li.jpg"; import whaleFluke from "../../assets/blog/conservation-x-labs-wildbook/34392859_francesco_ungaro.jpg"; import grevysZebra from "../../assets/blog/conservation-x-labs-wildbook/37518146_brett_aukburg.jpg"; import cxlLogo from "../../assets/blog/conservationxlabs-logo.svg"; import wildMeLogo from "../../assets/blog/WildMe-Logo.png";

The hardest conservation question is often the simplest one: how many animals are left?

For many species, the answer is still uncertain. Wildlife populations move across political borders, hide in remote habitats, appear only at night, or live underwater. Traditional monitoring methods — field surveys, tagging, aerial counts, and manual photo-identification — are scientifically powerful, but slow and expensive. In a world where biodiversity is declining faster than monitoring systems can often keep up, conservation needs better data at greater speed.

Conservation X Labs and the{" "} Wild Me / Wildbook project sit directly inside that gap. Conservation X Labs is a nonprofit innovation organization whose mission is to prevent the sixth mass extinction. Wildbook is an open-source wildlife research platform that uses computer vision, artificial intelligence, citizen science, and collaborative data management to identify individual animals and turn photographs into population knowledge.

This is not AI replacing conservation biology. It is AI being used as field infrastructure: helping researchers detect animals in images, compare markings, curate sightings, estimate abundance, and collaborate across borders, years, and institutions.

Source note: Metrics in this article come from public Conservation X Labs, Wild Me, Wildbook, GitHub, and research sources accessed on 14 June 2026. Live project metrics such as number of animals tracked, sightings, platforms, publications, and releases change over time.

Why Wildlife Monitoring Needs a Different Model

The biodiversity crisis is a measurement crisis as much as a protection crisis. The 2019 IPBES Global Assessment warned that around{" "} one million animal and plant species are threatened with extinction, many within decades. The IUCN Red List continues to expand its assessments, but for many species, conservation decisions are still constrained by limited population data, uneven field coverage, and long delays between data collection and analysis.

Conservation monitoring has three recurring bottlenecks:

Scale: camera traps, tourists, drones, researchers, and local communities can generate millions of photos, far more than human experts can review manually.
Individual identity: conservation often needs to know not just that a zebra, whale shark, manta ray, lynx, or leopard was seen, but whether it is the same animal seen before.
Collaboration: populations do not respect project boundaries. Long-lived datasets need to work across teams, regions, species, and decades.

Wildbook tackles those bottlenecks by combining structured ecological data with pattern recognition and collaborative workflows. It is especially useful for species whose bodies carry visible, individually distinctive patterns: stripes, spots, scars, flukes, fins, wrinkles, whisker spots, or shell markings.

Who Conservation X Labs Is

Conservation X Labs, often shortened to CXL, describes its mission as preventing the sixth mass extinction, the first extinction in Earth history driven by a single species: humans. Its model is not limited to grants or traditional field programs. It combines open innovation challenges, technology development, entrepreneurship, field partnerships, and market-facing solutions.

CXL reports impact across three broad modes of work:

Mode	What it means	Examples
Open innovation	Prizes, challenges, and support for external conservation innovators.	Grand challenges, startup support, funding for breakthrough solutions.
In-house invention	Building technology directly where a conservation bottleneck is not being solved fast enough.	Sentinel, Wild Me technologies, NABIT assays.
Field deployment	Working with conservationists, governments, and protected-area managers to make tools usable in real-world environments.	Protected-area monitoring, invasive-species detection, wildlife population assessment.

On its public website, CXL reports more than{" "} $12 million in funding given to breakthrough solutions,{" "} 165 innovations supported, 217{" "} wildlife species tracked using Wild Me and Sentinel, and{" "} 256,000+ animals re-identified. A later 2025 Meta AI profile reports updated figures, including 20 innovation challenges and more than 299,000 animals re-identified, which suggests the program metrics are continuing to grow.

The Wild Me Merger

In January 2024, Conservation X Labs and Wild Me{" "} announced a merger. Wild Me became part of Conservation X Labs, bringing with it a mature set of open-source conservation software platforms:{" "} Wildbook, Codex, and{" "} Scout.

At the time of the merger, Conservation X Labs said Wild Me platforms had served more than 1,800 researchers, tracked more than 200,000 individual animals, and recorded over{" "} one million sightings. CXL's current Wild Me page reports a larger footprint: 280,000 individual marine and terrestrial animals tracked, 250+ species,{" "} 14 million+ photos, 1.4 million{" "} sightings, and more than 100 scientific publications{" "} directly crediting Wildbook platforms and data.

The strategic reason for the merger is clear: CXL had field-facing conservation technology such as Sentinel, while Wild Me had deep experience in animal re-identification, species-specific platforms, citizen science, and long-term research data infrastructure. Together, they can push wildlife AI from offline analysis toward near-real-time conservation response.

What Wildbook Is

Wildbook is an open-source software framework for wildlife research. The project describes itself as a platform supporting mark-recapture, molecular ecology, and social ecology studies. It provides collaborative data storage, advanced search, APIs for export to external analysis tools, biodiversity database exposure, and animal biometric matching across multiple species.

In practical terms, a Wildbook installation is a web platform where researchers and contributors can upload animal encounters, attach time and location metadata, identify the species, run image analysis, compare the animal against known individuals, and maintain a long-term catalog.

The official Wildbook documentation defines it as a browser-based, multi-user platform for collaboratively tracking individual animals and estimating population sizes. Each installation can support multiple researchers and multiple species.

Concept	Meaning in Wildbook
Media asset	A photograph, video, or other uploaded media file.
Annotation	A marked animal or body region inside a media asset.
Encounter	A record of one animal observation, usually including media, time, place, and biological metadata.
Sighting	A grouped observation event that may include multiple encounters.
Individual	The identified animal that links repeated encounters over time.
Relationship or social unit	Information about associations among individuals, such as groups, pods, or social structures.

Emma Li on Pexels`} />

How Wildbook Turns Photos Into Conservation Data

Wildbook is best understood as a pipeline that moves from raw images to usable ecological information.

Collect: images come from field researchers, camera traps, drones, tourists, citizen scientists, or partner organizations.
Upload: contributors submit photos with metadata such as date, time, location, species, sex, behavior, or observer notes.
Detect: image analysis tools locate animals or relevant body regions in the photograph.
Identify: matching algorithms compare visual features against known individuals in the catalog.
Review: researchers or trained reviewers accept, reject, or investigate AI-suggested matches.
Analyze: the curated data supports mark-recapture, movement, survival, abundance, social ecology, and conservation planning.

This workflow matters because individual identification is the bridge between "we saw an animal" and "we know how this population is changing." When the same animal can be recognized across years, scientists can estimate survival, recruitment, migration, site fidelity, social relationships, and population size.

The Technical Architecture

Wildbook is not just a model. It is a software ecosystem that combines a research database, web application, user interface, API layer, and machine-learning backend.

Layer	Role	Technology notes
Wildbook web application	Stores encounters, individuals, metadata, users, projects, and search workflows.	The public GitHub repository describes the app as Java and JavaScript / JSP, licensed under GPL-2.0.
Database	Maintains long-term structured wildlife records.	Official documentation describes a high-performance PostgreSQL database for multiple wildlife-related data types.
WBIA	Runs computer vision and machine-learning workflows for image analysis.	Wildbook Image Analysis is a Python backend service, licensed under Apache-2.0.
Species-specific algorithms	Support detection, localization, classification, and individual re-identification.	Techniques include convolutional neural network detection, keypoint methods, SIFT descriptors, LNBNN matching, and plugins such as CurvRank or fluke-matching tools.
Human review	Validates or corrects automated suggestions.	Essential for data quality, especially when images are low quality, animals are occluded, or species markings are subtle.

The architecture also explains why Wildbook is more than a photo-ID app. It is designed to preserve ecological context: who observed the animal, where it was seen, when it was seen, what evidence supports the identification, and how that record fits into a broader population database.

Wildbook Is a Family of Platforms

Wildbook began as a collaborative platform for globally coordinated whale shark research. After requests to use the software beyond whale sharks, it became an open-source, community-maintained standard for mark-recapture studies.

Wild Me lists multiple Wildbook-based platforms, each adapted to a species or species group. Examples include:

Platform	Focus
Flukebook	Whales and dolphins.
Sharkbook	Sharks, including whale sharks and other species using fins, gill markings, scars, and other patterns.
MantaMatcher	Giant mantas and rays.
GiraffeSpotter	Giraffe species and individual spot-pattern matching.
Internet of Turtles	Sea and terrestrial turtles.
Zebra Wildbook	Grevy's and plains zebras.
Wildbook for Lynx	Iberian lynx and lynx photo-identification.
African Carnivore Wildbook	Multiple large carnivore species, including use cases for broad-landscape demography and dispersal.
Amphibian and Reptile Wildbook	Multiple amphibian and reptile species.
Whiskerbook	Species whose whisker patterns support individual identification.

This species-specific model is important. A whale fluke, zebra stripe pattern, manta belly pattern, giraffe coat, shark fin, and lynx coat all require different data practices and sometimes different matching techniques. Wildbook provides the shared infrastructure, while individual platforms adapt to the biology of the species.

Francesco Ungaro on Pexels`} />

Impact Case Study: The Great Grevy's Rally

One of Wildbook's clearest proof points is the{" "} Great Grevy's Rally in Kenya. In 2016, the Grevy's Zebra Technical Committee enlisted the public in a citizen-science survey across the species' range. Participants photographed visible Grevy's zebras over two days, producing more than{" "} 25,000 usable images. Those images were analyzed to estimate a Kenyan population of around{" "} 2,350 Grevy's zebras.

The broader Wildbook research literature describes this as an example of crowdsourcing, computer vision, and ecological statistics coming together: ordinary citizens collected the images, algorithms and expert workflows helped match individuals, and the results informed official conservation knowledge.

The lesson is not that AI alone saved the zebra. The lesson is that conservation data systems can be redesigned so that more people can contribute useful observations and researchers can turn those observations into rigorous estimates faster.

Brett Aukburg on Pexels`} />

How Wildbook Connects to CXL's Wider AI Stack

Wildbook is part of a wider Conservation X Labs technology strategy. Another major CXL system is Sentinel, an AI toolkit for field monitoring that combines a device, AI models, and a dashboard to deliver near-real-time information from remote conservation sites.

The Wild Me merger announcement described the long-term opportunity clearly: Sentinel could notify conservation authorities not only that a jaguar was seen by a device, but which individual jaguar passed in front of the camera. That is the difference between species detection and individual-level intelligence.

CXL's 2025 collaboration around the SA-FARI dataset and Meta's Segment Anything models shows where this field is moving next: video, behavior, tracklets, segmentation masks, and open benchmark datasets for ecological monitoring. Wildbook's individual-ID heritage fits naturally into this future stack, but it remains most valuable when paired with ecological questions and human review.

Limits, Risks, and Caveats

Wildbook is powerful, but no automated identification platform removes the hard parts of conservation science. Recent research on automated individual animal identification argues that the main barrier is often not simply model performance. The bigger issue is whether the system is aligned with ecological goals: what question is being asked, what data are available, what mistakes matter, and how decisions are reviewed.

Challenge	Why it matters
Image quality	Blur, lighting, angle, occlusion, partial bodies, and repeated burst images can make individual ID unreliable.
Species differences	Some species have obvious individual patterns; others need different body parts, multi-image evidence, genetics, or manual confirmation.
Dataset bias	Citizen-science and tourism photos are not random samples. They cluster around roads, boats, tourist sites, seasons, and charismatic individuals.
False matches and missed matches	Errors can distort abundance estimates, movement histories, survival analyses, and management decisions.
Transparency	Researchers need to know why a match was suggested, how confident the system is, and what evidence supports a decision.
Operational sustainability	Long-term wildlife monitoring needs hosting, maintenance, model updates, data governance, and user support, not just a one-off model.

The best framing is therefore{" "} AI-assisted conservation science. Wildbook accelerates curation and makes large-scale collaboration possible, but the conservation value depends on well-designed sampling, expert validation, transparent workflows, and careful interpretation.

Why Wildbook Is Important

Wildbook matters because it makes wildlife populations legible at the level conservation often needs: the individual. Species-level detection tells you what is present. Individual re-identification tells you who is present, whether that animal survived, where it moved, who it associates with, and whether the population is growing or shrinking.

Its larger significance is that it treats conservation data as shared infrastructure. A single whale shark, manta ray, zebra, lynx, turtle, or leopard may be photographed by many people in many places over many years. Without a collaborative system, those observations remain scattered. With Wildbook, they can become a life history.

That is the promise of Conservation X Labs and Wildbook: not technology for its own sake, but technology that helps conservationists move at the speed of the threats they are trying to stop.

Snapshot

Question	Answer
What is Conservation X Labs?	A nonprofit conservation-technology and innovation organization focused on preventing human-driven extinction.
What is Wild Me?	The conservation AI and open-source software team now operating as part of Conservation X Labs.
What is Wildbook?	An open-source, web-based platform for collaborative wildlife photo-identification, mark-recapture, and population monitoring.
What does it identify?	Individual animals, usually by natural markings such as spots, stripes, scars, fins, flukes, shells, or other visible features.
What technologies does it use?	Web databases, PostgreSQL, Java/JSP, Python, computer vision, machine learning, species-specific matching algorithms, APIs, and human review workflows.
Why is it useful?	It turns scattered photos into structured records of individuals, sightings, movements, population estimates, and long-term conservation insight.

Sources and Further Reading

GBIF: The Global Biodiversity Information Facility

Sun, 14 Jun 2026 00:00:00 GMT

import { Image } from "astro:assets"; import BlogImage from "../../components/blog/BlogImage.astro"; import fieldResearcher from "../../assets/blog/gbif-global-biodiversity-information-facility/5036758_katrin__bolovtsova.jpg"; import pinnedButterflies from "../../assets/blog/gbif-global-biodiversity-information-facility/6055287_tamula_aura.jpg"; import butterflyCollection from "../../assets/blog/gbif-global-biodiversity-information-facility/29751458_giulia_botan.jpg";

Biodiversity science depends on a deceptively simple question: what species was seen, where, and when?

For centuries, the answers lived in museum drawers, herbarium sheets, field notebooks, local databases, government surveys, environmental assessments, and now smartphone apps and automated sensors. Each record may be small — a plant specimen collected in 1890, a bird checklist from yesterday, a camera-trap image, a DNA-derived detection, a preserved insect, a coral observation — but together they form the evidence base for understanding life on Earth.

GBIF, the{" "} Global Biodiversity Information Facility, is one of the core pieces of global biodiversity infrastructure. It is not just a website and not merely a database. It is an international network, standards ecosystem, data publishing system, API platform, and citation framework designed to make biodiversity occurrence data discoverable, reusable, and attributable.

Its importance is easiest to understand this way: GBIF does for species occurrence data what public web infrastructure does for documents. It gives many independent data holders a shared way to publish, index, search, cite, and reuse records without forcing every institution into a single database or software stack.

Source note: This article was prepared from public GBIF documentation, GBIF.org pages, technical documentation, GBIF publishing guidance, data quality guidance, and related biodiversity informatics sources accessed on 14 June 2026. Live counters for occurrence records, datasets, publishers, papers, and participants change continuously, so exact operational totals should be checked on GBIF.org at the time of use.

Why GBIF Matters

Conservation decisions are only as good as the evidence behind them. A government deciding where to place a protected area, a researcher modelling species distributions under climate change, a conservation NGO assessing invasive species risk, or a company screening biodiversity impacts all need reliable information about the distribution of species.

The problem is that biodiversity knowledge is scattered. Natural history museums may hold millions of specimens. Universities maintain research datasets. Birders and naturalists contribute observations through citizen-science platforms. National agencies publish monitoring data. Environmental consultants generate survey records. Molecular labs produce DNA-derived detections. Without shared standards and open infrastructure, these datasets remain isolated and difficult to combine.

GBIF reduces that fragmentation by giving the biodiversity community a common publishing and discovery layer. Its value comes from three linked functions:

Mobilization: helping institutions publish biodiversity data using shared standards, tools, metadata, and licences.
Integration: indexing records from thousands of datasets so they can be searched by taxon, place, time, basis of record, institution, licence, coordinate quality, and other fields.
Reuse: supporting downloads, APIs, cloud access, DOI-based citation, and literature tracking so data can flow into research, conservation, policy, and decision-making.

This makes GBIF especially important for global-scale questions. No single research group can collect enough records to represent global biodiversity. GBIF works because it federates many sources: museums, herbaria, government agencies, NGOs, citizen-science platforms, research projects, and thematic networks.

KATRIN BOLOVTSOVA on Pexels`} />

What GBIF Is

GBIF describes itself as an international network and data infrastructure funded by governments, with the goal of giving anyone, anywhere, open access to data about all types of life on Earth. It is coordinated by a Secretariat in Copenhagen and works through participating countries, organizations, and participant nodes.

The important phrase is network and data infrastructure . GBIF does not own most of the data it serves. Data remain associated with their publishers. GBIF provides the shared infrastructure that lets those records be standardized, indexed, discovered, downloaded, cited, and reused.

GBIF is	GBIF is not
A global network of participating countries and organizations.	A single research institute collecting all biodiversity records itself.
A publishing and indexing infrastructure for biodiversity datasets.	A guarantee that every record is correct, complete, unbiased, or fit for every analysis.
A standards-based platform using Darwin Core and related biodiversity data standards.	A replacement for taxonomic experts, collection managers, field ecologists, or local knowledge.
A discovery, download, API, cloud, and citation layer for biodiversity data reuse.	A closed commercial data product or paywalled repository.

GBIF began from international recognition that biodiversity data were too fragmented to support science and sustainable development properly. The Organization for Economic Cooperation and Development's Megascience Forum recommended a global mechanism in 1999, and GBIF was established in 2001. In 2026, GBIF marks 25 years of building global biodiversity informatics infrastructure.

What Kind of Data GBIF Handles

The core unit in GBIF is the species occurrence record: evidence that a taxon occurred at a particular place and time. A record may come from a museum specimen, a herbarium sheet, a human observation, a machine observation, a fossil, a living specimen, a material citation, an environmental sample, or a published dataset.

Occurrence records are powerful because they combine three basic facts:{" "} taxon, location, and{" "} date. When those facts are standardized and linked to metadata, they can support mapping, trend analysis, species distribution modelling, invasive species alerts, conservation prioritization, and environmental risk assessment.

Source type	What it contributes	Typical strengths	Typical caveats
Museum and herbarium specimens	Physical evidence collected over centuries.	Verifiable material, historical depth, taxonomic value.	May have imprecise locality data, colonial collection bias, or outdated taxonomy.
Citizen-science observations	High-volume recent records from volunteers and naturalists.	Fresh, wide coverage, often image-supported.	Uneven sampling near roads, cities, parks, and popular species.
Monitoring and survey datasets	Structured observations from agencies, researchers, NGOs, and field projects.	Protocol-driven, repeatable, often designed for trend detection.	May be geographically limited or difficult to compare across protocols.
Machine observations	Records from sensors, camera traps, acoustic devices, or automated systems.	Scalable and increasingly important for near-real-time monitoring.	May require careful validation of automated identifications.
DNA-derived and molecular records	Detections from DNA barcodes, metabarcoding, eDNA, or sequence-based workflows.	Can reveal cryptic, microscopic, or hard-to-observe biodiversity.	Requires careful handling of sequence provenance, taxonomic assignment, and sampling context.

GBIF also supports checklist, sampling-event, metadata-only, and other dataset classes. The long-term direction is toward a richer data model that can better represent surveys, interactions, material samples, DNA-derived data, and policy-ready biodiversity indicators.

Giulia Botan on Pexels`} />

How Data Moves Through GBIF

GBIF works because it separates publishing from{" "} indexing. A museum, university, government department, citizen-science network, or research organization publishes its dataset. GBIF crawls and indexes it. Users then discover and download data through GBIF.org, the API, cloud snapshots, or other tools.

Prepare: the data holder cleans its dataset, maps fields to standard terms, chooses a licence, and writes metadata.
Publish: the organization publishes through the Integrated Publishing Toolkit, a hosted IPT, a Living Atlas installation, an API workflow, or another endorsed route.
Index: GBIF processes the dataset, interprets taxonomy and geography, flags issues, and makes records searchable.
Discover: users find records through GBIF.org, species pages, dataset pages, maps, filters, APIs, or literature links.
Download: serious research use normally creates an occurrence download with a DOI, making the exact data extraction citable and reproducible.
Cite and improve: users cite the DOI, data publishers receive credit, and errors or improvements can flow back to the original data holders.

This workflow is why GBIF is more than a search engine. It creates a feedback loop between data publishers, data users, standards, tools, and citation practices.

Darwin Core: The Shared Language

GBIF's interoperability depends heavily on Darwin Core, a community-developed biodiversity data standard maintained through Biodiversity Information Standards (TDWG). Darwin Core provides a stable vocabulary for describing biodiversity records from many sources.

The standard matters because biodiversity data are messy. One database might store a species name as scientific_name, another as{" "} taxon, another as latinName. One dataset might call latitude lat; another might store it inside a locality string. Darwin Core gives publishers and aggregators a shared set of terms such as scientificName, eventDate,{" "} decimalLatitude, decimalLongitude,{" "} basisOfRecord, occurrenceID,{" "} institutionCode, and catalogNumber.

GBIF's occurrence data quality requirements show how basic Darwin Core terms become practical publishing rules. For occurrence-only datasets, fields such as occurrenceID, basisOfRecord,{" "} scientificName, and eventDate are required, while country codes, coordinates, geodetic datum, coordinate uncertainty, kingdom, taxon rank, and abundance fields are strongly recommended.

Darwin Core term	Why it matters
`occurrenceID`	Provides a persistent identifier for the record, helping avoid duplication and ambiguity.
`scientificName`	Links the record to a taxon and enables taxonomic interpretation.
`eventDate`	Places the record in time, enabling historical and trend analysis.
`decimalLatitude` / `decimalLongitude`	Supports mapping, modelling, and spatial analysis.
`coordinateUncertaintyInMeters`	Helps users decide whether a record is spatially precise enough for a given use case.
`basisOfRecord`	Indicates whether the evidence is a preserved specimen, human observation, machine observation, fossil, living specimen, or other source.

For publishers, Darwin Core is a discipline: describe the data well enough that someone else can reuse it safely. For users, it is a map: know which fields to trust, filter, inspect, and cite.

Tamula Aura on Pexels`} />

The Integrated Publishing Toolkit

The Integrated Publishing Toolkit, or{" "} IPT, is GBIF's widely used open-source tool for publishing biodiversity datasets. An IPT installation lets organizations prepare metadata, map their data to Darwin Core, publish datasets, and register them with GBIF.

The IPT is important because many biodiversity data holders are not software companies. A herbarium, national park agency, university collection, or NGO may have valuable records but limited capacity to build custom publishing infrastructure. IPT gives them a common path to publish data in a standards-compliant way.

GBIF documentation describes several publishing routes: running an institutional IPT, using hosted IPT services through a national or thematic node, publishing through Living Atlases infrastructure, or using more customized programmatic publishing workflows. Individuals generally publish through affiliated organizations, citizen-science platforms, or data papers rather than directly as standalone publishers.

Publishing route	Best fit
Institutional IPT	Organizations that can host and maintain their own publishing server.
Hosted IPT	Institutions that need a national, regional, or thematic node to host the publishing infrastructure.
Living Atlases	National or regional biodiversity portals aligned with GBIF-style data publishing and discovery.
API or custom workflows	Large, technically mature publishers with automated data pipelines.
Data papers	Researchers who want peer-reviewed recognition for curated datasets.

How Researchers and Analysts Use GBIF Data

GBIF-mediated data support a wide range of scientific and policy uses. Common applications include species distribution modelling, climate-change impact studies, invasive species risk assessment, protected-area planning, pollinator research, crop wild relative mapping, disease vector studies, red-list assessments, environmental impact screening, and national biodiversity reporting.

GBIF reports that its data are used in peer-reviewed studies at a rate of more than six papers per day. The platform also maintains a literature-tracking programme and Science Review series that surface examples of how open biodiversity data are being used in research.

Use case	What GBIF contributes	Important caution
Species distribution modelling	Occurrence points across broad geography and time.	Sampling bias, duplicate records, coordinate uncertainty, and pseudo-absence design matter.
Climate impact research	Historical and recent species records for range-shift analysis.	Observation effort changes over time, so raw counts are not automatically population trends.
Invasive species early warning	New and historical occurrences of alien or expanding species.	Taxonomic misidentifications and reporting delays can affect confidence.
Protected-area planning	Known species occurrences inside and outside candidate areas.	Absence of records is not evidence of absence without sampling-effort context.
Corporate biodiversity screening	Open species occurrence data around sites and supply chains.	Screening outputs should be treated as risk signals, not complete ecological assessments.
National reporting and policy	Shared infrastructure for mobilizing and reusing country-level biodiversity records.	Countries differ in data mobilization capacity, digitization history, and publishing coverage.

The best use of GBIF data is rarely a simple download-and-map exercise. Good analysis usually includes taxonomic cleaning, duplicate handling, coordinate filtering, uncertainty filtering, temporal filtering, licence checks, bias correction, and a clear citation trail.

The Developer View

For developers and data scientists, GBIF is unusually useful because it exposes biodiversity infrastructure through web services, documented APIs, download formats, and cloud-accessible data products.

The main API families include species services, occurrence services, occurrence image services, maps, literature, registry, and validation. The occurrence API supports real-time paged search, but serious research downloads should use GBIF's asynchronous occurrence download system or cloud snapshots rather than attempting to page through massive search results.

GBIF's technical documentation describes several download formats:

Simple CSV / tab-separated text: a practical table for spreadsheet and scripting workflows.
Darwin Core Archive: a richer zipped package with interpreted records, verbatim records, metadata, and optional extensions.
Species list: a summary export of distinct species names returned by a filter.
Occurrence cubes: aggregated occurrence outputs by taxonomic, temporal, and spatial dimensions.
Parquet / cloud access: formats suitable for large analytical workflows.

A simple API exploration might look like this:

```bash # Match a scientific name to the GBIF Backbone Taxonomy curl "https://api.gbif.org/v1/species/match?name=Panthera%20leo" # Search occurrence records for a known taxon key curl "https://api.gbif.org/v1/occurrence/search?taxonKey=5219404&limit=10" # For serious research, use occurrence downloads rather than deep paging. # Downloads generate a DOI and make the dataset extract citable. ```

R users commonly access GBIF through rgbif, while Python users often use pygbif or direct API calls. A good rule is: use small search calls for exploration, but create a download for reproducible research.

Citation and Credit

GBIF data are free to access, but reuse comes with responsibilities. GBIF's citation guidance states that users who download individual datasets or search results and use them in research or policy agree to cite them using a DOI.

DOI citation is central to the whole system. It lets other researchers reproduce the exact download, gives credit to data-publishing institutions and individuals, and demonstrates the impact of open data sharing to funders, collection managers, and governments.

Data access pattern	Responsible citation approach
Occurrence download from GBIF.org or API	Cite the download DOI generated by GBIF.
Individual dataset	Cite the dataset DOI or recommended citation from the dataset page.
Individual occurrence	Cite the occurrence, dataset, and media source as appropriate.
Third-party package such as rgbif or pygbif	Cite the GBIF-mediated data DOI and, where relevant, the software package.
Images and other media	Respect the media licence and credit the creator, dataset, and licence.

This is one of GBIF's most important design decisions. Open biodiversity data only remains sustainable if the people and institutions doing the hard work of collecting, curating, digitizing, publishing, and maintaining records receive visible credit.

Governance and the Node Model

GBIF is governed through its participating countries and organizations. The Governing Board is the main decision-making body, with one representative from each participant country and organization. Voting rights are reserved for voting participant countries that contribute financially to GBIF's central fund, while associate participants and organizations can participate in discussion.

The node model is central. A GBIF participant node is usually a team or institution designated to coordinate biodiversity data mobilization and use within a country, organization, region, or thematic community. Nodes help publishers, support training, connect stakeholders, build capacity, promote standards, and align national or institutional priorities with the global infrastructure.

This federated model is important because biodiversity data is local before it is global. Countries, museums, universities, Indigenous communities, agencies, and NGOs hold different kinds of expertise and authority. GBIF works best when global infrastructure supports local stewardship rather than replacing it.

Data Quality: Powerful, Not Perfect

GBIF is often described as the world's largest open biodiversity data infrastructure, but large does not mean complete or unbiased. The platform exposes what has been collected, digitized, standardized, licensed, and published. It does not represent an evenly sampled census of life on Earth.

GBIF itself emphasizes data quality requirements and recommendations for publishers. Scientific studies of GBIF-mediated data repeatedly show spatial, temporal, and taxonomic biases. Records are denser in places with stronger research infrastructure, richer collection histories, active citizen-science communities, and better digitization capacity. Charismatic and easy-to-observe species are often overrepresented compared with small, cryptic, marine, tropical, soil, microbial, or poorly studied groups.

Issue	Why it matters	Good practice
Spatial bias	Records cluster near roads, cities, research stations, protected areas, and wealthy regions.	Use sampling-bias correction, spatial thinning, background sampling strategy, or effort covariates.
Temporal bias	Historical specimens and recent citizen-science observations reflect different collection processes.	Filter by year, model time explicitly, and avoid treating raw record counts as abundance.
Taxonomic bias	Birds, mammals, plants, and charismatic groups are generally better represented than many invertebrates, fungi, microbes, and cryptic taxa.	Check taxon-specific coverage and avoid broad claims from unevenly sampled groups.
Coordinate error	Coordinates may be missing, rounded, generalized, transposed, centred on a country, or imprecise.	Use coordinate uncertainty fields, GBIF issue flags, country checks, and spatial validation.
Taxonomic interpretation	Names may be synonyms, misspellings, outdated combinations, homonyms, or uncertain identifications.	Use taxon keys, inspect the backbone match, and document taxonomic decisions.
Duplicates	The same observation or specimen may appear through multiple routes.	Deduplicate cautiously using occurrence IDs, institution/catalog numbers, coordinates, dates, and dataset context.
Sensitive species	Precise locations can increase risk of poaching, disturbance, or exploitation.	Use generalized data where appropriate and respect publisher restrictions and conservation ethics.

The key point is not that GBIF data are unreliable. The key point is that GBIF data are evidence, and evidence needs interpretation. Used carefully, GBIF is extraordinarily valuable. Used naively, it can produce misleading maps, biased models, or false confidence.

A Practical Workflow for Responsible Use

A responsible GBIF workflow should be reproducible, transparent, and explicit about uncertainty.

Define the question: decide whether you need presence records, a checklist, a survey dataset, a time series, or a modelling input.
Resolve taxonomy: use scientific names carefully, prefer GBIF taxon keys, and document synonym handling.
Create a reproducible download: use occurrence downloads for research-scale work so GBIF generates a DOI.
Inspect licences: confirm whether records are CC0, CC BY, CC BY-NC, or otherwise constrained.
Filter quality issues: inspect coordinates, dates, basis of record, coordinate uncertainty, geospatial flags, and taxonomic issues.
Deduplicate: remove repeated records only when you understand the source context.
Account for sampling bias: do not treat record density as species abundance unless the dataset design supports that interpretation.
Cite properly: include the GBIF download DOI, dataset citations, software citations, and date accessed.
Report limitations: describe geographic, temporal, taxonomic, and methodological limitations clearly.
Feed corrections back: when possible, notify publishers of clear errors rather than silently fixing only your local copy.

GBIF should be treated like infrastructure for evidence-based biodiversity work, not like a magic answer machine. It gives users access to enormous biodiversity evidence; the analysis still has to be scientifically careful.

How GBIF Relates to Other Biodiversity Platforms

GBIF sits in a broader biodiversity informatics ecosystem. Its role overlaps with, but differs from, platforms such as iNaturalist, OBIS, BOLD, Catalogue of Life, eBird, national biodiversity portals, Living Atlases, and museum collection systems.

Platform or system	Main role	Relationship to GBIF
iNaturalist	Citizen-science observations and community identifications.	Research-grade observations are among the major sources of recent GBIF occurrence records.
eBird	Bird observations and checklist-based bird monitoring.	The eBird Observation Dataset is a major contributor to bird occurrence data in GBIF.
OBIS	Marine biodiversity occurrence data.	Closely aligned through biodiversity standards, especially Darwin Core, with a marine focus.
Catalogue of Life	Taxonomic checklist infrastructure.	Important for taxonomic backbone and name alignment.
BOLD / barcode repositories	DNA barcode and molecular reference data.	Relevant to GBIF's growing support for DNA-derived biodiversity records.
Natural history collection systems	Institutional specimen and collection management.	Publish specimen-derived occurrence datasets into GBIF through IPT or other routes.
Living Atlases	National or regional biodiversity portals.	Often interoperate with GBIF as publishing and discovery infrastructure.

The ecosystem works because specialization is preserved. GBIF does not need to replace specialist platforms. It needs to make their data easier to discover, combine, cite, and reuse.

Where GBIF Is Going

GBIF's strategic direction for 2023–2027 focuses on science and research, policy and partnerships, community and capacity, technical infrastructure, and stronger data mobilization. Several trends stand out.

Richer data models: moving beyond simple occurrence records toward better support for surveys, samples, interactions, and monitoring datasets.
DNA-derived data: building pipelines for eDNA, metabarcoding, sequence-derived occurrence evidence, and reference libraries.
Cloud-scale analytics: making occurrence snapshots, Parquet outputs, SQL downloads, and data cubes easier to use for large analyses.
Policy-ready outputs: supporting biodiversity indicators, national reporting, invasive species work, protected-area planning, and global biodiversity frameworks.
Equity and capacity: strengthening nodes, training, and data mobilization in underrepresented regions and taxonomic communities.
Data governance: responding to debates around digital sequence information, Indigenous data governance, sensitive species, CARE principles, and equitable benefit sharing.

The next phase of GBIF is therefore not just "more records." It is better context, better provenance, better uncertainty, better governance, and better translation of biodiversity data into decisions.

Quick Reference

Question	Answer
What does GBIF stand for?	Global Biodiversity Information Facility.
What does GBIF mainly provide?	Open access to biodiversity data, especially species occurrence records, through a global network and shared infrastructure.
What is an occurrence record?	Evidence that a species or other taxon was recorded at a place and time.
What standard is central to GBIF publishing?	Darwin Core, usually packaged as Darwin Core Archive for many datasets.
What tool do many publishers use?	The Integrated Publishing Toolkit, a free open-source GBIF publishing tool.
Can I use GBIF data commercially?	It depends on the record and dataset licences. Always check licence terms, especially CC BY-NC restrictions.
Should I use the API or downloads?	Use API searches for exploration; use occurrence downloads for serious research so your data extract gets a DOI.
Does GBIF prove absence?	No. A lack of records usually means lack of published evidence, not confirmed absence.
Is GBIF data population abundance data?	Usually no. Occurrence density is influenced by sampling effort and should not be treated as abundance without appropriate survey design.

Sources and Further Reading

```

Google Earth Engine: What It Is, Why It Matters for Science, and How to Use It

Sun, 14 Jun 2026 00:00:00 GMT

import { Image } from "astro:assets"; import BlogImage from "../../components/blog/BlogImage.astro"; import codeEditorDiagram from "../../assets/blog/google-earth-engine/Code_editor_diagram.png"; import codeEditorQuickstart from "../../assets/blog/google-earth-engine/Code_editor.png"; import dynamicWorldSample from "../../assets/blog/google-earth-engine/GOOGLE_DYNAMICWORLD_V1_sample.png"; import hansenForestSample from "../../assets/blog/google-earth-engine/UMD_hansen_global_forest_change_2025_v1_13_sample.png"; import appsPublishing from "../../assets/blog/google-earth-engine/publish-project-owned-app.png";

Google Earth Engine is one of the most important scientific computing platforms built for looking at the planet from above. It lets researchers, conservation teams, governments, NGOs, and developers analyse satellite imagery and other geospatial datasets without first downloading terabytes of files.

The short version is this:{" "} Earth Engine brings the data and the computation together . Instead of downloading satellite images onto your laptop and trying to process them locally, you write a script that tells Earth Engine what data you need, what filters to apply, what calculation to run, and what output you want. Google’s servers do the heavy lifting, then return a map, chart, table, image export, or application.

This guide explains what Earth Engine is, what people mean by its “model,” why it matters for science, and how to start using it. It is written for readers who understand basic science or conservation work but may be new to remote sensing and cloud geospatial computing.

Source note: This article was prepared using Google Earth Engine’s official website, Google for Developers documentation, Google Cloud Earth Engine pages, the Earth Engine Data Catalog, Google Research’s publication page for Gorelick et al. 2017, Dynamic World documentation, Hansen Global Forest Change documentation, and selected peer-reviewed references accessed on 14 June 2026.

What Is Google Earth Engine?

Google Earth Engine is a{" "} cloud-based geospatial processing service. Google describes it as a platform that combines a multi-petabyte catalog of satellite imagery and geospatial datasets with planetary-scale analysis capabilities, used to detect changes, map trends, and quantify differences on Earth’s surface.

In practice, Earth Engine gives you three things in one place:

Part	What it means	Why it matters
Data catalog	A large public archive of satellite imagery and geospatial datasets, including Landsat, Sentinel, MODIS, climate, terrain, water, land cover, forest, fire, and population products.	You can start analysis without sourcing, downloading, cleaning, and storing raw imagery yourself.
Compute platform	Earth Engine runs geospatial operations on Google infrastructure.	You can process large areas and long time periods that would be difficult on a normal laptop.
Developer tools	A web Code Editor, JavaScript API, Python API, REST API, apps framework, exports, and links to Google Cloud services.	You can prototype, analyse, automate, publish apps, and connect results into bigger systems.

The classic peer-reviewed paper is{" "} Google Earth Engine: Planetary-scale geospatial analysis for everyone {" "} by Gorelick, Hancher, Dixon, Ilyushchenko, Thau, and Moore. Google Research summarises the platform as a cloud-based system for planetary-scale geospatial analysis applied to deforestation, drought, disaster, disease, food security, water management, climate monitoring, and environmental protection.

Official sources: Earth Engine homepage , Google for Developers Earth Engine docs , Google Cloud Earth Engine overview , and{" "} Google Research publication page .

What Is “the Model”?

Google Earth Engine is not one AI model. It is better understood as a platform with several “models” working together: a data model, a programming model, a computation model, and optional machine-learning models.

Model	Meaning in Earth Engine	Example
Data model	Earth observation data are represented as server-side objects such as images, image collections, features, feature collections, geometries, reducers, and dictionaries.	A Landsat scene is an `ee.Image`; a time series of Sentinel-2 scenes is an `ee.ImageCollection`; a park boundary is an `ee.FeatureCollection`.
Programming model	You write JavaScript or Python code that builds an analysis recipe rather than directly processing every pixel on your own machine.	`filterDate()`, `filterBounds()`,{" "} `map()`, `median()`,{" "} `reduceRegion()`, and{" "} `Export.image.toDrive()`.
Computation model	Earth Engine uses server-side, deferred execution. Your script describes what should happen; Earth Engine decides how and when to compute the result.	A chain of operations may only run when you add a layer to the map, print a result, request a chart, or start an export task.
Science model	Researchers encode scientific assumptions into workflows: vegetation index thresholds, cloud masks, classification models, training data, change-detection rules, or uncertainty thresholds.	Using NDVI to estimate vegetation greenness or using Dynamic World class probabilities to estimate land-cover change.
Machine-learning model	Earth Engine includes built-in classifiers and can connect to custom models hosted on Vertex AI through `ee.Model`.	Random Forest land-cover classification in Earth Engine, or a hosted deep-learning model that returns predictions as Earth Engine images.

This distinction matters because people often say “the model says forest was lost” or “the model detected crops.” In Earth Engine, the platform may be executing a workflow, while the actual scientific model could be a threshold, a Random Forest classifier, a published dataset such as Dynamic World, or a deep-learning model hosted elsewhere.

Real Image References

These are real image links from official Google Earth Engine / Google for Developers pages that can be used as article visuals. They are included here as live image links rather than placeholders.

Google Earth Engine Code Editor docs`} />

Get started with Earth Engine in the Code Editor`} />

Dynamic World V1 Earth Engine Data Catalog`} />

Hansen Global Forest Change v1.13 Data Catalog`} />

Earth Engine Apps documentation`} />

What Earth Engine Means for Science

Earth Engine changed environmental science because it lowered the barrier between a scientific question and a continental or global-scale answer. Before platforms like Earth Engine, a researcher often needed to discover satellite scenes, download them, store them, pre-process them, mosaic them, mask clouds, tile large rasters, run local software, and manage storage. That work still matters, but Earth Engine collapses much of the data-management burden into a shared cloud platform.

1. It makes planetary-scale analysis more accessible

A student, small conservation NGO, or researcher without a supercomputer can still ask large questions: how vegetation changed over a decade, where forest loss accelerated, how surface water shifted after drought, or where urban expansion is replacing natural habitat. The 2017 Earth Engine paper explicitly framed the platform as a way to empower not only remote-sensing specialists but also a broader audience lacking traditional supercomputing capacity.

2. It improves reproducibility

Earth Engine scripts can describe the dataset ID, date range, spatial filter, cloud mask, index calculation, reducer, visualization, and export process in one place. That makes it easier to rerun the same workflow for another year, region, or field site. It does not guarantee good science, but it gives teams a clear computational recipe.

3. It changes the scale of field science

Field surveys remain essential. Satellite data cannot tell you everything that is happening on the ground. But Earth Engine can help field teams decide where to sample, identify areas of rapid change, compare local observations to broader patterns, and generate spatial context for patrols, restoration planning, fire history, grazing pressure, water availability, habitat connectivity, and land-cover change.

4. It supports rapid environmental response

Because many datasets are updated regularly, Earth Engine can support post-fire analysis, flood mapping, drought monitoring, crop-condition mapping, storm damage mapping, and near-real-time land-cover monitoring when appropriate datasets are available. The speed advantage comes from having both current data and compute resources close together.

5. It encourages shared applications, not only papers

Earth Engine Apps allow researchers to publish interactive interfaces for non-technical users. This is important for science communication: a result can become a usable tool for land managers, park staff, policymakers, teachers, or community partners.

The Data Catalog: What You Can Analyse

The Earth Engine Data Catalog is the platform’s scientific foundation. Google Cloud describes the catalog as containing more than 90 petabytes of analysis-ready geospatial data and more than 1,000 curated geospatial datasets. Common datasets include satellite imagery, climate data, environmental indices, terrain, land cover, forest, water, fire, population, and administrative data.

Dataset family	Common use	Example Earth Engine ID
Landsat	Long-term change since the 1980s; forest loss; water change; fire scars; vegetation trends.	`LANDSAT/LC08/C02/T1_L2`
Sentinel-2	Higher-resolution optical monitoring for vegetation, land cover, agriculture, wetlands, restoration, and habitat condition.	`COPERNICUS/S2_SR_HARMONIZED`
MODIS	Frequent global observations for vegetation, land surface temperature, fire, snow, and phenology.	`MODIS/061/MOD13Q1`
Dynamic World	Near-real-time 10 m land-cover probabilities for nine classes derived from Sentinel-2.	`GOOGLE/DYNAMICWORLD/V1`
Hansen Global Forest Change	Global forest extent and forest loss/gain time-series analysis from Landsat.	`UMD/hansen/global_forest_change_2025_v1_13`
Climate and weather reanalysis	Temperature, rainfall, evapotranspiration, drought, and climate context.	`ECMWF/ERA5_LAND/MONTHLY_AGGR`

The important habit is to read each dataset’s catalog page before using it. Check the date range, spatial resolution, bands, units, scale factors, masks, update frequency, licensing, and citations. Satellite products are not interchangeable; a Landsat surface reflectance product, a Sentinel-2 top-of-atmosphere product, and a derived land-cover product answer different questions.

Core Concepts You Need to Know

Concept	Plain meaning	Example
Raster	Grid of pixels. Each pixel stores values such as reflectance, temperature, NDVI, elevation, or probability.	Satellite image, elevation model, rainfall grid.
Vector	Points, lines, or polygons with attributes.	Camera-trap points, reserve boundary, river line, transect path.
Band	One layer inside an image.	Red, green, blue, NIR, SWIR, NDVI, land-cover probability.
ImageCollection	A group or time series of images.	All Sentinel-2 images over a park between January and March.
Filter	Select only the data you need.	Filter by date, location, cloud percentage, or metadata.
Map	Apply the same function to each item in a collection.	Calculate NDVI for every image in a time series.
Reduce	Summarise many values into fewer values.	Median image, mean NDVI per reserve, yearly forest-loss area.
Mask	Exclude pixels from analysis.	Remove clouds, shadows, no-data areas, or low-confidence classifications.
Export	Save results out of Earth Engine.	Export a GeoTIFF to Drive or a CSV table to Cloud Storage.

How to Use Google Earth Engine

There are several ways to interact with Earth Engine. The best place to start is usually the browser-based Code Editor because it gives you immediate feedback: write code, run it, see the map, inspect pixels, print outputs, and start export tasks.

Step 1: Get access

Create or register a Google Cloud project for Earth Engine access. Google’s current access documentation requires projects to state whether their purpose is commercial or noncommercial. Academic, nonprofit, and some government users can use Earth Engine for noncommercial work, while operational commercial use requires paid commercial configuration. Google also introduced noncommercial eligibility verification and quota tiers, so users should check the current access page before starting a new project.

Start here:{" "} Earth Engine access {" "} and{" "} JavaScript Code Editor quickstart .

Step 2: Open the Code Editor

Open{" "} code.earthengine.google.com . The Code Editor contains the script editor, map, console, task manager, asset manager, dataset search, and drawing tools.

Step 3: Choose a question before choosing a dataset

Do not begin with “I want to use Sentinel-2.” Begin with a scientific question:

Has vegetation cover changed around this waterhole?
Where has forest loss occurred inside a buffer around the reserve?
Which restoration blocks have the strongest NDVI recovery?
Did a flood alter surface water extent?
Where did fire scars appear after the dry season?

Once the question is clear, choose data that match the required time period, resolution, geography, and uncertainty tolerance.

Step 4: Build a basic workflow

Most beginner Earth Engine workflows follow this pattern:

Define a region of interest.
Load an image or image collection.
Filter by date and place.
Mask clouds or poor-quality pixels.
Create a composite or index.
Display it on the map.
Summarise values over a polygon.
Export the result.

First Example: Sentinel-2 NDVI in the Code Editor

NDVI is a simple vegetation index based on near-infrared and red light. Healthy green vegetation usually reflects more near-infrared and absorbs more red light, so NDVI is often used as a first-pass measure of vegetation greenness. It is not a perfect measure of biodiversity, biomass, grazing pressure, or habitat quality, but it is useful for screening broad vegetation patterns.

Copy this into the Earth Engine Code Editor. Replace the point with your own reserve or study area.

```javascript // 1. Define a simple region of interest. // Example point is in northeastern South Africa. Replace with your own site. var roi = ee.Geometry.Point([31.5, -24.5]).buffer(10000); // 2. Load Sentinel-2 surface reflectance imagery. var s2 = ee .ImageCollection("COPERNICUS/S2_SR_HARMONIZED") .filterBounds(roi) .filterDate("2025-01-01", "2025-12-31") .filter(ee.Filter.lt("CLOUDY_PIXEL_PERCENTAGE", 20)); // 3. Create a median composite. var composite = s2.median().clip(roi); // 4. Calculate NDVI = (NIR - Red) / (NIR + Red). // Sentinel-2: B8 is NIR, B4 is Red. var ndvi = composite.normalizedDifference(["B8", "B4"]).rename("NDVI"); // 5. Display the result. Map.centerObject(roi, 11); Map.addLayer(roi, { color: "white" }, "Region of interest"); Map.addLayer( ndvi, { min: 0, max: 0.8, palette: ["brown", "yellow", "green"], }, "NDVI 2025", ); // 6. Summarise mean NDVI for the region. var meanNdvi = ndvi.reduceRegion({ reducer: ee.Reducer.mean(), geometry: roi, scale: 10, maxPixels: 1e9, }); print("Mean NDVI", meanNdvi); ```

This workflow loads Sentinel-2 imagery, filters it to a year and region, makes a median composite, calculates NDVI, displays it, and prints a regional average. The same structure can be adapted to fire, water, urban growth, bare ground, crop monitoring, or habitat-screening questions.

Python Example: Using Earth Engine with Geemap

Python is useful when you want Earth Engine inside notebooks, data pipelines, dashboards, or reproducible research projects. The common beginner setup is the Earth Engine Python API plus geemap for interactive maps in notebooks.

```python import ee import geemap # Authenticate once, then initialize your Earth Engine project. ee.Authenticate() ee.Initialize(project="your-google-cloud-project-id") # Define a region of interest. roi = ee.Geometry.Point([31.5, -24.5]).buffer(10000) # Load Sentinel-2 surface reflectance imagery. s2 = ( ee.ImageCollection("COPERNICUS/S2_SR_HARMONIZED") .filterBounds(roi) .filterDate("2025-01-01", "2025-12-31") .filter(ee.Filter.lt("CLOUDY_PIXEL_PERCENTAGE", 20)) ) # Median composite and NDVI. composite = s2.median().clip(roi) ndvi = composite.normalizedDifference(["B8", "B4"]).rename("NDVI") # Display an interactive map. Map = geemap.Map() Map.centerObject(roi, 11) Map.addLayer(ndvi, {"min": 0, "max": 0.8, "palette": ["brown", "yellow", "green"]}, "NDVI 2025") Map ```

Python is a good choice when your Earth Engine output needs to be combined with pandas, GeoPandas, scikit-learn, xarray, local field data, camera-trap station tables, biodiversity observations, or reporting notebooks.

Example Science and Conservation Projects

Question	Possible data	Earth Engine method	Important caution
Where has forest loss occurred?	Hansen Global Forest Change, Landsat	Mask forest loss pixels, summarise by reserve, watershed, or buffer.	Forest loss does not always equal illegal deforestation; validate context locally.
Is vegetation recovering after restoration?	Sentinel-2, Landsat NDVI/EVI	Build annual or seasonal composites and compare trends.	Greenness can increase because of invasive plants or rainfall, not necessarily ecological recovery.
Where did water availability change?	JRC Global Surface Water, Sentinel-2, Landsat NDWI	Map open water over time and compare seasonal periods.	Clouds, shadows, floating vegetation, and sediment can confuse water detection.
How is land cover changing near protected areas?	Dynamic World, Sentinel-2, national land-cover maps	Class probabilities, land-cover labels, change maps, buffer summaries.	Use probability thresholds and local validation; do not blindly trust class labels.
Where are fire scars visible?	MODIS fire products, Sentinel-2, Landsat NBR/dNBR	Use burn indices, active-fire products, and before/after composites.	Fire impact depends on season, fuel, intensity, and ecosystem type.
Can habitat suitability be mapped?	Remote sensing layers, elevation, water distance, vegetation indices, occurrence points	Export predictor layers or train a classifier/regression model.	Species observations are often biased toward roads, trails, and accessible places.

Dynamic World: A Good Example of Earth Engine + AI

Dynamic World is a near-real-time, 10 m land use / land cover dataset created by Google and the World Resources Institute. It is available in Earth Engine as GOOGLE/DYNAMICWORLD/V1. The dataset provides probabilities and labels for nine classes: water, trees, grass, flooded vegetation, crops, shrub and scrub, built, bare, and snow/ice.

Dynamic World is useful because it shows how Earth Engine can host AI-derived environmental datasets. You do not have to train the deep-learning model yourself to use the output. But you still need to use it carefully: a class label is not the same as ecological truth, and the catalog recommends using probability thresholds when selecting pixels confidently belonging to a class.

```javascript // Dynamic World example: display most likely land-cover label. var roi = ee.Geometry.Point([31.5, -24.5]).buffer(10000); var dw = ee .ImageCollection("GOOGLE/DYNAMICWORLD/V1") .filterBounds(roi) .filterDate("2025-01-01", "2025-12-31"); var labelComposite = dw.select("label").mode().clip(roi); var dwVis = { min: 0, max: 8, palette: [ "#419BDF", // water "#397D49", // trees "#88B053", // grass "#7A87C6", // flooded vegetation "#E49635", // crops "#DFC35A", // shrub and scrub "#C4281B", // built "#A59B8F", // bare "#B39FE1", // snow and ice ], }; Map.centerObject(roi, 11); Map.addLayer(labelComposite, dwVis, "Dynamic World most common label"); ```

Use Dynamic World for screening and monitoring, then validate important conclusions with field data, higher-resolution imagery, expert interpretation, or locally trained models.

Machine Learning in Earth Engine

Earth Engine supports several levels of machine learning:

Built-in supervised classifiers: Earth Engine’s classifier package includes algorithms such as CART, Random Forest, Naive Bayes, and SVM.
Unsupervised clustering: Useful for exploring spectral groups when labelled training data are limited.
Pre-made AI-derived datasets: Dynamic World is a strong example of a deep-learning-derived land-cover product hosted in Earth Engine.
Custom models through Vertex AI: Earth Engine can send image or table data to models hosted on Vertex AI using{" "} ee.Model, then bring predictions back as images or tables.

For most conservation teams, the best starting point is not custom deep learning. Start with good data hygiene, clear labels, good reference data, and simple models. A carefully validated Random Forest classifier can be more useful than an impressive neural network trained on weak labels.

A simple supervised classification workflow

Choose classes: for example water, woodland, grassland, bare ground, cropland, built area.
Collect training polygons or points from field data or careful image interpretation.
Build predictor bands: Sentinel-2 bands, NDVI, NDWI, elevation, slope, texture, seasonality.
Split training and validation data.
Train a classifier.
Classify the image.
Calculate an error matrix and accuracy metrics.
Inspect errors visually and revise training data.

Never publish a classification map without reporting what data trained it, how accuracy was assessed, what classes were confused, and where the model is likely to fail.

Exports, Assets, and Apps

Earth Engine analysis often ends in one of four ways:

Output	Use it when	Example
Map layer	You want to inspect or demonstrate a result visually.	NDVI map, land-cover layer, forest-loss layer.
Chart	You want to show a time series or relationship.	Monthly NDVI trend for a reserve.
Table export	You need statistics in CSV, BigQuery, or another analysis tool.	Mean NDVI per restoration block per month.
Image export	You need a raster result for GIS, reports, modelling, or archiving.	GeoTIFF of classified land cover.
Earth Engine App	You want non-technical users to interact with your analysis.	Dashboard for fire history, water points, vegetation trend, or forest change.

Earth Engine Apps can be public or access-controlled depending on how you configure them. For team science, project-owned apps are useful because collaboration can be managed through a Google Cloud project rather than one person’s account.

Good Scientific Practice

Earth Engine makes large-scale analysis easier, but it does not remove the need for scientific caution. The most common mistakes are not technical; they are conceptual.

Do not confuse pixels with proof

A satellite pixel is a measurement or model output. It is not the same as a direct field observation. For example, a red pixel in a forest-loss layer may represent a stand-replacement disturbance, but the cause could be logging, fire, storm damage, mining, plantation harvesting, or classification error.

Always match scale to the question

If your study area is a small wetland, a 500 m MODIS product may be too coarse. If your study period is 40 years, Sentinel-2 is too recent on its own. If your question is species habitat quality, a single NDVI value is usually not enough.

Record uncertainty

Good Earth Engine outputs should state the dataset version, date range, spatial resolution, masks used, thresholds used, model assumptions, and validation method. For AI-derived products, include confidence or probability thresholds where available.

Use local knowledge

Field rangers, ecologists, farmers, community members, reserve managers, and local researchers often understand land-use history that satellite data alone cannot explain. The best science usually combines remote sensing with local evidence.

Protect sensitive information

When mapping threatened species, nests, dens, rhino habitat, pangolin sightings, rare plants, archaeological sites, or anti-poaching infrastructure, do not publish exact coordinates without a clear data-sensitivity review.

Conservation Use Cases

Earth Engine is especially useful for conservation because it can connect field operations to landscape context. A conservation team can combine satellite-derived layers with camera-trap stations, ranger patrols, wildlife sightings, restoration plots, water points, fences, fire scars, community land-use zones, and protected-area boundaries.

Use case	Earth Engine role	Field validation
Anti-poaching patrol planning	Map access routes, vegetation cover, burn scars, water availability, and recent disturbance.	Ranger intelligence, patrol tracks, incident reports.
Habitat monitoring	Track vegetation seasonality, woody cover, bare ground, water extent, and land-cover change.	Vegetation plots, ecological surveys, drone imagery.
Restoration monitoring	Compare before/after NDVI, canopy, wetness, bare ground, and fire history.	Species composition, survival counts, soil condition, invasive species checks.
Camera-trap site selection	Stratify sampling by habitat type, distance to water, slope, fire history, and land-cover class.	Ground scouting, access/safety checks, animal sign.
Wildlife corridor analysis	Map habitat continuity, settlement expansion, crop conversion, barriers, and riparian strips.	Movement data, spoor, telemetry, local knowledge.

Limits and Risks

Earth Engine is powerful, but it has important limits.

It is not a substitute for fieldwork. Satellite data cannot identify every species, explain every cause, or replace local ecological knowledge.
Clouds and shadows matter. Optical satellite images can be badly affected by clouds, smoke, haze, or shadows.
Resolution matters. A 10 m pixel, 30 m pixel, and 500 m pixel can tell different stories.
Dataset versions change. Re-run analyses when data products update, and cite exact versions.
Quotas and exports can fail. Very large tasks may hit memory, time, scale, or quota limits.
Commercial/noncommercial rules matter. Check access, billing, verification, and terms before operational or paid work.
Privacy and sensitive-location risks are real. Apps and map layers can expose information if published carelessly.

A Practical Learning Path

A beginner does not need to master everything at once. A good learning sequence is:

Run the Code Editor quickstart. Learn{" "} Map.addLayer(), print(), raster layers, vector points, and simple charts.
Learn Image and ImageCollection. Understand bands, metadata, date filters, spatial filters, and composites.
Learn masking. Clouds, shadows, no-data pixels, and quality bands are often the difference between good and bad analysis.
Learn reducers. Reducers let you calculate summaries over time, space, polygons, and regions.
Learn exports. Practise exporting small results before attempting national-scale maps.
Use one real local project. Pick a reserve, wetland, restoration site, or catchment you know well.
Add validation. Compare outputs with field notes, GPS points, drone images, camera-trap locations, or expert interpretation.
Publish carefully. Build an app only after checking privacy, licensing, and sensitive-location risks.

Earth Engine Cheat Sheet

Task	Common function	Example
Load one image	`ee.Image()`	`ee.Image('UMD/hansen/global_forest_change_2025_v1_13')`
Load time series	`ee.ImageCollection()`	`ee.ImageCollection('COPERNICUS/S2_SR_HARMONIZED')`
Filter by date	`filterDate()`	`.filterDate('2025-01-01', '2025-12-31')`
Filter by place	`filterBounds()`	`.filterBounds(roi)`
Calculate NDVI	`normalizedDifference()`	`image.normalizedDifference(['B8', 'B4'])`
Create a composite	`median()`, `mean()`, `mosaic()`	`collection.median()`
Summarise an area	`reduceRegion()`	`image.reduceRegion({reducer: ee.Reducer.mean(), geometry: roi, scale: 10})`
Add a layer to map	`Map.addLayer()`	`Map.addLayer(ndvi, vis, 'NDVI')`
Export image	`Export.image.toDrive()`	Save a GeoTIFF for GIS.
Export table	`Export.table.toDrive()`	Save regional summaries as CSV.

Example Export

Exports are handled as tasks. After running this code, open the Tasks tab in the Code Editor and click Run to start the export.

```javascript Export.image.toDrive({ image: ndvi, description: "ndvi_2025_roi", folder: "earth_engine_exports", fileNamePrefix: "ndvi_2025_roi", region: roi, scale: 10, maxPixels: 1e9, }); ```

Keep exports small while learning. Once your workflow is correct, scale up gradually.

Before You Trust an Earth Engine Result

Use this checklist before reporting a result:

Did I use the correct dataset version?
Did I read the dataset documentation?
Did I check band units and scale factors?
Did I mask clouds, shadows, no-data pixels, or low-quality observations?
Did I choose a resolution that matches my question?
Did I compare the result against field evidence or high-resolution imagery?
Did I record uncertainty and limitations?
Did I cite the dataset and Earth Engine correctly?
Did I avoid publishing sensitive coordinates or protected species locations?

Sources and Further Reading

Offline Communications for Wildlife Field Teams: LoRa, Radio, Satcom, Mesh, Starlink and Long-Range Connectivity

Sun, 14 Jun 2026 00:00:00 GMT

import BlogImage from "../../components/blog/BlogImage.astro"; import fieldCommsOverview from "../../assets/blog/offline-comms-for-wildlife-field-teams/30623992_aleksandr_sochnev.jpg"; import satelliteConnectivity from "../../assets/blog/offline-comms-for-wildlife-field-teams/33153_pixabay.jpg";

Remote conservation work has a simple communications problem: the places that most need monitoring are often the places with the least infrastructure. Rangers patrol valleys where mobile coverage disappears. Camera traps sit beyond Wi‑Fi. Research teams work from temporary camps. Anti-poaching teams need voice comms now, not when a cloud dashboard eventually syncs. Communities near protected areas need fast warnings when elephants, lions, buffalo or wildfire move toward people.

The answer is not one magic network. It is a layered communications stack. Voice radio keeps teams safe during patrols. LoRa and Meshtastic move small messages and location packets when there is no cell service. Satellite messengers and satellite phones provide an emergency backstop. Point-to-point Wi‑Fi and microwave links connect outposts. Cellular routers and directional antennas exploit weak coverage at the edge of the network. Starlink and other low-Earth-orbit satellite internet systems can turn a remote camp into an online operations centre. Field platforms such as EarthRanger or SMART then turn all those signals into an operational picture.

This article is written for conservation teams, field operations managers, wildlife researchers, reserve owners, ranger units and students building technology for field deployment. It explains what each communication layer does, where it works, where it fails, and how to combine the pieces into a practical system.

Source note: This post was researched from official Meshtastic, LoRa Alliance, The Things Network, Garmin, Iridium, Starlink, Ubiquiti, EarthRanger, African Parks, ICASA, GSMA and related conservation-technology sources. Source links are included throughout and listed again at the end.

The Big Idea: Build for Failure

A good field communications plan assumes things will break. Batteries die. Repeaters lose power. Vehicles park behind ridges. Rain degrades links. Solar panels get dusty. Antennas are mounted too low. Mobile towers go down. A satellite terminal cannot see the sky. A phone app works at base but not under tree cover. A ranger forgets to charge the device that was supposed to save the day.

That is why resilient field comms should be designed like aviation or emergency response: layered, simple, rehearsed and redundant. The goal is not to create the fanciest network. The goal is to answer four questions reliably:

Where is the team? Location and status.
Can the team call for help? Distress, injury, breakdown, security incident.
Can the control room coordinate a response? Voice, dispatch, map, incident log.
Can data get out eventually? Patrol reports, camera-trap alerts, collar data, sensor data, photos and field notes.

Different technologies answer different parts of that problem. Radio is still king for immediate group voice. LoRa is excellent for tiny low-power packets. Wi‑Fi bridges are good for fixed links. Cellular is cheap where coverage exists. Starlink is powerful for remote internet. Sat messengers are good for last-resort SOS and check-ins. None of them is enough alone.

Aleksandr Sochnev on Pexels`} />

The Communications Stack at a Glance

Layer	Best for	Typical payload	Strength	Weakness
VHF/UHF radio	Immediate patrol voice, dispatch, safety.	Voice, short text, GPS on digital systems.	Fast, simple, shared situational awareness.	Needs spectrum planning, repeaters and radio discipline.
LoRa / Meshtastic	Off-grid text, location beacons, simple team tracking, low-power sensors.	Small messages, GPS position, telemetry.	Cheap, low power, peer-to-peer, works without cell towers.	Low bandwidth, terrain-sensitive, not for voice or images.
LoRaWAN	Structured sensor networks across farms, reserves and research sites.	Telemetry from collars, gates, tanks, weather stations, traps.	Long range, low power, private/public network options.	Designed for IoT data, not team chat.
Satellite messenger	Emergency SOS, two-way text, check-ins for isolated teams.	Text, location, SOS, sometimes weather/photos/voice snippets.	Works far beyond mobile networks.	Subscription cost, slow messages, sky view needed.
Satellite phone / PTT	Remote command, voice call, emergency coordination.	Voice, SMS, low-rate data.	Human voice when nothing else works.	Expensive, operationally fragile if not charged/tested.
Starlink / LEO satellite internet	Remote base camps, mobile command posts, uploads, dashboards, video calls.	Broadband internet.	High bandwidth in remote areas.	Needs power, clear sky, regulatory approval and subscription.
Point-to-point Wi‑Fi	Fixed links between lodges, towers, gates, outposts and offices.	IP data, cameras, dashboards, VoIP, file sync.	High bandwidth and low recurring cost once installed.	Needs line of sight, towers, alignment and lightning protection.
Long-range cellular	Using weak edge coverage with proper routers, SIMs and antennas.	Normal internet, alerts, apps, photos, reports.	Cheap and familiar where coverage exists.	Coverage is patchy; boosters/repeaters are regulated.
Offline phone apps / P2P Wi‑Fi	Nearby device-to-device messages, local forms, data handoff.	Messages, forms, files, sync bundles.	Useful when the internet is down but phones are nearby.	Short range unless combined with other radios or access points.

1. VHF/UHF Radio: Still the Backbone of Field Safety

For ranger teams, two-way radio remains the most important layer. It is immediate, shared and operationally simple. When a vehicle breaks down, a rhino security unit changes route, a fire line shifts, or a dangerous animal is near a walking team, push-to-talk voice is faster than typing into an app.

Most reserves use some combination of VHF, UHF, analogue repeaters, digital mobile radio (DMR), trunked radio, vehicle radios and handheld radios. The details depend on terrain, licensing, budget and security requirements. In open savanna, VHF often travels well. Around buildings, vehicles and broken terrain, UHF can be useful. Digital systems can add cleaner audio, GPS tracking, emergency buttons, man-down alerts, text messaging and dispatch consoles.

The modern pattern is radio plus map. A digital radio system can show ranger positions in a control room, while an operations platform such as EarthRanger or SMART stores incidents, tracks patrols and integrates collar/sensor feeds. EarthRanger describes its role as bringing field data, wildlife, teams and technology together in real time, and African Parks describes using EarthRanger as a real-time operational tool for ranger patrols, wildlife monitoring and human-wildlife conflict response.

Radio is not glamorous, but it is often the difference between a good technology deployment and a dangerous one. If a new drone, AI camera, LoRa tracker or satellite terminal does not improve the radio command picture, it may not improve field safety.

Good field practice

Give every team a simple call sign.
Keep a daily radio check schedule.
Use standard words for emergency, medical, vehicle, wildlife and security incidents.
Do not share sensitive wildlife locations on open or poorly controlled channels.
Mount repeaters high, power them with solar/battery backup and protect them from lightning.
Train new rangers on radio discipline before they need it in an emergency.

2. LoRa and LoRaWAN: Tiny Packets Over Long Distances

LoRa is a radio modulation designed for low-power, long-range communication. LoRaWAN is the network protocol built on top of LoRa for structured sensor networks. The Things Network explains the distinction clearly: LoRa is the physical radio technique, while LoRaWAN defines how devices communicate, how messages are formatted and how network servers handle devices.

In conservation, LoRa-style systems are useful when you need small packets, not broadband. Think water-tank levels, fence alarms, gate sensors, weather stations, acoustic sensor health checks, trail-counter data, camera-trap status messages, GPS pings from vehicles, or low-frequency collar telemetry. LoRa is not for sending images, voice or video.

The LoRa Alliance describes LoRaWAN as a low-power wide-area technology for affordable sensor connectivity, including private, public and hybrid network options. The Things Network notes that LoRaWAN is suited to small payloads over long distances, often more than 10 km in rural areas under good conditions, while end devices can run for years because they transmit very little data.

Where LoRa shines

Low power: small solar or battery-powered devices can last a long time.
Long distance: good range when antennas are high and terrain is kind.
Low cost: devices and gateways can be much cheaper than cellular or satellite hardware.
Local ownership: a reserve can run a private network without waiting for a mobile operator.

Where LoRa fails

It cannot move much data.
Dense bush, ridges and valleys can block links.
Duty-cycle and local spectrum rules matter.
Network planning is still engineering, not magic.

3. Meshtastic: LoRa for Off-Grid Team Messaging

Meshtastic is an open-source project that uses inexpensive LoRa radios as a long-range, off-grid communication platform for places without reliable communications infrastructure. Instead of connecting through a tower, devices can relay messages through other devices in a mesh. A phone connects to a small radio over Bluetooth; the radio sends short messages and GPS position packets over LoRa.

Meshtastic is attractive for wildlife teams because it is cheap, hackable and independent. A research team can place router nodes on hills, vehicles or camp buildings. Rangers can carry small nodes. Volunteers can send text check-ins without cell coverage. A field technician can see rough positions of nearby team members. A base-station node can bridge messages to MQTT or an operations system when internet is available.

It is important not to oversell it. Meshtastic is not a replacement for licensed mission-critical voice radio. It does not give you broadband. It is not guaranteed emergency infrastructure. The mesh must be planned, tested and maintained. Antenna height matters more than marketing range claims. A node in a backpack under tree cover is not the same as a node on a hill mast.

Security note

Meshtastic supports AES256-CTR encryption for message payloads, but its documentation warns that the default primary channel uses a known key and should be changed for real privacy. Packet headers are also not encrypted because nodes need to relay traffic. Treat Meshtastic as useful operational messaging, not as a covert anti-poaching command system unless it has been configured and threat-modelled carefully.

Best conservation uses

Volunteer groups spread across a reserve.
Research assistants working within a known valley or transect network.
Backup check-ins for hiking, snare patrols or camera-trap servicing.
Low-cost vehicle or team beacons inside a private mesh.
Education projects showing how radio mesh networks work.

Bad uses

Sending photos, videos or large files.
Replacing emergency radio for armed or high-risk patrols.
Assuming encryption is safe without changing default keys.
Using imported radios without checking local frequency plans and type approval.

4. Satellite Messengers: The Personal Safety Backstop

A satellite messenger is a small device for two-way text, location sharing and SOS. It is not a replacement for a radio network, but it is very useful when one person or one team is beyond all terrestrial coverage. Garmin’s inReach products, for example, advertise two-way text messaging and SOS functionality with a satellite subscription, while Iridium positions its satellite network for mission-essential communications across the planet.

In the field, the best use is boring and disciplined: each remote team sends a scheduled check-in, the control room knows when to escalate, and the device is kept charged, labelled and paired to the correct phone. An SOS button is not a plan by itself. The plan is who receives the alert, who can respond, which helicopter/vehicle/medical contact is called, and what happens if the satellite message is delayed.

Best uses

Solo researchers, remote camera-trap technicians or anti-poaching observation teams.
Emergency SOS in areas with no radio or cell coverage.
Simple daily check-ins from remote camps.
Backup communications for expedition vehicles.

Limitations

Messages are slow compared with cellular chat.
Subscriptions must be active and paid.
Tree cover, cliffs, vehicles and buildings can block sky view.
Staff must know how to trigger SOS and how to cancel false alarms.

5. Satellite Phones and Push-to-Talk: Voice Beyond the Tower Map

Satellite phones and satellite push-to-talk systems are used when a team needs voice communication far outside terrestrial coverage. They are more expensive than handheld radios and satellite messengers, but in remote marine work, desert surveys, transfrontier patrols, mountain deployments or disaster response they can be worth the cost.

The main decision is whether you need occasional emergency voice calls or structured operational group comms. A single satphone at base camp is useful for escalation. Satellite push-to-talk can support distributed teams, but cost, training, device management and operating procedures become more serious.

Operational advice

Keep numbers and procedures printed in the vehicle and in the control room.
Test calls monthly, not only during emergencies.
Carry a spare battery or power bank.
Store devices where staff can actually access them.
Do not assume satellite voice is private or immune to disruption; use it for safety and coordination, not sensitive tactical details unless properly secured.

6. Starlink and LEO Satellite Internet: Broadband for Camps and Command Posts

Low-Earth-orbit satellite internet has changed what a remote field camp can do. A site that once had only VHF radio can now upload camera-trap batches, sync SMART or EarthRanger data, run video calls, download maps, send drone imagery, update software and back up field photos.

Starlink is the most visible example. SpaceX’s Direct to Cell materials describe a separate satellite-to-mobile path for ordinary LTE phones through operator partnerships, while standard Starlink terminals provide broadband through a dish/terminal. These are different products. For conservation operations today, the practical use case is usually a Starlink terminal at a lodge, reserve office, research camp, mobile command vehicle or field ranger base.

The catch is power, regulation and governance. A terminal needs reliable electricity and clear sky. It needs a legal service plan in the country of operation. It can become a single point of failure if every dashboard, camera, patrol report and emergency workflow depends on it. It can also create cyber-security and staff-management issues if the field internet becomes an unmanaged open Wi‑Fi network.

A good Starlink deployment looks like this

Solar/battery or generator backup sized for the terminal, router and key devices.
Router with firewall rules, separate staff/ops/guest networks and bandwidth controls.
Offline workflow for when the link drops.
Mounting that survives wind, dust, animals and curious humans.
Documented legal approval and billing ownership.

Southern Africa note

Starlink availability in southern Africa changes by country and depends on licensing. Reuters reported Zimbabwe approval in 2024 and Lesotho licensing in 2025, while South Africa has faced licensing obstacles. For a South African reserve or NGO, do not assume a kit purchased elsewhere is lawful to import, resell or operate. Check the official availability map, the local communications regulator and your organisation’s compliance requirements before deployment.

Pixabay on Pexels`} />

7. Point-to-Point Wi‑Fi and Microwave: The Fixed-Link Workhorse

Point-to-point wireless links are one of the most useful underappreciated tools in conservation tech. If you can see from one hill, lodge or tower to another, you may be able to move high-speed IP data across kilometres without trenching fibre or relying on a mobile operator. Vendors such as Ubiquiti publish point-to-point and long-range Wi‑Fi configuration guides and devices with directional antennas for fixed wireless links.

This is not the same as normal campsite Wi‑Fi. Long-range Wi‑Fi uses directional antennas aimed at each other. It needs line of sight, careful alignment, clean power, correct channels, weatherproofing and lightning protection. Done well, it can connect a gate camera, ranger post, rhino boma, remote office, lodge, airstrip, hill repeater or Starlink backhaul to the main network.

Best uses

Linking reserve headquarters to a remote gate or tower.
Backhauling camera feeds from a waterhole or fence line.
Connecting a hilltop radio repeater site to the internet.
Sharing Starlink or fibre from one site to nearby operational buildings.
Connecting LoRaWAN gateways or EarthRanger hardware integrations.

Failure modes

No true line of sight.
Antennas mounted too low.
Wrong frequency choice in a noisy area.
Lightning damage.
Poor grounding.
Water ingress.
No one knows the admin password two years later.

8. Long-Range Cellular: Use the Edge of the Network Properly

In many reserves, mobile coverage is not absent; it is just weak, patchy or high on ridgelines. A normal phone may fail, while an industrial LTE/5G router with an outdoor directional MIMO antenna, good cable and the right SIM can work. This is often the cheapest useful data link when coverage exists.

Cellular is excellent for camera alerts, staff messaging, app sync, remote sensors, telemetry gateways and visitor operations. It is also familiar: everyone understands SIM cards and data bundles. But it can collapse during power outages, storms, tourist peaks or tower maintenance. It also depends on the operator’s coverage, bands, fair-use policy and roaming rules.

Practical advice

Survey signal using multiple networks, not one phone on one day.
Test at height: roof, mast, koppie, water tower, repeater site.
Use directional outdoor antennas for fixed sites.
Use dual-SIM routers where possible.
Do not use illegal boosters or repeaters; they can interfere with licensed networks.
Log uptime and data use before declaring the link reliable.

For IoT, LTE‑M and NB‑IoT are also important. GSMA describes LTE‑M and NB‑IoT as low-power wide-area cellular technologies designed for IoT devices with lower complexity, extended coverage and long battery life. They are not everywhere, and they require operator support, but they matter for future conservation telemetry where licensed-spectrum reliability is preferred over unlicensed LoRa.

9. Offline Phone Apps and Peer-to-Peer Wi‑Fi

Sometimes the problem is not long-distance communication. It is local data movement. A ranger has forms on a phone. A camera-trap team has photos on a laptop. A researcher has species notes. The base camp internet is down, but the team is physically together.

Peer-to-peer tools can help. Briar, for example, describes itself as a secure peer-to-peer messaging app that can connect via Bluetooth, Wi‑Fi or Tor and can sync without central servers. In conservation, the same principle matters beyond chat: field apps should be able to store data offline, sync locally, export files and recover from intermittent connectivity.

The range is usually short. Phone-to-phone Bluetooth or Wi‑Fi is not a reserve-wide radio network. But for camps, vehicles, training groups, data handovers, emergency notebooks and offline-first patrol systems, local sync is valuable.

Good uses

Offline patrol forms that sync when a team returns to base.
Local sharing of maps, SOPs, incident forms and species guides.
Short-range peer messaging during internet outages.
Bulk transfer from field laptop to office server without cloud connectivity.

10. The Operations Layer: Turning Signals Into Decisions

Communications hardware is only useful if people know what to do with the information. A LoRa sensor that reports a broken fence, a Starlink link that syncs patrol forms, a radio GPS ping, a satellite check-in and a camera-trap alert all need to land somewhere. That somewhere may be a control room whiteboard, a WhatsApp group, a radio dispatcher, a spreadsheet, SMART, EarthRanger, QGIS, a custom database or an incident management system.

EarthRanger says it integrates field data, technology, wildlife and teams in real time, with more than 900 supported conservation sites, more than 23,000 animals tracked via GPS and more than 150 integrations. African Parks describes EarthRanger as combining ranger patrols, remote imaging and sensors into a real-time tool for anti-poaching, wildlife monitoring and human-wildlife conflict response.

The important design question is not “What network should we buy?” It is “What decision should this network help us make?” A fence alarm should create a maintenance task. A person-detection camera should create a patrol response. A ranger SOS should trigger a medical protocol. A collar breach should alert the human-wildlife conflict team. A failed gateway should create a technician ticket.

Reference Architectures

Small reserve or conservancy

Need	Suggested layer	Notes
Daily patrol voice	VHF/UHF handhelds + one repeater if terrain requires it.	Start with voice safety before adding data toys.
Remote team check-ins	Meshtastic or satellite messenger.	Use Meshtastic where teams are within a planned mesh; satellite for true remote fallback.
Office internet	Cellular router, fibre, WISP or Starlink if legal/available.	Use dual-WAN if budget allows.
Gate or waterhole camera	Cellular camera or point-to-point Wi‑Fi.	Use Wi‑Fi bridge when fixed site has line of sight.
Sensor alerts	LoRaWAN gateway + sensors.	Good for tanks, gates, fences and weather.

Large protected area

Need	Suggested layer	Notes
Mission-critical patrol comms	Licensed digital radio network with repeaters, GPS and dispatch.	Design professionally; test dead zones and incident workflows.
Outpost connectivity	Point-to-point microwave/Wi‑Fi, cellular router or Starlink.	Use the cheapest reliable backhaul per site.
Ranger/personnel safety	Radio emergency button + satellite messenger for isolated teams.	Do not depend on one path.
Wildlife and asset tracking	GPS collars, LoRaWAN, cellular IoT, satellite IoT.	Choose based on species, terrain, update rate and battery life.
Operations dashboard	EarthRanger/SMART/custom GIS.	The dashboard is only useful if someone is assigned to act on alerts.

Mobile research expedition

Need	Suggested layer	Notes
Vehicle-to-vehicle	VHF/UHF mobile radios or handhelds.	Keep this simple and trained.
Camp internet	Starlink, cellular router or local WISP.	Power budget matters more than bandwidth claims.
Team text and location	Meshtastic nodes with spare batteries.	Works best when team stays in a known operating area.
Emergency fallback	Satellite messenger or satphone.	Test before departure.
Data management	Offline-first forms + local backup + delayed cloud sync.	Do not make cloud connectivity a requirement for data capture.

Regulatory Reality: Radios Are Not Just Gadgets

Every country regulates radio spectrum. In South Africa, ICASA is the communications regulator. ICASA explains that radio equipment generally needs type approval before it may be used, supplied, sold, leased or imported, unless exempted. ICASA also states that equipment used under spectrum licences must be type approved and that radio frequency spectrum licensing follows the national frequency plan and relevant regulations.

This matters for all the technologies in this article: VHF/UHF radios, repeaters, LoRa devices, Wi‑Fi bridges, cellular routers, boosters, satellite terminals and marine radios. Some operate in licence-exempt bands but still have power, antenna, duty-cycle and type-approval constraints. Some need licences. Some are legal in one country and not in a neighbour. Some devices sold online are configured for the wrong region.

The practical rule: before deploying at scale, ask three questions. Is the device type-approved? Is the frequency legal for this use? Is the power/antenna configuration compliant? If the answer is unclear, get help from a licensed radio dealer, WISP, network engineer or regulator-facing installer.

Security, Privacy and Sensitive Wildlife Data

Conservation communications often carry sensitive information: rhino locations, patrol routes, ranger positions, informant reports, fence breaches, camera alerts, anti-poaching movements and community conflict incidents. A network that works technically can still be unsafe operationally.

Minimum rules

Do not broadcast exact rhino, pangolin, elephant-carcass or snare locations on open channels.
Separate guest internet from operations internet.
Change default passwords and default Meshtastic channel keys.
Keep device inventories: serial number, SIM, owner, firmware, antenna, charger, assigned user.
Remove access when volunteers or contractors leave.
Do not let every alert go to every WhatsApp group.
Have a clear data-retention rule for ranger tracks and sensitive incidents.

Also remember the human side. A GPS-tracked radio can improve ranger safety, but it can also feel like surveillance if introduced badly. Explain why tracking exists, who can see it, when it is used, and how it protects staff.

The Hidden Layer: Power

Most failed field communication systems are not radio failures. They are power failures. A brilliant hilltop repeater is useless if the solar panel is shaded. A Starlink terminal is useless if the inverter trips. A LoRaWAN gateway is useless if the battery dies after three cloudy days. A satellite messenger is useless in the drawer at base camp.

Design power before you design the network. Count watts. Size batteries for bad weather. Protect against lightning. Label chargers. Keep spare cables. Use low-voltage disconnects. Log outages. Give someone ownership of every remote power system.

Technology	Power profile	Field implication
Handheld VHF/UHF radio	Battery powered, daily charging.	Needs charging discipline and spare batteries.
Radio repeater	Continuous site power.	Needs solar/battery backup and lightning protection.
Meshtastic node	Low power.	Can run on battery/solar if configured carefully.
LoRaWAN gateway	Moderate continuous power.	Needs stable power and backhaul.
Cellular router	Moderate continuous power.	Good for solar sites if sized properly.
Point-to-point Wi‑Fi	Low-to-moderate continuous power.	Often practical on solar towers.
Starlink	Higher continuous power.	Needs serious power planning for off-grid camps.

How to Choose: A Simple Decision Framework

Start with the job, not the gadget.

Question	If yes	Likely technology
Do people need immediate group voice?	Safety or patrol coordination depends on it.	VHF/UHF radio, digital radio, repeaters.
Do you only need tiny messages or GPS pings?	Low data, low power, low cost.	LoRa, Meshtastic, LoRaWAN.
Is there weak but usable mobile coverage?	Coverage improves at height or with an antenna.	Industrial cellular router, directional antenna, dual SIM.
Can two fixed points see each other?	You need high-bandwidth site-to-site data.	Point-to-point Wi‑Fi or microwave.
Do you need broadband far from infrastructure?	Camp, command vehicle or office needs real internet.	Starlink or other satellite internet where legal.
Could a person be alone beyond all coverage?	Emergency rescue path is required.	Satellite messenger or satphone.
Do devices need to sync locally without internet?	Teams are physically near each other.	Offline-first apps, peer-to-peer Wi‑Fi, local servers.

Common Mistakes

Buying hardware before mapping the problem. Walk the terrain, mark dead zones and define workflows first.
Ignoring spectrum law. Imported radios can be illegal or badly configured for local bands.
Assuming Starlink replaces radio. Broadband is not the same as mission-critical push-to-talk.
Putting LoRa nodes too low. Antenna height and line of sight are everything.
Using consumer Wi‑Fi like infrastructure. Field networks need weatherproofing, grounding and access control.
Forgetting power. Power is the network.
Not training staff. A ranger under stress should not be learning a menu system for the first time.
No escalation protocol. An SOS alert without a response plan is just a panic button.
Over-sharing sensitive locations. Communications can create poaching risk if badly governed.

Deployment Checklist

Before buying

Map patrol routes, camps, gates, ridges, valleys and high sites.
Define use cases: voice, SOS, sensor, internet, image upload, location, dashboard.
Survey existing radio, cellular and Wi‑Fi coverage.
Check legal spectrum, type approval and satellite service availability.
Estimate power at every fixed site.

Before deployment

Label every device, battery, charger, SIM and antenna.
Write a one-page SOP for each device type.
Train users in normal and emergency use.
Run a field test in the actual terrain.
Create a dead-zone map.
Change default passwords, keys and admin accounts.

After deployment

Log outages and near misses.
Update maps after every significant field test.
Replace failed cables and batteries before they become emergencies.
Review incident response monthly.
Audit who still has access to apps, dashboards and devices.

Where This Is Going

The future of conservation communications is not one network. It is convergence. Radios will carry GPS and emergency data. LoRa sensors will bridge into dashboards. Satellite-to-cell will make ordinary phones useful in more dead zones. Starlink and similar systems will make remote camps feel online. Edge AI devices will send only the important alerts instead of every image. Field platforms will increasingly combine patrol, sensor, collar, camera, acoustic and satellite data into a single map.

The best conservation teams will not be the ones with the most expensive equipment. They will be the ones with the clearest operational design: which data matters, who receives it, who acts, what happens when the link fails, and how the system protects both wildlife and people.

In the field, good communications are not about technology for its own sake. They are about trust. A ranger trusts that the radio will work. A researcher trusts that the check-in will reach base. A community member trusts that a conflict alert will be answered. A reserve manager trusts that an incident will not disappear into a dead zone. That is the real goal: fewer blind spots, faster response, safer people and better decisions for wildlife.

Sources and Further Reading

Planetary Boundaries Explained

Sun, 14 Jun 2026 00:00:00 GMT

import { Image } from "astro:assets"; import BlogImage from "../../components/blog/BlogImage.astro"; import PlanetaryB2 from "../../assets/blog/planetary-boundaries/30596208_zelch_csaba.jpg"; import PlanetaryB1 from "../../assets/blog/planetary-boundaries/12498815_zelch_csaba.jpg"; import earthHorizon from "../../assets/blog/planetary-boundaries/30596227_zelch_csaba.jpg"; import doughnut from "../../assets/blog/planetary-boundaries/planetary-boundaries-doughnut.svg"; import earthFromSpace from "../../assets/blog/planetary-boundaries/32961160_zelch_csaba.jpg"; import { CO2Chart, TemperatureChart } from "@components/blog/charts";

For roughly 10,000 years, Earth stayed inside an unusually stable Holocene envelope. Climate, water, forests, oceans, nutrient cycles, and the biosphere varied — sometimes dramatically at regional scale — but the planetary state remained stable enough for agriculture, cities, and every civilisation you have ever heard of to emerge.

That stability is now gone. Human activity has pushed Earth beyond the conditions that sustained our species' entire recorded history.

In 2009, a team of 28 Earth system scientists led by Johan Rockström proposed a way to measure exactly how far we have pushed. They identified nine critical processes that regulate the planet's stability and asked a simple question: where are the guardrails?

They called it the planetary boundaries framework. In their first assessment, three boundaries were already breached. By 2015, four. By 2023, six. As of the 2025 Planetary Health Check,{" "} seven of nine boundaries are transgressed.

Here is every boundary, what it measures, the specific numbers, and where we stand today.

Data note: the status table uses the latest planetary-boundary assessment values available from the 2025 Planetary Health Check and the 2023 Science Advances update. Continuous observations, such as atmospheric CO₂, change month by month. Where those live data are useful, they are called out separately rather than mixed into the boundary assessment tables.

Zelch Csaba on Pexels`} />

The Framework at a Glance

The nine boundaries cover the biophysical systems that kept Earth in a Holocene-like state for the past 10 millennia. Each has a quantitative control variable, a boundary value, and an assessed value. The boundaries are set at the lower end of a zone of increasing risk — a precautionary approach. Cross the boundary and you enter a yellow-to-red zone where the probability of non-linear, irreversible change rises. Go deep enough and you hit the purple zone: high risk with high confidence.

The big update since the 2023 framework paper is ocean acidification. In 2023, six of nine boundaries were assessed as transgressed and ocean acidification was still described as close to the boundary. The 2025 Planetary Health Check assesses it as breached for the first time, lifting the count to seven of nine. The two boundaries still inside the global safe zone are stratospheric ozone depletion and atmospheric aerosol loading, although aerosols remain regionally dangerous over parts of South and East Asia.

The nine planetary boundaries — 2025 status, 7 breached, 2 nearing, 1 safe

The planetary boundaries framework. Green = safe. Yellow to red = zone of increasing risk. Purple = high risk. Seven wedges now extend beyond the safe zone. Credit: Azote for Stockholm Resilience Centre, based on Sakschewski and Caesar et al. 2025. Licensed under CC BY-NC-ND 3.0.

Zelch Csaba on Pexels`} />

1. Climate Change BREACHED

What it measures: Human disruption of Earth's energy balance through greenhouse gas emissions. Two control variables: atmospheric CO₂ concentration and total radiative forcing.

Variable	Boundary	Assessed value	Preindustrial
Atmospheric CO₂	350 ppm	423 ppm	280 ppm
Radiative forcing	+1 W/m²	+2.97 W/m²	0

The 350 ppm boundary corresponds to roughly 1°C of warming — slightly below the Paris Agreement's 1.5°C target. The planet is now in the high-risk zone on both measures. Climate change is designated a core boundary: its transgression amplifies risk across every other boundary in the framework.

One caveat matters: the table above is the boundary-assessment view, not a live CO₂ dashboard. NOAA's Mauna Loa record reached 432.34 ppm in May 2026, compared with 430.51 ppm in May 2025. Mauna Loa is a seasonal, site-specific monthly reading; the planetary-boundary assessment uses broader global mean metrics. The conclusion is the same either way: we are far beyond 350 ppm.

2. Biosphere Integrity BREACHED

What it measures: The health of Earth's living systems. Two dimensions: genetic diversity (measured as extinction rate) and functional integrity (measured as human appropriation of net primary production, or HANPP).

Dimension	Variable	Boundary	Assessed value
Genetic diversity	Extinction rate (E/MSY)	<10	>100
Functional integrity	HANPP (% of Holocene NPP)	<10%	30%

HANPP measures the share of the planet's photosynthetic energy that humanity consumes or displaces. Through agriculture, forestry, grazing, and urbanisation, we now appropriate roughly 30% of the biosphere's energy flow — triple the boundary. This boundary was crossed in the late 19th century, a generation before anyone noticed. Of an estimated 8 million animal and plant species, roughly 1 million are threatened with extinction. The 2024 Living Planet Index reports a 73% average decline in monitored vertebrate wildlife populations between 1970 and 2020; that is a population-index measure, not a count of individual animals lost.

3. Land System Change BREACHED

What it measures: Conversion of natural ecosystems — primarily forests — to agricultural and urban land.

Variable	Boundary	Assessed value
Global forest area intact	75% (85% tropical, 85% boreal, 50% temperate)	59%

Forests are the land biome with the strongest coupling to the climate system. They regulate carbon, water, and energy fluxes at the planetary scale. Land conversion also connects directly to the freshwater and biosphere boundaries: clear forests, and you alter rainfall recycling, soil moisture, habitat, and carbon storage at once.

4. Freshwater Change BREACHED

What it measures: Human modification of the entire terrestrial water cycle — surface and groundwater (blue water) and soil moisture available to plants (green water). The 2023 update fundamentally revised this boundary to capture far more than simple water consumption.

Component	Variable	Boundary	Assessed value
Blue water	% land area with streamflow deviations	12.9%	22.6%
Green water	% land area with soil moisture deviations	12.4%	22.0%

Under the old metric — 4,000 km³/year of consumptive water use — freshwater appeared safely within limits. The revised assessment split the system into blue water and green water, revealing that blue-water deviations crossed the boundary around 1905 and green-water deviations around 1929. We simply had not been measuring the right thing.

5. Biogeochemical Flows BREACHED

What it measures: Disruption of natural nitrogen and phosphorus cycles through industrial fertiliser production and agricultural runoff.

Element	Variable	Boundary	Assessed value
Nitrogen	Industrial N fixation (Tg/yr)	62	165
Phosphorus (regional)	P to erodible soils (Tg/yr)	6.2	18.2

The Haber-Bosch process feeds roughly half of humanity. It has also doubled the global nitrogen cycle. Excess nitrogen and phosphorus from fertilisers drain into waterways, driving eutrophication, oxygen loss, and hundreds of coastal dead zones. This is one of the hardest boundaries politically: reducing excess nutrient pollution is essential, but doing it badly would threaten food security. The solution is not simply “less fertiliser everywhere”; it is better nutrient efficiency, lower losses, circular nutrient recovery, and farming systems that can feed people without overwhelming rivers, lakes, and coasts.

6. Ocean Acidification BREACHED

What it measures: Decreasing pH of ocean surface waters as they absorb atmospheric CO₂. The 2025 Planetary Health Check assessed this boundary as transgressed for the first time.

Variable	Boundary	Assessed value	Preindustrial
Aragonite saturation state (Ω)	2.86	2.84	3.44

Aragonite is the form of calcium carbonate that corals, molluscs, pteropods, and shell-building plankton use. The boundary is set at roughly 80% of preindustrial saturation. At 2.84 — just below 2.86 — the 2025 assessment places the ocean outside the safe zone. Surface ocean pH has already dropped by about 0.1 units since the industrial era, corresponding to roughly a 30–40% increase in acidity. A separate 2025 study found that the boundary may have entered its uncertainty range by 2020, which is why this should be read as a delayed formal diagnosis rather than a sudden one-year collapse.

Acidification is not acting alone. It combines with marine heatwaves, deoxygenation, nutrient runoff, and overfishing. NOAA reports that the fourth global coral bleaching event, which ran from early 2023 to mid-2025, exposed 84% of the world's coral reef area to bleaching-level heat stress across the Pacific, Atlantic, and Indian Ocean basins.

Zelch Csaba on Pexels`} />

7. Atmospheric Aerosol Loading NOT BREACHED (globally)

What it measures: Microscopic particles from both natural and human sources that affect climate, monsoon systems, and human health.

Variable	Boundary	Assessed value (global)
Interhemispheric AOD difference	0.1	0.063

Globally within limits — but the global number hides regional transgression. South Asia (AOD ~0.3–0.35) and East China (AOD ~0.4) both exceed the regional boundary of 0.25. Aerosol loading can disrupt monsoon rainfall patterns that billions of people depend on for agriculture. It also complicates climate communication: some reflective aerosols have masked a portion of greenhouse warming, so cleaning the air improves health while revealing warming that pollution had been temporarily hiding.

Zelch Csaba on Pexels`} />

8. Stratospheric Ozone Depletion NOT BREACHED

What it measures: Thinning of the stratospheric ozone layer that shields us from UV radiation.

Variable	Boundary	Assessed value	Preindustrial
Stratospheric O₃ (Dobson Units)	277 DU	285.7 DU	290 DU

This is the single success story in the framework, and it proves that planetary-scale environmental problems can be solved. The Montreal Protocol (1987) phased out ozone-depleting substances, and the WMO reports that production and consumption of controlled ozone-depleting substances have been cut by more than 99%. If current policies remain in place, ozone is expected to recover to 1980 values by around 2040 for most of the world, 2045 over the Arctic, and 2066 over Antarctica. It remains depleted over Antarctica during Austral spring, but globally the boundary is respected.

9. Novel Entities BREACHED

What it measures: Synthetic chemicals and human-made materials released into the environment without adequate testing for Earth system impacts.

Variable	Boundary	Assessed value
% synthetics released without safety testing	0%	>0%

Over 350,000 synthetic chemicals and mixtures are on the global market. Under the EU REACH regulation, approximately 80% of registered chemicals had been in use for more than 10 years without completing a safety assessment. Plastic production exceeds 400 million tonnes per year and is still rising. The "cocktail effects" of chemical mixtures in the environment are poorly understood. This boundary is not a neat threshold like 350 ppm CO₂; it is a warning that the rate, volume, and novelty of chemical release have outrun society's ability to test, monitor, and govern systemic risk.

Status Summary

Boundary	Control Variable	Boundary Value	Assessed value	Status
Climate Change	CO₂ / Radiative forcing	350 ppm / +1 W/m²	423 ppm / +2.97	High risk
Biosphere Integrity	Extinction rate / HANPP	<10 E/MSY / <10% HANPP	>100 E/MSY / 30% HANPP	High risk
Land System Change	Forest area intact	75%	59%	Increasing risk
Freshwater Change	Blue/green water deviation	12.9% / 12.4%	22.6% / 22.0%	Increasing risk
Biogeochemical Flows	N fixation / P to soil	62 / 6.2 Tg/yr	165 / 18.2 Tg/yr	High risk
Ocean Acidification	Aragonite saturation	2.86 Ω	2.84 Ω	Just breached (2025)
Aerosol Loading	Interhemispheric AOD	0.1	0.063	Safe (globally)
Ozone Depletion	Stratospheric O₃	277 DU	285.7 DU	Safe
Novel Entities	Untested synthetics released	0%	>0%	Breached

What Happens When Boundaries Interact

These nine systems do not operate in isolation. They are coupled — sometimes in ways that amplify risk, sometimes in ways that mask it.

Cascading effects are the rule, not the exception. The framework identifies climate change and biosphere integrity as core boundaries — transgress either one and the safe operating space for every other boundary shrinks. Some examples:

Climate → Biosphere: Warming drives species extinction, which reduces ecosystem carbon storage, which accelerates warming. A feedback loop that amplifies both.
Land use → Freshwater: Deforestation reduces evapotranspiration and rainfall recycling. In the Amazon, this can push forest regions toward hotter, drier conditions, increasing fire risk and weakening the forest's carbon sink.
N/P flows → Ocean acidification → Biosphere: Nutrient runoff fuels algal blooms. When algae decompose, they consume oxygen, creating dead zones. Add warming and acidification, and marine ecosystems face three simultaneous, mutually reinforcing stressors.
Aerosols → Climate: Aerosols have masked roughly 0.5°C of warming. As countries reduce coal burning and aerosol pollution (good for health, good for air quality), this masking effect diminishes — unmasking additional warming. The clean air we want comes with a climate debt.
All boundaries → Tipping points: Each boundary carries its own tipping elements. The Greenland ice sheet, the West Antarctic ice sheet, the Amazon, the Atlantic overturning circulation, permafrost carbon — each has a threshold beyond which change becomes self-sustaining and essentially irreversible on human timescales.

What This Is Not Saying

Crossing a planetary boundary does not mean the world ends the next morning. It means the system has moved outside the Holocene-like safe operating space and into a zone where risks rise, feedbacks become harder to predict, and recovery becomes more expensive. Think of it as crossing a medical risk threshold: the diagnosis is serious before the organ fails.

It also does not mean every place is equally responsible or equally exposed. Some boundaries are global, like climate and ocean acidification. Others are deeply regional, like freshwater, aerosols, land-system change, and nutrient runoff. A globally safe number can still hide local danger, while a globally breached boundary can still be driven disproportionately by a subset of countries, industries, and consumption patterns.

This is why newer work on safe and just Earth system boundaries matters. It asks not only what keeps Earth stable, but what prevents significant harm to people. In that assessment, seven of eight quantified global safe-and-just boundaries were already crossed, and many local hotspots face several transgressions at once.

Is This Framework Useful?

An honest assessment.

What the framework does well

It treats the Earth as one system. Most environmental policy addresses problems in isolation — climate here, biodiversity there, water somewhere else. The framework forces integration. It makes explicit what scientists have long understood: the planet is a single, coupled system.
It provides quantitative targets. "Protect the environment" is not an operational goal. "Keep atmospheric CO₂ below 350 ppm" is. The framework gives policymakers, corporations, and civil society something to measure against.
It's being used. The EU's 8th Environment Action Programme is legally bound to the framework. Several nations have produced planetary boundary assessments. The World Business Council for Sustainable Development incorporated it into Vision 2050. This is not an academic curiosity — it is becoming operational infrastructure for environmental governance.
It communicates clearly. The doughnut diagram — nine coloured wedges radiating from a safe green centre — is one of the most effective visualisations in environmental science. You can look at it and immediately understand: green good, red bad, we're in the red.

Limitations

Global aggregation hides regional reality. The aerosol boundary is globally safe but regionally breached over South and East Asia, where over 2 billion people live. A global "safe" reading can be dangerously misleading.
Some boundaries are poorly quantified. Novel entities (chemical pollution) and aerosol loading remain difficult to pin to a precise planetary threshold. The 0% boundary for untested synthetics is aspirational, not empirical.
It doesn't address social foundations. The original framework is purely biophysical. Kate Raworth's Doughnut Economics (2012) addressed this by adding an inner ring representing minimum social standards — food, water, health, education, political voice. A safe planet that is unjust is not a stable planet.
Boundary interactions are not well-modelled. The framework acknowledges that boundaries interact, but comprehensive Earth system models that capture all nine simultaneously do not yet exist. Our understanding of cascading effects is largely qualitative.
Nitrogen is genuinely difficult. Roughly 4 billion people are alive today because of synthetic nitrogen fertiliser. The nitrogen boundary is deeply transgressed — but pulling back is not a simple technical problem. It's a problem of how humanity feeds itself.

The framework's authors have been clear from the start: these are "rough, first estimates only, surrounded by large uncertainties and knowledge gaps" (Rockström et al., 2009). That does not make them useless. It makes them honest.

Who Uses This Framework

Sector	Examples
International bodies	UNEP, UN Sustainable Development Goals discussions, Global Environmental Outlook
Regional governance	EU 8th Environment Action Programme (legally binding), European Environment Agency assessments
National governments	Switzerland, Sweden, Netherlands, Germany, New Zealand — all have published national boundary assessments
Corporate / finance	World Business Council for Sustainable Development (Vision 2050), Science Based Targets Network, growing ESG adoption
Science	Potsdam Institute annual Planetary Health Check, Earth Commission "safe and just" boundaries (2023)

The Ozone Lesson

Two boundaries are not breached. One of them — aerosol loading — may be more a function of measurement limitation than genuine safety. The other — stratospheric ozone — is the genuine article: a planetary-scale environmental problem identified, agreed upon, and solved through coordinated international action.

The Montreal Protocol (1987) is the single most successful international environmental treaty ever enacted. It phased out CFCs. It worked. The ozone layer is healing. This proves that when the science is clear and the political will exists, humanity can operate within planetary boundaries.

The difference between the ozone boundary and the other eight is not the difficulty of the problem — it's that we chose to solve it.

For a complete, cited breakdown of every metric mentioned in this post — and the full data behind the projections — see{" "} A Letter to Humanity.

Sources and Method

This article separates three kinds of evidence: peer-reviewed boundary definitions, annual planetary-boundary status updates, and live monitoring datasets. The boundary values mostly come from the 2023 Science Advances update and the 2025 Planetary Health Check; live observational examples, such as Mauna Loa CO₂, come from NOAA.

Updated 14 June 2026. Values should be refreshed when the next Planetary Health Check or WMO/NOAA annual datasets are released.

Primary sources: Rockström et al. (2009), Steffen et al. (2015), Richardson et al. (2023), the 2025 Planetary Health Check from the Potsdam Institute for Climate Impact Research, Stockholm Resilience Centre, NOAA GML, WMO, IPBES, WWF/ZSL Living Planet Index, and the Earth Commission's safe-and-just boundary work.

The Field Co

Open-Source Conservation Technology

PyTorch-Wildlife

Sun, 14 Jun 2026 00:00:00 GMT

import { Image } from "astro:assets"; import BlogImage from "../../components/blog/BlogImage.astro"; import cameraTrapFox from "../../assets/blog/pytorch-wildlife/29619938_roderick_salatan.jpg";

Camera traps changed wildlife monitoring by making remote ecosystems visible. They also created a new bottleneck: millions of images that someone still has to review.

A single field campaign can produce months of empty frames, false triggers, half-visible animals, night-time infrared images, people walking through reserves, vehicles on tracks, and species that look different across regions and seasons. The ecological value is enormous, but the labour required to process the images can delay science and conservation decisions for months or years.

PyTorch-Wildlife is an open-source deep learning framework from the Microsoft AI for Good Lab built for exactly this problem. It gives conservation practitioners and developers a shared way to run, adapt, and combine wildlife AI models for detection, classification, video, batch processing, and emerging modalities such as bioacoustics and overhead imagery.

The easiest way to understand the project is this:{" "} PyTorch-Wildlife is the model zoo and workflow layer around modern conservation AI . It hosts MegaDetector, species classifiers, DeepFaune, HerdNet-style aerial detection, bioacoustic models, demo notebooks, and practical tools for turning raw wildlife media into usable evidence.

Source note: This article was prepared from public PyTorch-Wildlife documentation, the Microsoft Biodiversity and PyTorch-Wildlife GitHub repositories, PyPI metadata, MegaDetector documentation, and the 2024 PyTorch-Wildlife paper accessed on 14 June 2026. Package versions, model lists, APIs, and release plans change over time, so production users should check the official repository and documentation before deployment.

Why Wildlife AI Needs a Shared Framework

Conservation AI is not just a modelling problem. It is a workflow problem. The model has to fit into how ecologists, rangers, protected-area managers, NGOs, students, citizen scientists, and government agencies actually collect and use data.

Camera traps, drones, autonomous recorders, and field sensors can now produce data faster than human teams can label it. But many conservation teams do not have dedicated machine learning engineers. Even when strong models exist, the real-world barriers are installation, GPU setup, image formats, annotation formats, model weights, data cleaning, confidence thresholds, review tooling, species lists, local ecological context, and reproducibility.

The PyTorch-Wildlife paper frames this around three practical requirements:

Accessibility: models should be easy to install, run locally or in the cloud, and usable by non-specialists.
Scalability: the framework should support new datasets, species, sensors, models, and workflows.
Transparency: the code and model pipeline should be open enough for researchers to understand, audit, adapt, and extend.

PyTorch-Wildlife responds by packaging common conservation AI tasks behind a consistent Python interface while also supporting notebooks, Hugging Face demos, batch processing, and graphical tools. It does not remove the need for ecological judgement. It reduces the amount of repetitive sorting and makes it easier for expert review to focus where it matters.

What PyTorch-Wildlife Is

PyTorch-Wildlife is a collaborative deep learning framework for conservation, maintained by Microsoft's AI for Good Lab. It is built on{" "} PyTorch, the widely used open-source machine learning framework known for Python-based deep learning, GPU acceleration, and a large ecosystem of research and production tools.

In practical terms, PyTorch-Wildlife provides three things:

Layer	What it provides	Why conservation teams care
Model access	One-line loading of detection, classification, bioacoustic, and overhead models, with weights downloaded automatically.	Teams can run established models without rebuilding the model stack from scratch.
Workflow utilities	Single-image inference, batch processing, video support, image transforms, demos, and notebooks.	Projects can move from a few test images to large field datasets using the same toolkit.
Extension point	A modular codebase for adding new models, datasets, interfaces, and conservation tasks.	Researchers and developers can adapt the framework to new species, regions, sensors, and field constraints.

The current documentation describes PyTorch-Wildlife as the unified open-source AI framework for wildlife monitoring, hosting detection models, species classifiers, and tools for single-image inference through large-scale batch processing.

Where It Fits in Microsoft's Biodiversity Ecosystem

PyTorch-Wildlife is part of a broader open-source biodiversity ecosystem from the Microsoft AI for Good Lab. The umbrella documentation describes a move away from a single large repository toward focused repositories for different projects, with PyTorch-Wildlife acting as the collaborative framework and model zoo.

Project	Role
PyTorch-Wildlife	The main Python framework and model zoo for conservation AI workflows.
MegaDetector	Open-source model family for detecting animals, people, and vehicles in camera-trap imagery.
MegaDetector-Classifier	Species-classification fine-tuning tools for adapting classifiers to local datasets and regions.
MegaDetector-Acoustic	Bioacoustic models for audio-based wildlife monitoring.
MegaDetector-Overhead	Point-based detection models for aerial and overhead imagery.
SPARROW	Solar-Powered Acoustic and Remote Recording Observation Watch, an edge-AI device for remote field deployments.
SPARROW Studio	A graphical interface built on top of the PyTorch-Wildlife ecosystem for data management, inference, analysis, and annotation.

This structure matters because conservation AI is no longer only about one camera-trap detector. The ecosystem is expanding across camera traps, species classifiers, acoustic recorders, aerial imagery, sonar-like modalities, edge devices, and non-coder interfaces.

The Basic Workflow

A typical PyTorch-Wildlife workflow starts with raw media and ends with structured predictions that can be reviewed, filtered, summarized, or passed into downstream ecological analysis.

Step	What happens	Output
1. Input	Load a camera-trap image, folder, video, audio file, or other supported media type.	File path, image tensor, video frames, or audio segments.
2. Detection	Run a detector such as MegaDetector to find animals, people, vehicles, or other relevant objects.	Bounding boxes, class labels, and confidence scores.
3. Filtering	Remove blanks, low-confidence predictions, or irrelevant detections depending on project needs.	A smaller set of candidate images or detections for review.
4. Classification	Run a downstream classifier on detections or cropped animals to assign likely species or categories.	Species/category predictions and confidence scores.
5. Human review	Experts validate uncertain records, correct errors, and decide which predictions are fit for analysis.	Cleaned, reviewed, and project-ready observation data.
6. Analysis	Use reviewed outputs for occupancy modelling, activity patterns, invasive-species detection, trend analysis, or reporting.	Ecological evidence, monitoring reports, and conservation decisions.

The key design idea is separation of tasks: detection tells you whether and where something is in the image; classification tells you what it probably is; human review decides what should be trusted for a specific scientific or management purpose.

Roderick Salatan on Pexels`} />

Quick Start

The package is published on PyPI as PytorchWildlife. As of 14 June 2026, PyPI lists version 1.3.0, released on 22 April 2026.

pip install PytorchWildlife

A minimal image-detection example looks like this:

```python from PytorchWildlife.models import detection as pw_detection model = pw_detection.MegaDetectorV6() results = model.single_image_detection("path/to/camera_trap_image.jpg") ```

A basic species-classification example uses the classification module:

```python from PytorchWildlife.models import classification as pw_classification classifier = pw_classification.AI4GAmazonRainforest() results = classifier.single_image_classification("path/to/image.jpg") ```

The official installation docs recommend Python 3.8 or later, with Python 3.10 or later preferred. GPU acceleration is optional, but an NVIDIA GPU with CUDA can provide a major speedup for large image batches. Users who do not want to install locally can try the Hugging Face demo or the Google Colab notebook linked from the documentation.

The Model Zoo

The model zoo is the heart of PyTorch-Wildlife. It turns the framework from a code library into a practical toolkit. The current documentation lists detection models, classification models, detection-plus-classification pipelines, and bioacoustic models.

Detection models

Model	Purpose	Notes
MegaDetector V6	Animal, person, and vehicle detection in camera-trap imagery.	Current generation, with YOLOv10, YOLOv9, and RT-DETR variants that trade off speed, accuracy, and licence type.
MegaDetector V5	Previous-generation animal/person/vehicle detector.	Widely deployed and still available for existing workflows.
DeepFaune Detector	Camera-trap detection model trained for European ecosystems.	Integrated as a third-party detection model.
HerdNet	Point-based localization for aerial and overhead imagery.	Useful where animals are counted or localized from overhead views rather than boxed in camera-trap frames.

Classification models

Model	Geographic/ecological focus	What it classifies
AI4G Amazon Rainforest	Amazon rainforest	Approximately 36 animal species in the documented model zoo.
AI4G Snapshot Serengeti	African savanna	Approximately 48 species in the documented model zoo.
AI4G Opossum	Americas / Galápagos use case	Opossum versus non-opossum classification.
DeepFaune Classifier	Europe	Approximately 44 species in the documented model zoo.
DFNE	Northeastern North America	A fine-tuned DeepFaune classifier for regional species recognition.

The important point is that model choice is ecological, not only technical. A classifier trained for the Amazon should not be assumed to work in South Africa, Europe, the Galápagos, or a fenced reserve without validation. Detection models often generalize better than species classifiers, but every deployment still needs local testing.

The MegaDetector Relationship

MegaDetector is the best-known model in the PyTorch-Wildlife ecosystem. It detects animals, people, and vehicles in camera-trap imagery, making it especially useful for removing blank images and reducing the manual review burden.

MegaDetector is not a species classifier. It can say, in effect, "there is an animal here" and draw a box around it. It does not reliably say "this is a leopard" or "this is a jackal" on its own. For that, a project pairs MegaDetector with a species classifier or a human review workflow.

Question	MegaDetector	Species classifier
Is there something in this image?	Yes, this is the core task.	Usually not the first step.
Where is the animal/person/vehicle?	Returns bounding boxes and confidence scores.	Usually works on a crop or detected region.
What species is it?	No — not its main job.	Yes, if trained and validated for that region/species set.
Can it remove blanks?	Yes, by filtering images with no confident detections.	No, unless paired with detection or a separate filtering step.

The current MegaDetector documentation recommends V6 for new projects and explains that V5 and earlier weights remain available for workflows that depend on them. It also notes that MegaDetector is used by more than 80 conservation programs and organizations worldwide, including government agencies, NGOs, universities, museums, zoos, and technology platforms.

Documented Applications

The 2024 PyTorch-Wildlife paper describes two real-world classification applications: species recognition in the Amazon Rainforest and invasive opossum recognition in the Galápagos Islands. In the arXiv abstract, the authors report 98% accuracy for the opossum model and{" "} 92% recognition accuracy for the Amazon model across{" "} 36 animals in 90% of the data.

These examples show the framework's intended use: not a universal species oracle, but a practical way to train, package, and share models for specific conservation needs.

Use case	Conservation problem	AI role
Amazon rainforest classification	Large, species-rich camera-trap datasets that are slow to annotate manually.	Classify animals from camera-trap imagery after detection and preprocessing.
Galápagos opossum recognition	Detecting or monitoring invasive species that threaten island ecosystems.	Binary or targeted classification: opossum versus non-opossum.
European camera-trap workflows	Regional monitoring where species lists and habitats differ from global benchmark datasets.	Use DeepFaune detector/classifier or regional fine-tuning such as DFNE.
Overhead wildlife localization	Aerial imagery where animals are visible from drones, aircraft, or overhead sensors.	Use point-based localization models such as HerdNet or the emerging overhead model work.
Bioacoustic monitoring	Species that are easier to detect by sound than by image, especially birds, frogs, insects, or nocturnal species.	Run audio classifiers through the bioacoustics module or related MegaDetector-Acoustic work.

Sparrow Studio and the Move Toward Friendlier Interfaces

The PyPI project page for version 1.3.0 highlights a major direction for the project: many conservation users prefer a graphical interface over writing code. In response, the team describes{" "} Sparrow Studio as a unified UI built on top of PyTorch-Wildlife.

The planned role of Sparrow Studio is broader than running a detector. It is intended to support local and cloud data management, model inference from the PyTorch-Wildlife model zoo, post-inference statistics, analysis, annotation, bounding-box editing, category editing, embedding visualization, and feature retrieval.

This matters because the last mile of conservation AI is often not the model. It is the interface used by the field team. A strong detector that requires a fragile command-line pipeline may be less useful than a slightly less flexible tool that project staff can run, inspect, correct, and trust.

PW-Engine and the Future Architecture

The version 1.3.0 project description also introduces{" "} PW-Engine, described as a model-agnostic inference core written in Rust. The stated goal is to make PyTorch-Wildlife evolve into a stable API and high-quality model zoo layered on top of a more general inference engine, while Sparrow Studio becomes the user-friendly frontend.

The described consumption surfaces are important for developers:

HTTP REST API: useful for web services, internal dashboards, and pipeline integration.
Single-binary CLI: useful for batch workflows, field laptops, and automated processing.
Python bindings: useful for notebooks, research code, and data-science pipelines.
Native C library: useful for desktop applications and lower-level integrations.

If this direction matures, PyTorch-Wildlife becomes less like a single Python package and more like an ecosystem: a model registry, a Python API, a desktop UI, an inference engine, and a shared set of conservation AI conventions.

Strengths

Strength	Why it matters
Open source	Researchers and developers can inspect code, reproduce workflows, adapt models, and contribute improvements.
Model zoo	Teams can start with established detectors and classifiers rather than training everything from scratch.
Python-first	Fits naturally into data-science workflows using notebooks, pandas, geospatial tools, and ecological analysis code.
Multiple entry points	Supports local installation, Hugging Face, Colab, notebooks, command-line-style workflows, and emerging graphical interfaces.
Detection plus classification	Reflects the way camera-trap analysis actually works: find objects first, then classify and review.
Extensible scope	The ecosystem is expanding beyond camera traps into bioacoustics, overhead imagery, edge deployments, and desktop review tools.

Limitations and Caveats

PyTorch-Wildlife is useful because it packages strong models and workflows, but it does not make wildlife monitoring automatic or error-free. The most important caveats are ecological and operational, not just technical.

Limitation	What can go wrong	Good practice
Domain shift	A model trained in one region may fail in another because backgrounds, camera angles, species, seasons, lighting, and vegetation differ.	Validate on local images before trusting predictions.
Classifier scope	A classifier can only predict categories it has been trained to recognize. Unknown species may be mislabelled as the closest known class.	Use project-specific species lists, confidence thresholds, and an "unknown" review path.
False negatives	Rare, small, obscured, nocturnal, or partially visible animals may be missed.	Audit blank-removal workflows and sample supposedly empty images.
False positives	Vegetation, shadows, heat signatures, or camera artefacts can be mistaken for animals.	Set thresholds according to the cost of missing animals versus reviewing extra images.
Bias in training data	Models often perform best on species, camera types, environments, and geographies well represented in training data.	Report validation results by species, site, season, and camera setup where possible.
Sensitive species data	Automated pipelines can accelerate the exposure of locations for threatened or poached species.	Protect coordinates, access controls, and publication workflows for sensitive records.
Operational maintenance	Python, CUDA, model versions, and package dependencies can break pipelines if not pinned and documented.	Use versioned environments, record model versions, and store inference parameters with outputs.

The safest framing is: PyTorch-Wildlife can speed up review and improve consistency, but the output remains a prediction. Conservation decisions still need metadata, sampling design, uncertainty handling, and expert validation.

Responsible Use in Conservation Workflows

AI-assisted wildlife monitoring should be designed around the conservation question, not around the model. A project estimating occupancy, monitoring poaching access, detecting invasive species, counting animals from aerial imagery, or triaging camera-trap data will need different thresholds and review processes.

A responsible PyTorch-Wildlife deployment should document:

Model identity: exact model name, architecture variant, version, weights, and licence.
Data scope: camera locations, dates, species list, habitat types, camera models, and sampling design.
Inference settings: confidence thresholds, image resizing, batch settings, GPU/CPU environment, and preprocessing.
Review protocol: who checked predictions, what was sampled, how disagreements were resolved, and what confidence classes were accepted.
Error analysis: false positives, false negatives, species-level confusion, and performance on rare or priority species.
Sensitive data policy: whether locations or timestamps need to be hidden, generalized, or access-controlled.

This turns the framework from a convenient black box into a repeatable scientific workflow.

How It Differs from Other Wildlife Data Tools

PyTorch-Wildlife sits closer to the model and inference layer than to long-term ecological data management. It can work alongside platforms such as Wildlife Insights, Wildbook, Timelapse, Camelot, Agouti, Zooniverse, GBIF, and custom research databases, but it is not a one-for-one replacement for all of them.

Tool type	Main job	Relationship to PyTorch-Wildlife
Camera-trap AI framework	Run and combine detection/classification models.	This is PyTorch-Wildlife's core role.
Camera-trap management platform	Organize projects, images, deployments, labels, and review workflows.	Can consume or complement PyTorch-Wildlife predictions.
Citizen-science annotation platform	Use distributed human review to label large datasets.	AI can pre-filter images or prioritize human review.
Individual animal re-identification platform	Match individuals using stripes, spots, scars, fins, flukes, or other markings.	Different task: detection/classification can feed candidate media into re-ID workflows.
Biodiversity occurrence infrastructure	Publish and discover species occurrence records.	Reviewed outputs may later become occurrence records or monitoring evidence.

Technical Notes for Developers

For developers, PyTorch-Wildlife is useful because it gives a common abstraction around models that would otherwise require separate setup, preprocessing, output formats, and weight management.

Install path: the Python package is{" "} PytorchWildlife, with imports under{" "} PytorchWildlife.models.
Model loading: many model weights download automatically on first use.
V6 detector options: MegaDetector V6 variants include YOLOv10, YOLOv9, and RT-DETR choices, with AGPL, MIT, and Apache-licensed variants documented in the model zoo.
GPU setup: the docs provide CUDA installation guidance and note that GPU acceleration is optional but important for large jobs.
Interfaces: local Python, notebooks, Hugging Face, Colab, Docker examples, and the emerging Sparrow Studio/PW-Engine layers support different users.
Citation: the project asks users to cite Hernandez et al. 2024 for PyTorch-Wildlife and Beery, Morris, and Yang 2019 when MegaDetector is used specifically.

For production pipelines, pin package versions, record model architecture choices, save raw prediction JSON/CSV outputs, and keep a small validation dataset for regression testing when upgrading models.

Where the Project Is Going

The direction of PyTorch-Wildlife is broader than camera-trap classification. The public roadmap language around version 1.3.0 points toward four important trends:

More modalities: bioacoustics, overhead imagery, sonar-like imagery, and edge-device data alongside camera traps.
Better interfaces: graphical tools such as Sparrow Studio for users who do not want to write Python.
More flexible deployment: PW-Engine, REST APIs, CLIs, Python bindings, and desktop integration.
Local model adaptation: future tools for non-coders to fine-tune models on their own data.

That trajectory reflects the reality of biodiversity monitoring. The world does not need only a better benchmark score. It needs robust, reusable, auditable tools that can survive field conditions, work with imperfect data, and help conservation teams make decisions faster.

Why PyTorch-Wildlife Matters

PyTorch-Wildlife matters because it turns conservation AI from a collection of scattered models into a more coherent toolkit. It gives the field a shared way to load MegaDetector, run species classifiers, process images and videos, experiment with new model types, and move toward graphical and production-ready workflows.

Its biggest value is not that it removes humans from biodiversity monitoring. Its value is that it can move human expertise higher up the chain: away from endless blank-image sorting and toward validation, ecological interpretation, field response, and conservation action.

Used carefully, PyTorch-Wildlife is part of a larger shift in conservation technology: from isolated research scripts toward open, collaborative, reusable infrastructure for monitoring life on Earth.

Sources and Further Reading

SA-FARI — 11,609 Videos, 99 Species Categories, and a New Open Benchmark for Wildlife AI

Sun, 14 Jun 2026 00:00:00 GMT

import { Image } from "astro:assets"; import BlogImage from "../../components/blog/BlogImage.astro"; import SaFari2 from "../../assets/blog/sa-fari/34318597_amar__preciado.jpg"; import SaFari1 from "../../assets/blog/sa-fari/31001480_thilina_alagiyawanna.jpg"; import cxlLogo from "../../assets/blog/conservationxlabs-logo.svg"; import saFariPhoto from "../../assets/blog/sa-fari/33310474_ali_kazal.jpg"; import wildlifeFootage from "../../assets/blog/sa-fari/20339628_ryan_beirne.jpg";

For years, the bottleneck in AI for conservation has not been the algorithms. It has been the data. Training a computer vision model to detect, classify, segment, or track wildlife requires large numbers of verified examples — footage where animals are labelled, outlined, and followed across frames. Camera-trap projects generate this raw material every day, but annotation is slow, expensive, and difficult to standardise across species, continents, seasons, and field conditions.

SA-FARI is a major attempt to close that gap. The dataset, built by Conservation X Labs and Meta with a coalition of research and conservation partners, contains 11,609 camera-trap videos, 99 species categories, 16,224 masklet identities, and 942,702 individual bounding boxes, segmentation masks, and species labels. The videos span approximately ten years of field observations, from 2014 to 2024, across 741 independent sampling locations on four continents.

The full title is{" "} Segment Anything in Footage of Animals for Recognition and Identification . The name is important: SA-FARI is not only a wildlife image collection. It is a video benchmark for multi-animal tracking — the task of finding every relevant animal, preserving its identity through time, and handling the messy conditions that real camera traps produce.

What Is SA-FARI?

SA-FARI is an open-source multi-animal tracking dataset for wild animals. Each video-species pair is annotated with a spatio-temporal segmentation of all animals belonging to that category. Instead of only drawing one box around an animal in a single frame, the dataset provides masks and tracklets that preserve animal identity across video frames.

The dataset was built using Meta's Segment Anything Model 3 (SAM 3), a model designed to detect, segment, and track objects in images and videos from text or visual prompts. In the SA-FARI paper, SAM 3 is evaluated both with species-specific prompts and with generic animal prompts, then compared with wildlife-focused detector-plus-tracker baselines such as MegaDetector combined with ByteTrack, OCSort, and BoostSort++.

Metric	Training	Test	Total
Videos	10,776	833	11,609
Duration	2,545 min	202 min	2,747 min
Species categories	91	83	99
Masklets	15,141	1,083	16,224
Annotations	880,361	62,341	942,702
Video-species pairs	31,282	2,322	33,604
Independent sampling locations	650	91	741

The authors use the term species categories carefully. Labels are based on common names confirmed by local experts and mapped into taxonomy where possible. In aggregate, the dataset spans 4 classes, 23 orders, 53 families, and 82 genera. Like real field data, it is long-tailed: a few animals appear often, while many are rare. The paper reports that 29 species categories account for 90% of the data, with spider monkeys, collared peccaries, and agoutis among the most common.

Ali Kazal on Pexels`} />

Open, But Not Public Domain

SA-FARI is publicly listed on Hugging Face, but it should be described accurately. The dataset card lists the license as CC-BY-NC 4.0, which allows sharing and adaptation with attribution for non-commercial purposes. The Hugging Face page also requires users to log in and agree to share contact information before accessing the dataset content.

That still makes SA-FARI unusually open for conservation AI, where many camera-trap datasets remain private because of sensitive species locations, conservation risk, or partner restrictions. The published version includes anonymized camera-trap location identifiers rather than precise public coordinates, balancing reproducibility with field safety.

The Hugging Face repository contains the annotation files. The original videos and preprocessed 6 fps JPEG frames are hosted separately in a public Google Cloud Storage bucket referenced from the dataset card. The annotations follow a format similar to YouTube-VIS, with fields for videos, categories, annotations, and video-category pairs.

Who Built It

Conservation X Labs is a Washington, DC-based nonprofit founded in 2015 by Dr. Alex Dehgan and Dr. Paul Bunje. Its model combines conservation fieldwork, prizes, technology development, and data platforms. In Meta's November 2025 profile of the SA-FARI collaboration, CXL is described as having hosted 20 innovation challenges, provided more than $12 million in funding to breakthrough solutions, re-identified more than 299,000 animals, and supported the expansion of nearly 400,000 acres of protected areas.

SA-FARI was built by Conservation X Labs and Meta with footage and expertise from five additional partner organizations: Osa Conservation in Costa Rica, Los Amigos Biological Station in Peru, the Pan African Programme, the Institute for Game and Wildlife Research in Spain, and the Instituto Mixto de Investigación en Biodiversidad. The broader author list also includes researchers from the University of Bristol, Hasso Plattner Institute, University of Oviedo, Senckenberg Museum of Natural History, Max Planck Institute for Evolutionary Anthropology, Climate Corridors, CXL, and Meta.

This matters because a general wildlife model cannot be trained from one reserve, one country, or one charismatic species. SA-FARI's value comes from ecological variety: tropical rainforest, savanna, temperate woodland, day and night footage, single animals and groups, small masks, occlusion, and the long-tail distribution that conservationists actually see in the field.

Thilina Alagiyawanna on Pexels`} />

Why This Matters

Three things make SA-FARI significant beyond its headline size.

First, it is video. Camera-trap AI has historically been dominated by still-image detection and classification. Video captures movement, interactions, gait, occlusion, and behaviour. Multi-animal tracking is the bridge between knowing that an animal appeared and understanding what that animal did over time.

Second, it provides segmentation masks and tracklets. A bounding box can tell a model roughly where an animal is. A mask tells the model which pixels belong to the animal. A tracklet preserves that animal's identity across frames. Together, these annotations support stronger benchmarks for detection, segmentation, tracking, behaviour analysis, and eventually individual re-identification.

Third, it reflects real-world difficulty. Camera-trap data is messy. Animals enter partly cropped, overlap each other, appear at night, move quickly, or occupy a tiny fraction of the frame. The SA-FARI paper reports that small masklets are substantially harder to detect and track than large ones, and that challenging motion or occlusion makes tracking harder even when detection remains possible.

For context on how AI fits into the broader camera-trap workflow — from deployment to analysis — see our{" "} camera trap software comparison. SA-FARI is the kind of benchmark that can improve tools such as MegaDetector, Wildlife Insights, Wildbook, and other conservation AI systems that rely on robust detection and tracking.

Ryan Beirne on Pexels`} />

What the Benchmarks Show

The SA-FARI paper tests both species-specific and species-agnostic tracking. In the species-specific setting, SAM 3 trained or fine-tuned with SA-FARI substantially outperforms the baseline SAM 3 model. In the species-agnostic setting, SAM 3 trained on SA-FARI also outperforms MegaDetector paired with common tracking algorithms such as ByteTrack, OCSort, and BoostSort++.

The important takeaway is not that one model permanently wins. It is that wildlife video is a distinct domain. General-purpose computer vision models can be powerful, but they still need field-relevant, geographically diverse, species-rich training and evaluation data. SA-FARI gives researchers a common basis for measuring progress rather than comparing results across incompatible private datasets.

Amar Preciado on Pexels`} />

What This Enables

A dataset like SA-FARI is not an end product. It is infrastructure. Here is what it makes possible:

Better wildlife video models. Researchers can train and evaluate detection, segmentation, and tracking systems on a shared benchmark instead of relying only on private field datasets.

Behavioural analysis. Many conservation questions are temporal: feeding, social interaction, avoidance, predator-prey dynamics, disease signals, and animal responses to human disturbance. Video tracklets are a foundation for building those downstream models.

Open-world recognition. Several species categories are deliberately present only in the test split. That forces models to face a real deployment problem: a camera trap will eventually record animals the model has not seen during training.

Lower barriers for conservation organizations. Smaller NGOs and field teams may not have the budget to annotate thousands of videos from scratch. SA-FARI provides a starting point for building or fine-tuning systems that are better aligned with field conditions.

Caveats for Responsible Use

SA-FARI is a benchmark, not a finished conservation product. Models trained on it still need local validation before they are used for management decisions. Species distributions, camera hardware, habitat, season, weather, and field protocols can all shift model performance.

The non-commercial license also matters. SA-FARI can accelerate research and conservation prototyping, but commercial users need to review the CC-BY-NC 4.0 terms carefully. Sensitive species data should also remain protected: anonymized locations reduce risk, but downstream users should avoid publishing precise locations for threatened or trafficked species.

Finally, AI should reduce annotation burden, not remove ecological expertise. The best use of SA-FARI is as a shared foundation: models do the repetitive work faster, while local experts remain responsible for taxonomy, survey design, uncertainty, and conservation interpretation.

Get the Dataset

SA-FARI is listed on Hugging Face at{" "} huggingface.co/datasets/facebook/SA-FARI. The dataset card describes the license, access conditions, file layout, annotation format, and links to the public cloud storage location for videos and 6 fps JPEG frames.

The paper is available at{" "} arxiv.org/abs/2511.15622 and as a CVPR 2026 open-access paper. It documents the data collection, annotation pipeline, train/test split, taxonomy, benchmark settings, and evaluation results.

Conservation X Labs continues to build open infrastructure for biodiversity monitoring. SA-FARI sits naturally alongside its work on Wild Me and Wildbook for animal re-identification, Sentinel for protected area monitoring, and other conservation technology programs. If SA-FARI is the dataset layer, those tools are part of the application layer.

Dataset: Conservation X Labs & Meta (2025). SA-FARI: Segment Anything in Footage of Animals for Recognition and Identification.{" "} Hugging Face. License listed on the dataset card: CC-BY-NC 4.0. Paper:{" "} arXiv:2511.15622 and CVPR 2026 open-access proceedings. CXL project page:{" "} conservationxlabs.com/sa-fari . Meta profile:{" "} How Conservation X Labs Is Using Segment Anything Model 3 for Endangered Wildlife Monitoring . SAM 3: Meta AI SAM 3.

The Field Co

Open-Source Conservation Technology

Wildlife Volunteering in South Africa: Comparison Guide for International Students

Sun, 14 Jun 2026 00:00:00 GMT

import { Image } from "astro:assets"; import BlogImage from "../../components/blog/BlogImage.astro"; import penguinColony from "../../assets/blog/south-africa-wildlife-volunteer-comparison-guide/7177262_jeffrey_eisen.jpg"; import elephantRhino from "../../assets/blog/south-africa-wildlife-volunteer-comparison-guide/6818980_magda_ehlers.jpg"; South Africa has a huge volunteer market: true conservation monitoring, reserve maintenance, penguin rehabilitation, shark and marine research, wildlife sanctuaries, primate rescue centres, big-cat facilities, community education projects, and large international placement agencies. That variety is exciting, but it also makes the choice risky. A student can easily pay for a programme that looks like conservation on Instagram but is mostly tourism, animal handling, or reserve labour. This guide compares the most visible and researchable wildlife volunteering options in South Africa as of **14 June 2026**. It is written for a student coming from overseas who wants useful experience, a safe and ethical placement, and a realistic understanding of what each programme offers. > **Scope note:** “All places” is difficult because many reserves, NGOs, and sanctuaries take volunteers informally or only through agents. This guide includes the major public-facing programmes and organisations I could validate through official websites, independent review platforms, and Instagram/social presence. It favours programmes with clear public information, field relevance, review history, and traceable conservation or animal-welfare purpose. --- ## Fast recommendation For most students, the strongest options are: | Student goal | Best-fit programmes | | ---------------------------------------------------- | ------------------------------------------------------------------------------------------------------------- | | Serious endangered-species field monitoring | **Wildlife ACT**, **GVI Karongwe**, **Siyafunda**, **African Impact Big 5** | | Reserve life plus conservation and community work | **Kariega Foundation**, **Shamwari Conservation Experience** | | Marine biology / shark / seabird experience | **Oceans Research**, **Marine Dynamics Academy**, **SANCCOB**, **APSS / Dyer Island Conservation Trust** | | Penguin and seabird rehabilitation | **SANCCOB**, **APSS** | | Wildlife rehabilitation / husbandry | **Moholoholo**, **Tenikwa**, **DAKTARI**, **Vervet Monkey Foundation**, **C.A.R.E. Baboon Sanctuary** | | Community education plus animals | **DAKTARI**, **Kariega Foundation** | | Big-cat sanctuary experience | **Panthera Africa** — but only if the student understands this is sanctuary care, not wild field conservation | | Best “CV-building” options for conservation students | **Wildlife ACT**, **GVI**, **Oceans Research**, **SANCCOB**, **Kariega**, **Siyafunda** | If the student wants the closest thing to practical ranger-style conservation work, start with **Wildlife ACT**, **GVI Karongwe**, **Siyafunda**, **African Impact Big 5**, **Kariega**, or **Shamwari**. If the student wants animal-care experience, look at **SANCCOB**, **APSS**, **Moholoholo**, **VMF**, **C.A.R.E.**, or **DAKTARI**. If the student wants “touching lions, walking cheetahs, cub photos, elephant rides, or predator selfies,” do **not** book that experience. --- ## Ethical filter: how to avoid bad voluntourism Before comparing programmes, use this filter. A programme should be able to answer these questions clearly. ### Green flags - The work is linked to **monitoring, research, rehabilitation, education, reserve management, habitat work, or community conservation**. - Volunteers are supervised by qualified staff. - The organisation explains how volunteer fees support conservation, animal care, staff, equipment, food, vehicles, or operations. - Wild animals remain wild; captive animals are there because of rescue, injury, rehabilitation, or lifelong sanctuary need. - The programme is honest that students may do basic work: cleaning, data entry, early mornings, maintenance, food prep, camera-trap sorting, alien plant removal, or observation. ### Red flags - Lion cub petting, bottle-feeding large carnivores, walking with big cats, predator selfies, elephant riding, or performing animals. - Claims that captive lion cubs will be released into the wild. - Breeding predators for tourism, trade, or vague “conservation”. - No named reserve, no project partner, no explanation of data collection, or no local staff leadership. - The Instagram feed is mostly glamorous selfies and animal contact, with little fieldwork, education, research, or local impact. The SATSA captive wildlife decision tree flags tactile interactions with infant wild animals, predators, cetaceans, walking with predators or elephants, riding wild animals, performing animals, breeding, trade, and unclear sanctuary status as major issues. South African Tourism has also publicly stated that it does not promote or endorse wild-animal interaction experiences such as petting wild cats, interacting with elephants, or walking with lions and cheetahs. --- ## Comparison table | Programme / organisation | Main type | Location | Typical fit | Minimum stay / age | Review signal | Instagram / social signal | Ethical note | | -------------------------------------------------------------------------------------- | ---------------------------------------------- | ------------------------------------------ | ---------------------------------------------------- | ---------------------------------------------------------- | ------------------------------------------------------------------------ | ---------------------------------------------------------- | ------------------------------------------------------------------------------------------------------- | | **Wildlife ACT** | Endangered species monitoring | Zululand, KwaZulu-Natal; multiple reserves | Best overall for real field conservation | 18–70+, 2 weeks+ | Go Overseas: 4.88 / 40 reviews; strong alumni interviews | `@wildlife_act` | Strong: small groups, WWF-linked, no animal handling focus | | **GVI Wildlife Research Expedition** | Structured research volunteering | Karongwe / Limpopo, near Kruger | Student wanting structure, training, support | 1–12 weeks; staff ratio listed | Go Overseas: 4.86 / 50 reviews | `@gvitravel` | Good for structured gap-year style; check project partner and data outputs | | **African Impact Big 5 Wildlife Conservation** | Reserve monitoring and conservation | Greater Kruger | Student wanting Big 5 field exposure | 2 weeks+; seasonal availability shown | GoAbroad reviews positive; GoAbroad review highlights research and staff | `@africanimpact` | Solid field option; also an agency/operator, so confirm reserve and task schedule | | **Siyafunda Wildlife & Conservation** | Wildlife monitoring / reserve support | Limpopo / Makalali area | More rugged bush monitoring | Programme details less polished online | TripAdvisor reviews mention professional, safety-focused staff | `@siyafundawildlife` | Promising direct field focus; verify dates, price, insurance, and accommodation | | **Shamwari Conservation Experience** | Reserve conservation + maintenance + education | Eastern Cape | Student wanting reputable reserve and Big 5 exposure | Min. 2 weeks often marketed | WorkingAbroad reviews: 4.9 / 17 reviews | `@shamwariprivategamereserve`; `@teamshamwari` | Reputable reserve; some tasks are maintenance and education, not only wildlife monitoring | | **Kariega Foundation Volunteer Programme** | Conservation + community | Eastern Cape | Best for mixed reserve/community experience | 2 weeks common; weekly schedule public | WorkingAbroad reviews very positive; Kariega posts testimonials | `@kariegavolunteers`, `@kariega_foundation` | Strong if student wants people + wildlife; not pure field research | | **SANCCOB** | Seabird rescue and rehabilitation | Cape Town or Gqeberha | Best penguin/seabird rehab | 18+, minimum 6 weeks | External testimonials positive | `@sanccob` | Strong conservation-care option; hard work, cleaning, feeding, rehab routines | | **African Penguin & Seabird Sanctuary / Dyer Island Conservation Trust** | Seabird rehabilitation / marine conservation | Gansbaai / Kleinbaai | Penguin/seabird and Dyer Island ecosystem | Check directly | TripAdvisor APSS reviews positive | `@apssza`, `@dyerict` | Good specialist seabird option; clarify whether applying via APSS, DICT, or Marine Dynamics | | **Marine Dynamics Academy** | Marine volunteer / internship | Gansbaai | Marine Big 5, shark tourism/research support | Check course; volunteer/internship options | Go Overseas Marine Dynamics reviews positive but some old | `@marinedynamics`, `@dyerict` | Good for marine exposure; tourism operations are part of the model | | **Oceans Research** | Marine field research | Mossel Bay | Strongest shark/marine research CV option | VolunteerWorld lists 4–12 weeks, 18+ | VolunteerWorld: 4.9 / 20 reviews | `@oceansresearch` | Strong research positioning; good for marine biology students | | **DAKTARI Bush School & Wildlife Orphanage** | Environmental education + animal care | Hoedspruit / Limpopo | Community education + wildlife orphanage | Check directly | Go Overseas and VolunteerWorld reviews positive | `@daktaribushschool` | Strong people-and-wildlife focus; less reserve research | | **Moholoholo Wildlife Rehabilitation Centre** | Wildlife rescue, rehab, education | Hoedspruit | Hands-on rehab / animal-care learning | Minimum 2 weeks stated | TripAdvisor volunteer reviews positive | Moholoholo Instagram/social present; verify current handle | Rehab centre; good for animal care, but clarify contact rules and release pathways | | **Vervet Monkey Foundation** | Primate rehabilitation / sanctuary | Tzaneen, Limpopo | Primate care and sanctuary operations | 4+ weeks for longer stays; short stays possible | Go Overseas / VolunteerWorld positive | `@vervetmonkeyfoundation` | Good direct charity option; vegan meals; labour-intensive sanctuary work | | **C.A.R.E. Baboon Sanctuary** | Baboon rescue, rehab, release | Phalaborwa / Limpopo | Primate rehabilitation | Check directly | VolunteerWorld reviews mixed-positive | `@primate_care` | Valuable if student specifically wants primates; physically and emotionally demanding | | **International Primate Rescue** | Primate sanctuary | Pretoria / Gauteng | Primate care near city | Check directly | VolunteerWorld and TripAdvisor reviews | `@international_primate_rescue` | More sanctuary than field conservation; ask about welfare policies and daily duties | | **Free To Be Wild Sanctuary** | Wildlife rescue and sanctuary | KwaZulu-Natal area | Broad animal care / sanctuary work | Check directly | Own volunteer stories and social proof | `@freetobewildsanctuary` | Needs direct due diligence; good candidate if rehabilitation and release evidence is clear | | **Tenikwa Wildlife Rehabilitation & Awareness Centre** | Wildlife rehab / awareness / tours | Plettenberg Bay, Garden Route | Shorter stay, rehab/tourism exposure | Accommodation/tours public; volunteer details via partners | TripAdvisor review mentions volunteering positively | `@tenikwa_wildlife` | Captive-wildlife/tourism setting; verify no harmful interactions and actual rehab role | | **Panthera Africa Big Cat Sanctuary** | Big-cat sanctuary care | Stanford, Western Cape | Big-cat sanctuary, not wild conservation | Check direct availability | VolunteerWorld and TripAdvisor volunteer reviews | `@pantheraafricasanctuary` | Better than cub-petting if true sanctuary rules apply; still captive big cats, not field conservation | | **Cheetah Outreach** | Education and ambassador cheetah programme | Somerset West / Cape Town area | Public education / facility work | Low/no-cost options listed | Reviews mostly visitor-focused | `@cheetahoutreach.za` | Includes public animal encounters; use high caution if student prioritises strict no-interaction ethics | | **Hoedspruit Endangered Species Centre (HESC)** | Cheetah/rhino conservation, rehab, education | Hoedspruit | Endangered species facility exposure | Check current volunteer route | Reviews mostly visitor/tour focused | `@hesc_endangeredspeciescentre` | Captive cheetah/rhino facility; ask detailed welfare and breeding questions | | **Cheetah Experience** | Captive cheetah/big-cat volunteer/internship | Free State | Big-cat facility experience | Volunteer/internship options public | Reviews mixed across web; mostly facility-focused | Check current handle | High due diligence required because captive big-cat volunteering is an ethical minefield | | **African Conservation Experience (ACE)** | Placement provider, not one project | South Africa + wider southern Africa | Students needing guided placement advice | 1–12 weeks for many projects | Go Overseas / VolunteerForever positive | `@africanconservationexperience` | Strong broker if they place with reputable partners; still verify exact project | | **IVHQ / GoEco / VolunteerWorld / WorkingAbroad / Pod Volunteer / The Great Projects** | Booking platforms / agencies | Multiple | Easier logistics, packaged trips | Varies by listed project | Many platform reviews | Platform-specific accounts | Useful for comparison and support, but always identify the actual host project | --- ## Detailed profiles ### 1. Wildlife ACT **Best for:** a student who wants the most field-relevant endangered-species monitoring experience. Wildlife ACT is one of the strongest fits for a serious conservation student. It is not a general voluntourism agency; it is focused on endangered and priority species monitoring across multiple protected areas in Zululand. The official programme states that volunteers join small teams of up to six, work in private, community-owned and government-managed reserves, and help monitor species such as African wild dog, cheetah, black rhino, vultures, elephants, white rhino, hyena, and leopard. **Why it stands out** - Strong field-monitoring model. - Small volunteer groups. - Clear species focus. - Multiple reserves, including Hluhluwe-iMfolozi and Manyoni. - Good review signal: Go Overseas lists 4.88 from 40 reviews. - Strong ethical positioning against animal-contact voluntourism. **Student fit** Choose Wildlife ACT if the student wants early starts, telemetry, GPS, behaviour notes, species ID, data collection, and a more ranger-like experience. It is less suitable for someone who mainly wants animal cuddling, predictable safari-style viewing, or city comfort. **Review signal** Go Overseas reviews are strong and include alumni interviews. Reviewers repeatedly frame the work as real conservation, learning from staff, and being in the bush rather than simply touring. **Instagram** - `@wildlife_act` - Useful to check: monitoring posts, wild dog/rhino/vulture work, field monitors, and whether content is fieldwork-heavy rather than selfie-heavy. **Sources** - Official: https://www.wildlifeact.com/ - Volunteer programme: https://www.wildlifeact.com/volunteer/program/endangered-species-conservation-south-africa - Why volunteer: https://www.wildlifeact.com/why-volunteer - Reviews: https://www.gooverseas.com/volunteer-abroad/south-africa/wildlife-act/21946 - Instagram: https://www.instagram.com/wildlife_act/ --- ### 2. GVI Wildlife Research Expedition, South Africa **Best for:** a student who wants structured support, a known international provider, and a research-oriented bushveld placement. GVI’s South Africa wildlife research programme is based around wildlife and conservation research in the bushveld near Kruger, with official materials listing 1–12 week durations, 35 fieldwork hours per week, and a 1:6 staff-to-participant ratio. **Why it stands out** - Good support structure for first-time international students. - Clear schedule and training model. - Strong review signal on Go Overseas. - Good option for students who want a managed experience rather than arranging everything directly. **Student fit** Choose GVI if the student is nervous about travelling alone or wants a polished programme with clear arrival support. It may feel more “international volunteer programme” than “local conservation NGO,” so the student should ask exactly what data is collected, who uses it, and how it supports reserve management. **Review signal** Go Overseas lists GVI South Africa wildlife volunteering at 4.86 from 50 reviews, with a review summary praising knowledgeable staff and immersive wildlife/community experiences. **Instagram** - `@gvitravel` - Useful to check: Karongwe/Limpopo posts, staff-to-student activity, and whether alumni posts show actual data collection rather than only safari shots. **Sources** - Official programme: https://www.gvi.ie/programs/wildlife-research-south-africa-expedition/ - GVI South Africa overview: https://www.gvi.ie/volunteer-in-south-africa/ - Reviews: https://www.gooverseas.com/volunteer-abroad/south-africa/global-vision-international/52250 - Instagram: https://www.instagram.com/gvitravel/ --- ### 3. African Impact Big 5 Wildlife Conservation **Best for:** a student who wants Greater Kruger, Big 5 exposure, structured learning, and conservation tasks. African Impact’s Big 5 project is based in the Greater Kruger region. The official programme describes wildlife monitoring, practical conservation work, structured learning, tracking, behaviour interpretation, erosion control, bush clearing, reserve maintenance, and training in species identification and wildlife management. The page listed availability May–October, 2 weeks+, and a price from £1,960 at the time checked. **Why it stands out** - Strong Greater Kruger appeal. - Balanced field monitoring + habitat work. - Good for students who want social volunteer life and structured activities. - Public pricing and project description are clear. **Student fit** Good for a student who wants Big 5 exposure and a guided, social placement. The student should understand that not every day will be dramatic wildlife work; reserve maintenance and practical labour are part of conservation. **Review signal** GoAbroad includes positive reviews, including a 2024 reviewer who said the guides were knowledgeable and that contributing to research and conservation work felt meaningful. **Instagram** - `@africanimpact` - Useful to check: fieldwork reels, Greater Kruger base, photography internship posts, and tagged alumni. **Sources** - Official: https://africanimpact.com/volunteer-projects/wildlife-conservation/african-big-5-wildlife-conservation/ - GoAbroad review page: https://www.goabroad.com/providers/african-impact/programs/african-big-5-wildlife-conservation-project-south-africa-108826 - VolunteerWorld listing/reviews: https://www.volunteerworld.com/en/volunteer-program/field-research-and-wildlife-internship-in-south-africa-hoedspruit - Instagram: https://www.instagram.com/africanimpact/ --- ### 4. Siyafunda Wildlife & Conservation **Best for:** a student who wants a more rugged, reserve-based wildlife monitoring project. Siyafunda’s official site describes wildlife monitoring and sustainable support for game reserves, research, conservation projects, researchers, interns, educational affiliations, and local communities. Review snippets from TripAdvisor praise the team as friendly, knowledgeable, professional, and safety-focused. **Why it stands out** - More direct “bush monitoring” feel. - Stronger fit for students comfortable with less polished online infrastructure. - Good for those who want to live in a reserve environment and learn from guides/monitors. **Student fit** Good for self-motivated students who want field exposure and are comfortable asking detailed logistical questions before booking. **Questions to ask** - Which reserve will I be based on? - What species are being monitored this season? - What exact data do volunteers collect? - What is the vehicle/walking split? - Who receives the data? - What are emergency procedures? **Instagram** - `@siyafundawildlife` **Sources** - Official: https://www.siyafundaconservation.com/ - TripAdvisor: https://www.tripadvisor.co.za/Attraction_Review-g312616-d5062609-Reviews-Siyafunda_Wildlife_Conservation-Hoedspruit_Kruger_National_Park.html - VolunteerWorld listing: https://www.volunteerworld.com/en/volunteer-program/assistant-at-siyafunda-game-reserve-in-south-africa-hoedspruit - Instagram: https://www.instagram.com/siyafundawildlife/ --- ### 5. Shamwari Conservation Experience **Best for:** a student who wants a reputable Eastern Cape reserve, Big 5 exposure, structured conservation tasks, and some community/education work. Shamwari is a well-known private game reserve in the Eastern Cape. Its Conservation Experience markets behind-the-scenes work in a free-roaming Big Five reserve. Third-party review pages describe a varied programme: maintenance, alien plant clearing, monitoring elephant herds, identifying lions and rhinos, and participating in night patrols. **Why it stands out** - Established, high-profile reserve. - Good accommodation and logistics by review reputation. - Strong for first-time South Africa visitors. - Good mix of fieldwork, maintenance, and education. **Student fit** Choose Shamwari if the student wants a safer, more structured, polished reserve experience. It may be less specialised than Wildlife ACT for endangered-species monitoring but broader and easier for a first-time student. **Review signal** WorkingAbroad lists 4.9/5 from 17 reviews. Reviews praise accommodation, food, hands-on student experience, lectures, and activity variety. **Instagram** - `@shamwariprivategamereserve` - `@teamshamwari` - Also search the Shamwari Conservation Experience location tag. **Sources** - Official PDF: https://www.shamwari.com/wp-content/uploads/pdf/nAE8Hr307Ti6oAK7BlIGjJZRf81PLA1RqmCGNjqrE1M8Ijw9.pdf - Reviews: https://www.workingabroad.com/projects/shamwari-game-reserve-conservation-volunteer-programme-south-africa/reviews/ - Instagram: https://www.instagram.com/shamwariprivategamereserve/ - Location tag: https://www.instagram.com/explore/locations/277719255/shamwari-conservation-experience/ --- ### 6. Kariega Foundation Volunteer Programme **Best for:** a student who wants both wildlife conservation and community engagement. Kariega Foundation’s public programme is unusually clear about the mix: each week includes **3 days conservation work and 2 days community engagement**. During two weeks, volunteers do 6 days on Kariega Game Reserve and 4 days of community work. Conservation projects include predator/prey monitoring, anti-poaching and rhino monitoring, endangered and priority species tracking, bird research, and elephant research. **Why it stands out** - Strong balance of conservation and community. - Good for students who understand conservation includes people. - Clear schedule. - Strong Eastern Cape location. **Student fit** Best for students interested in sustainable conservation, community education, reserve support, and not just animals. It may not suit a student who wants pure wildlife research every day. **Review signal** WorkingAbroad reviews are positive; one reviewer described feeling welcomed and called it a life-changing first South Africa experience. Kariega also publishes volunteer testimonials. **Instagram** - `@kariegavolunteers` - `@kariega_foundation` **Sources** - Official: https://kariegafoundation.com/volunteer - Foundation: https://kariegafoundation.com/ - Testimonials: https://www.kariega.co.za/blog/volunteer-testimonials-november-2024-february-2025 - Reviews: https://www.workingabroad.com/projects/kariega-big-five-game-reserve-volunteer-programme-south-africa/reviews/ - Instagram: https://www.instagram.com/kariegavolunteers/ - Instagram: https://www.instagram.com/kariega_foundation/ --- ### 7. SANCCOB **Best for:** a student who wants real animal-care experience with endangered African penguins and seabirds. SANCCOB is one of the clearest choices for students interested in wildlife rehabilitation rather than safari reserve work. The official international volunteer programme offers hands-on experience with endangered African penguins and other seabirds at centres in Cape Town or Gqeberha. Requirements include age 18+, no prior experience, a minimum of 6 weeks, and year-round placements. **Why it stands out** - Long-standing, specialist seabird organisation. - Strong conservation purpose. - Real rehabilitation work. - Clear age and minimum stay. - Cape Town or Gqeberha bases are easier than remote bush camps. **Student fit** Best for students considering veterinary nursing, animal care, marine biology, seabird conservation, wildlife rehabilitation, or NGO work. Expect cleaning, food prep, laundry, feeding support, and repetitive care routines — not glamorous wildlife tourism. **Review signal** External stories and Instagram posts frame it as meaningful, hands-on seabird care. SANCCOB has a visible public reputation and strong African penguin conservation relevance. **Instagram** - `@sanccob` **Sources** - Official volunteer page: https://sanccob.co.za/volunteer/ - Instagram: https://www.instagram.com/sanccob/ - External volunteer story: https://www.ourbetterworld.org/blog/volunteering-south-africa-sanccob-life-changing-experience Jeffrey Eisen on Pexels`} /> --- ### 8. African Penguin & Seabird Sanctuary / Dyer Island Conservation Trust **Best for:** a student interested in African penguins, seabirds, marine ecosystems, and Gansbaai-area conservation. The African Penguin & Seabird Sanctuary is linked to the Dyer Island ecosystem, with social channels showing seabird rehabilitation and volunteer support. It is a good fit for a student who wants seabird work but prefers the Gansbaai / Dyer Island ecosystem rather than SANCCOB’s Cape Town or Gqeberha centres. **Why it stands out** - Specialist seabird rehabilitation. - Strong connection to Dyer Island marine conservation. - Good add-on or alternative to Marine Dynamics Academy. **Student fit** Good for marine students, animal-care students, and anyone interested in African penguin conservation. Clarify whether the placement is directly with APSS/DICT or through a partner/agency. **Instagram** - `@apssza` - `@dyerict` **Sources** - APSS TripAdvisor: https://www.tripadvisor.com/Attraction_Review-g472522-d8722241-Reviews-African_Penguin_Seabird_Sanctuary-Gansbaai_Overstrand_Overberg_District_Western_C.html - Instagram: https://www.instagram.com/apssza/ - Dyer Island Conservation Trust Instagram: https://www.instagram.com/dyerict/ --- ### 9. Marine Dynamics Academy **Best for:** a student who wants marine volunteering with sharks, whales, seabirds, boats, ecotourism, and the Marine Big 5. Marine Dynamics Academy is based in Gansbaai and offers marine volunteer and internship pathways. Their official academy page says the goal is to help future marine ambassadors through skills-based training, specialist training, and learning from ocean experts. The marine volunteer page describes sea-based work across vessels, including shark cage diving, pelagic bird watching, whale watching, and marine research. **Why it stands out** - Strong marine identity. - Good if student wants ocean exposure rather than bush. - Links to Dyer Island Conservation Trust and APSS ecosystem. - Strong social media and tourism infrastructure. **Student fit** Best for students interested in marine conservation, ecotourism, sharks, boats, seabirds, and public education. Because ecotourism is part of the model, a student should ask what percentage of time is research, tourism support, APSS support, data entry, and boat work. **Instagram** - `@marinedynamics` - `@dyerict` - `@apssza` **Sources** - Official Marine Dynamics: https://www.marinedynamics.co.za/ - Academy: https://marinedynamics.org/academy/ - Marine volunteers: https://marinedynamics.org/academy/training/marine-volunteers/ - Sharkwatch volunteer programmes: https://sharkwatchsa.com/volunteer-programs/ - Instagram: https://www.instagram.com/marinedynamics/ --- ### 10. Oceans Research **Best for:** a student who wants the strongest marine research CV signal. Oceans Research is based in Mossel Bay and describes its purpose around research, discovery, education, and conservation. It exposes students and volunteers to species such as white sharks, Cape fur seals, bottlenose dolphins, and humpback dolphins, and VolunteerWorld lists logistics including vessels, acoustic tracking equipment, accelerometers and tags, SCUBA equipment, shark cage, aquarium lab, and research facilities. **Why it stands out** - Strong research positioning. - Good for marine biology students. - VolunteerWorld shows 4.9 from 20 reviews and $937/week for 4–12 weeks at time checked. - Reviews mention seamanship, data entry, species ID, mammal tracking, and shark population dynamics. **Student fit** Best for a student who wants skills relevant to marine research, not just an ocean adventure. It may be more expensive than some options but has strong career relevance. **Instagram** - `@oceansresearch` **Sources** - Official: https://www.oceans-research.com/ - VolunteerWorld reviews/listing: https://www.volunteerworld.com/en/review/oceans-research-institute - Instagram: https://www.instagram.com/oceansresearch/ --- ### 11. DAKTARI Bush School & Wildlife Orphanage **Best for:** a student who wants both environmental education and animal care. DAKTARI is a Bush School and Wildlife Orphanage near Hoedspruit. Its public materials emphasise teaching local children about wildlife and the environment, caring for injured/orphaned animals, and working with local communities. Review stories describe students who may never have seen wild animals before learning about injured animals, wildlife care, and environmental responsibility. **Why it stands out** - Combines community education and wildlife. - Good for students considering teaching, NGO work, conservation education, or community-based conservation. - Good social atmosphere by review signal. **Student fit** Choose DAKTARI if the student wants to work with people and animals. Do not choose it if the student wants pure reserve monitoring or predator telemetry. **Review signal** Go Overseas and VolunteerWorld reviews are positive. A VolunteerWorld 2026 review described feeling welcomed, supported, and included, and praised the students’ warmth and curiosity. **Instagram** - `@daktaribushschool` **Sources** - Official: https://daktaribushschool.org/ - Review/story: https://daktaribushschool.org/mobotse-review-jennyfer/ - Go Overseas reviews: https://www.gooverseas.com/organization/daktari-bush-school-wildlife-orphanage-reviews - VolunteerWorld reviews: https://www.volunteerworld.com/en/review/daktari-bush-school - Instagram: https://www.instagram.com/daktaribushschool/ --- ### 12. Moholoholo Wildlife Rehabilitation Centre **Best for:** a student who wants wildlife rescue, rehabilitation, husbandry, and unpredictable animal-care work. Moholoholo’s volunteer programme describes dynamic, hands-on conservation work, rescue call-outs, rehabilitation of orphaned animals, and work with experienced conservationists. The official page states a minimum stay of two weeks. **Why it stands out** - Long-established Hoedspruit rehabilitation centre. - Good for students interested in wildlife rehab, veterinary support, raptors, mammals, and education. - Stronger animal-care experience than reserve monitoring. **Student fit** Best for animal-care or pre-vet students who are comfortable with cleaning, feeding, enclosure work, emotional cases, and sometimes seeing animals that cannot be released. **Review signal** TripAdvisor includes volunteer reviews such as a two-week volunteer describing the experience positively and noting animal interaction. Because interaction can be ethically complicated, students should distinguish between necessary husbandry/rehab work and tourist handling. **Instagram** - Verify current Moholoholo official handle before booking. **Sources** - Official: https://www.moholoholo.co.za/ - Volunteer programme: https://www.moholoholo.co.za/facility/student-program/ - TripAdvisor volunteer review: https://www.tripadvisor.com/ShowUserReviews-g312616-d2284630-r118983173-Moholoholo_Wildlife_Rehabilitation_Centre-Hoedspruit_Kruger_National_Park.html --- ### 13. Vervet Monkey Foundation **Best for:** a student who wants direct primate rescue, sanctuary, and rehabilitation experience. The Vervet Monkey Foundation’s volunteer page is unusually direct: it says volunteering directly means 100% of the contribution goes into the operating budget, and tasks include working with monkeys, building/painting, preparing food, cleaning bottles and bowls, monitoring monkeys, sickbay work, building new enclosures, invader plant clearing, and other sanctuary tasks. It also states volunteers eat vegan meals. **Why it stands out** - Direct charity/sanctuary route. - Strong practical animal-care workload. - Good for students interested in primates, animal welfare, vegan/ethical living, and sanctuary operations. - Short and long stays are possible, with longer stays usually 4+ weeks. **Student fit** Good for self-motivated students who do not mind physical work and simple living. It is not a safari or Big 5 reserve experience. **Review signal** VolunteerWorld and Go Overseas have positive listings/reviews for VMF. Reviews often focus on the monkey-care experience and sanctuary community. **Instagram** - `@vervetmonkeyfoundation` **Sources** - Official volunteering: https://vervet.za.org/volunteering/ - Short-term stays: https://vervet.za.org/short-term-volunteers-and-tours/ - Go Overseas listing: https://www.gooverseas.com/volunteer-abroad/south-africa/program/165495 - Instagram: https://www.instagram.com/vervetmonkeyfoundation/ --- ### 14. C.A.R.E. Baboon Sanctuary **Best for:** a student specifically interested in baboons, primate rehabilitation, and sanctuary care. C.A.R.E. focuses on orphaned baboons and primate rehabilitation/release. VolunteerWorld reviews include strong praise from some volunteers, though also show mixed ratings. The Instagram account emphasises rescue, rehabilitation, release, and “not pets”. **Why it stands out** - Specialist primate focus. - Good for a student interested in baboons rather than general wildlife. - Strong educational value around human-primate conflict. **Student fit** Best for students prepared for emotional, repetitive, physical animal-care work. Ask about release success, volunteer supervision, safety protocols, and how volunteers interact with infant primates. **Instagram** - `@primate_care` **Sources** - VolunteerWorld reviews: https://www.volunteerworld.com/en/review/c.a.r.e. - Instagram: https://www.instagram.com/primate_care/ --- ### 15. International Primate Rescue **Best for:** primate sanctuary care near Gauteng/Pretoria. International Primate Rescue is another primate-focused option. Public listings describe rescued primates and daily volunteer care. It may be more accessible than remote bush placements for students already based in Gauteng. **Why it stands out** - Primate-specific. - Urban/provincial accessibility. - Good if student wants sanctuary care without going far into remote bush. **Student fit** A possible fit for animal-care students, but ask carefully about direct contact, enrichment, welfare policies, long-term sanctuary model, and release potential. **Instagram** - `@international_primate_rescue` **Sources** - VolunteerWorld reviews: https://www.volunteerworld.com/en/review/international-primate-rescue - TripAdvisor: https://www.tripadvisor.com/Attraction_Review-g312583-d7693451-Reviews-International_Primate_Rescue-Pretoria_Gauteng.html - Instagram: https://www.instagram.com/international_primate_rescue/ --- ### 16. Free To Be Wild Sanctuary **Best for:** broad wildlife rescue/sanctuary experience if the student verifies the current programme directly. Free To Be Wild appears as a sanctuary/rescue option with public social presence and volunteer stories. Because the online trail is less standardised than Wildlife ACT, GVI, SANCCOB, or Oceans Research, it should be treated as a promising but “ask more questions first” option. **Why it could be interesting** - Sanctuary/rescue experience. - Volunteer stories online. - Good for students who want animal care and are not focused only on Big 5. **Questions to ask** - Are animals released where possible? - Which species are permanently resident and why? - Is there public handling? - What is the volunteer schedule? - Who supervises animal care? - What veterinary partners are involved? **Instagram** - `@freetobewildsanctuary` **Sources** - Volunteer stories: https://freetobewildsanctuary.org/tag/free-to-be-wild-sanctuary/ - Instagram: https://www.instagram.com/freetobewildsanctuary/ --- ### 17. Tenikwa Wildlife Rehabilitation & Awareness Centre **Best for:** Garden Route wildlife awareness and rehabilitation exposure. Tenikwa is based near Plettenberg Bay and describes itself as a wildlife awareness centre with rehabilitation facilities for injured and abandoned wildlife from the Garden Route. It also offers accommodation, tours, and conservation-based experiences. That makes it more tourism-facing than a purely field-research project. **Why it stands out** - Garden Route location. - Wildlife rehabilitation and public education. - Easier logistics than remote reserves. - Good for students combining travel and shorter wildlife exposure. **Student fit** Better for wildlife education, rehab interest, and tourism/conservation learning than for strict research experience. Students should ask whether there is a volunteer placement, what tasks are available, and whether the role is through Tenikwa directly or a partner. **Review signal** TripAdvisor includes positive volunteer comments, but many reviews are from day visitors. **Instagram** - `@tenikwa_wildlife` **Sources** - Official: https://tenikwa.com/ - Give & Grow / vet programme listing: https://giveandgrow.world/project/tenikwa-wildlife-rehabilitation-vet-program - TripAdvisor volunteer review: https://www.tripadvisor.com/ShowUserReviews-g1382495-d1025869-r574189747-Tenikwa_Wildlife_Awareness_Centre-The_Crags_Plettenberg_Bay_Western_Cape.html - Instagram: https://www.instagram.com/tenikwa_wildlife/ --- ### 18. Panthera Africa Big Cat Sanctuary **Best for:** a student who specifically wants sanctuary care with rescued big cats, not wild conservation. Panthera Africa markets itself as an ethical big cat sanctuary. Its volunteer page lists daily tasks such as cleaning and maintaining enclosures, food preparation, enrichment, farm work, fence testing, perimeter walking, and educational visits. VolunteerWorld and TripAdvisor show positive volunteer reviews. **Why it stands out** - Big-cat sanctuary experience without claiming to be a wild reserve. - Good if student wants captive sanctuary care and enrichment. - Stronger ethical framing than cub-petting facilities if sanctuary claims are followed in practice. **Student fit** Choose only if the student understands the difference between sanctuary care and conservation fieldwork. It will not teach wild predator monitoring in the same way as Wildlife ACT or GVI. Ask about no breeding, no trade, no cub petting, no walking with predators, and no public handling. **Instagram** - `@pantheraafricasanctuary` **Sources** - Official volunteering: https://pantheraafrica.com/volunteering/ - VolunteerWorld reviews: https://www.volunteerworld.com/en/review/panthera-africa - TripAdvisor: https://www.tripadvisor.com/Attraction_Review-g1187779-d6902930-Reviews-Panthera_Africa_Big_Cat_Sanctuary-Stanford_Overstrand_Overberg_District_Western_.html - Instagram: https://www.instagram.com/pantheraafricasanctuary/ --- ### 19. Cheetah Outreach **Best for:** conservation education and facility work near Cape Town — but not for strict no-contact wildlife ethics. Cheetah Outreach’s volunteer page says volunteers raise awareness, communicate the conservation message to guests, assist with public cheetah and small-animal encounters, prepare food, feed animals, clean enclosures, greet guests, lead tours, and walk dogs. **Why it stands out** - Accessible Cape Town-area location. - Low/no-cost options for some volunteers. - Useful for education, public engagement, and facility operations. - Cheetah conservation messaging and livestock guardian dog work are part of the broader model. **Ethical caution** Because the page explicitly mentions public cheetah and small animal encounters, this is not a fit for students who want strict no-contact wildlife volunteering. Students should ask exactly what “encounters” mean, whether direct contact is involved, and how the programme aligns with South African Tourism/SATSA no-interaction principles. **Instagram** - `@cheetahoutreach.za` **Sources** - Official: https://cheetah.co.za/become-a-volunteer/ - Instagram: https://www.instagram.com/cheetahoutreach.za/ --- ### 20. Hoedspruit Endangered Species Centre **Best for:** endangered species facility exposure, especially cheetah/rhino education — with due diligence. HESC states that it focuses on the survival of endangered species through cheetah bloodlines, rhino rehabilitation, education, and research. It is an established Hoedspruit institution, but a student should verify whether the current opportunity is a volunteer placement, internship, tour, or partner placement. **Why it may interest a student** - Well-known Hoedspruit conservation facility. - Cheetah and rhino focus. - Educational and rehabilitation language. **Ethical caution** Because it is a captive endangered species centre, the student should ask about breeding, contact, ambassador animals, release pathways, and the exact purpose of volunteer labour. **Instagram** - `@hesc_endangeredspeciescentre` **Sources** - Official: https://hesc.co.za/ - Instagram: https://www.instagram.com/hesc_endangeredspeciescentre/ --- ### 21. Cheetah Experience **Best for:** only if the student specifically wants captive cheetah/big-cat facility experience and accepts the ethical due-diligence burden. Cheetah Experience publicly offers volunteer and internship programmes and says more than 1,000 volunteers from 33 countries have joined. Because it is a captive big-cat facility, it sits in a high-diligence category rather than the first-choice conservation list. **Why it might interest a student** - Cheetah/big-cat exposure. - Volunteer/internship programmes are public. - Potentially useful for students interested in captive animal care. **Ethical caution** This should not be recommended as wild conservation unless the facility can clearly explain its welfare, breeding, release, no-trade, and no-public-contact policies. Ask the SATSA decision-tree questions before booking. **Sources** - Official: https://www.cheetahexperience.com/ --- ### 22. African Conservation Experience **Best for:** a student who wants help choosing a placement and wants a broker with conservation-specific positioning. African Conservation Experience is not one volunteer site. It is a conservation travel/placement organisation that places volunteers at partner projects in South Africa and southern Africa. It has been operating for many years and positions itself around genuine conservation projects, animal welfare standards, and matching students to the right placement. **Why it stands out** - Good for students who do not know which project fits them. - Useful for pre-vet, wildlife rehab, field research, and career-focused placements. - Helps with preparation and logistics. **Student fit** A good option if the student wants advice and support, but they should still evaluate the exact host project, not only the broker brand. **Review signal** Go Overseas and VolunteerForever list positive signals for ACE. ACE also publishes traveller stories and references animal welfare standards. **Instagram** - `@africanconservationexperience` **Sources** - Official: https://www.conservationafrica.net/ - Go Overseas: https://www.gooverseas.com/organization/african-conservation-experience-reviews - VolunteerForever: https://www.volunteerforever.com/program/african-conservation-experience/ - Animal welfare standards: https://www.conservationafrica.net/information/animal-welfare - Instagram: https://www.instagram.com/africanconservationexperience/ --- ## Placement agencies and comparison platforms These platforms can be useful, but they are not the conservation project itself. A student should always identify the real host organisation, reserve, sanctuary, or NGO. | Platform | Useful for | Caution | | ----------------------------------- | -------------------------------------------- | ------------------------------------------------------------ | | **VolunteerWorld** | Comparing prices, dates, photos, and reviews | Listings can be polished; verify host directly | | **Go Overseas** | Verified reviews and alumni interviews | Some listings are old or availability may change | | **GoAbroad** | Programme discovery and reviews | Ratings may combine different projects | | **WorkingAbroad** | Conservation-focused project packaging | Ask whether booking direct is possible | | **Pod Volunteer** | Supported volunteering and review pages | Verify the reserve/project name | | **The Great Projects** | Wildlife-focused packages | Strong marketing; verify ethics and exact host | | **IVHQ** | Budget-friendly structure | Often places through local partners; identify actual project | | **GoEco** | Broad range of animal/wildlife projects | Check sanctuary ethics carefully | | **African Impact** | Operator/provider with own project pages | Good support, but still ask for data-use detail | | **African Conservation Experience** | Conservation-specific placement advice | Strong option, but still evaluate exact host | --- ## Programme categories ### A. Best for real field conservation 1. Wildlife ACT 2. GVI Karongwe Wildlife Research 3. African Impact Big 5 4. Siyafunda 5. Kariega Foundation 6. Shamwari Conservation Experience Choose these if the student wants monitoring, reserve work, camera traps, tracking, data collection, habitat work, anti-poaching support, or predator/prey observation. Magda Ehlers on Pexels`} /> ### B. Best for animal care and rehabilitation 1. SANCCOB 2. APSS 3. Moholoholo 4. Vervet Monkey Foundation 5. C.A.R.E. Baboon Sanctuary 6. DAKTARI 7. Tenikwa Choose these if the student wants feeding, cleaning, rehab routines, animal husbandry, nursery/orphan season, and public education. ### C. Best for marine conservation 1. Oceans Research 2. Marine Dynamics Academy 3. SANCCOB 4. APSS / Dyer Island Conservation Trust 5. African Impact Shark, Penguin & Marine Conservation Choose these for marine biology, sharks, seabirds, whales, dolphins, boats, photo-ID, data entry, and ocean fieldwork. ### D. Best for community and education 1. DAKTARI 2. Kariega Foundation 3. Wildlife ACT community/coexistence exposure 4. African Impact community-linked programmes Choose these if the student is interested in conservation education, youth outreach, coexistence, or the human side of conservation. ### E. Best for sanctuary/captive animal care — high diligence 1. Panthera Africa 2. Cheetah Outreach 3. HESC 4. Cheetah Experience 5. Tenikwa 6. International Primate Rescue 7. Free To Be Wild These may be valuable for animal-care experience, but they must be evaluated carefully. Captive wildlife volunteering is the area where greenwashing is most common. --- ## Instagram research guide Instagram is useful, but it is not proof. Use it to check consistency. ### What to look for - Repeated posts showing staff, monitoring, data, releases, rehab routines, education, habitat work, and conservation outcomes. - Tagged alumni who describe actual duties. - Recent posts showing the programme is active. - No cub petting, big-cat walks, elephant rides, predator selfies, or bottle-feeding large carnivores for tourists. - Transparent captions explaining species, data, partners, or conservation purpose. ### Useful handles to review | Programme | Instagram | | ------------------------------- | -------------------------------------------------------- | | Wildlife ACT | https://www.instagram.com/wildlife_act/ | | GVI | https://www.instagram.com/gvitravel/ | | African Impact | https://www.instagram.com/africanimpact/ | | Siyafunda | https://www.instagram.com/siyafundawildlife/ | | Shamwari | https://www.instagram.com/shamwariprivategamereserve/ | | Kariega Volunteers | https://www.instagram.com/kariegavolunteers/ | | Kariega Foundation | https://www.instagram.com/kariega_foundation/ | | SANCCOB | https://www.instagram.com/sanccob/ | | DAKTARI | https://www.instagram.com/daktaribushschool/ | | Marine Dynamics | https://www.instagram.com/marinedynamics/ | | Dyer Island Conservation Trust | https://www.instagram.com/dyerict/ | | APSS | https://www.instagram.com/apssza/ | | Oceans Research | https://www.instagram.com/oceansresearch/ | | Tenikwa | https://www.instagram.com/tenikwa_wildlife/ | | Panthera Africa | https://www.instagram.com/pantheraafricasanctuary/ | | Cheetah Outreach | https://www.instagram.com/cheetahoutreach.za/ | | HESC | https://www.instagram.com/hesc_endangeredspeciescentre/ | | Vervet Monkey Foundation | https://www.instagram.com/vervetmonkeyfoundation/ | | C.A.R.E. Baboon Sanctuary | https://www.instagram.com/primate_care/ | | International Primate Rescue | https://www.instagram.com/international_primate_rescue/ | | Free To Be Wild Sanctuary | https://www.instagram.com/freetobewildsanctuary/ | | African Conservation Experience | https://www.instagram.com/africanconservationexperience/ | --- ## Review research guide Reviews are helpful, but they are biased toward people who had a strong positive or negative experience. Treat reviews as a pattern, not proof. ### Strong review signals found - **Wildlife ACT:** Go Overseas 4.88 / 40 reviews; reviews and alumni interviews repeatedly emphasise real conservation and field monitoring. - **GVI South Africa:** Go Overseas 4.86 / 50 reviews; review summary praises knowledgeable staff and immersive experience. - **African Impact Big 5:** GoAbroad review page includes a 10/10 review praising knowledgeable guides and research/conservation work. - **Shamwari:** WorkingAbroad lists 4.9/5 from 17 reviews; reviews praise accommodation, food, lectures, and activity variety. - **Kariega:** WorkingAbroad reviews positive; Kariega publishes recent volunteer testimonials. - **SANCCOB:** External stories and Instagram show meaningful seabird rehabilitation experiences. - **DAKTARI:** Go Overseas and VolunteerWorld reviews praise community, children, animals, and support. - **Oceans Research:** VolunteerWorld lists 4.9 / 20 reviews; reviewers mention research skills, data entry, seamanship, species ID, and marine-field relevance. - **Panthera Africa:** VolunteerWorld and TripAdvisor reviews are positive, but this is sanctuary care, not wild conservation. - **C.A.R.E.:** VolunteerWorld has positive and some mixed review signals; read carefully. - **VMF:** Public volunteer materials and external listings show a strong sanctuary-care model. ### How to read reviews Look for reviews that mention: - actual daily tasks; - staff supervision; - safety briefings; - data collection; - animal welfare; - accommodation realism; - food and transport; - what was disappointing; - whether the experience matched marketing. Be cautious if reviews only say “amazing animals” and “best time ever” without explaining the conservation work. --- ## Questions a student should ask before paying Send these to every programme: 1. What exact tasks will I do in a normal week? 2. Which species are monitored or cared for? 3. Is the work field research, reserve management, rehabilitation, education, tourism support, or animal husbandry? 4. Who uses the data volunteers collect? 5. Are volunteers ever allowed to touch, hold, walk with, or pose with wild animals? 6. Do you breed, trade, sell, loan, or exchange animals? 7. What happens to animals that cannot be released? 8. What qualifications do supervisors have? 9. What is included in the fee: accommodation, meals, transport, training, insurance, laundry, airport transfers? 10. What is not included? 11. What are the emergency and medical procedures? 12. Is Wi-Fi available? 13. What vaccinations or health precautions are recommended? 14. What visa category do you recommend for my nationality and length of stay? 15. Can you provide a recent example of how volunteer work contributed to conservation? --- ## Student logistics ### Visa and length of stay Visa rules depend on nationality and trip length. South African government guidance says standard visitor visas are generally for up to 90 days for tourism or business, while South African mission guidance states that people volunteering for under 90 days apply for a visitor’s visa and longer unpaid volunteer/charitable work may require the volunteer visa/permit route. Always check with the nearest South African embassy, consulate, or official visa centre before booking. ### Insurance The student should have travel insurance that covers: - volunteering; - wildlife/rural work; - 4x4 travel; - boat work if marine; - medical evacuation; - trip cancellation; - lost baggage; - personal liability. ### Health Common considerations: - malaria risk in parts of Limpopo, Mpumalanga, and KwaZulu-Natal; - heat and dehydration; - tick bite fever; - sun exposure; - rabies risk if working with wildlife; - seasickness for marine programmes; - allergies and asthma in dusty bush camps. A travel clinic should advise based on exact location and season. ### Safety Good programmes should have: - airport transfer instructions; - emergency contact numbers; - induction and safety briefing; - rules for dangerous wildlife; - no walking alone in unfenced bush; - PPE where relevant; - vehicle and radio protocols; - safe accommodation. --- ## Choosing by student personality | Student type | Best choices | | ----------------------------------------- | --------------------------------------------------------------- | | Wants serious conservation career | Wildlife ACT, GVI, Oceans Research, SANCCOB, Kariega, Siyafunda | | First time overseas and nervous | GVI, African Impact, Shamwari, Kariega, SANCCOB | | Wants remote bush and fieldwork | Wildlife ACT, Siyafunda, African Impact Big 5, GVI | | Wants marine biology | Oceans Research, Marine Dynamics, SANCCOB, APSS | | Wants animal-care/vet-adjacent experience | SANCCOB, Moholoholo, VMF, C.A.R.E., DAKTARI | | Wants community education | DAKTARI, Kariega | | Wants Cape Town/Garden Route convenience | SANCCOB, Cheetah Outreach, Tenikwa, Panthera Africa | | Wants budget/direct charity option | SANCCOB, VMF, Cheetah Outreach, DAKTARI — verify current fees | | Wants lots of animal contact/photos | Reconsider. Ethical wildlife work usually limits contact. | --- ## My ranked shortlist for a student ### 1. Wildlife ACT — best overall for conservation credibility Most suitable for a student who wants real field monitoring, endangered species, and ranger-style learning. Strong review profile, clear field model, and small teams. ### 2. SANCCOB — best animal-care conservation placement Best if the student wants hands-on rehab and can commit to six weeks. Strong conservation mission and practical daily work. ### 3. Oceans Research — best marine research placement Best for marine biology students who want research skills and shark/marine ecosystem exposure. ### 4. Kariega Foundation — best balanced conservation/community programme Best for students who understand that conservation includes local communities. ### 5. GVI Karongwe — best structured first-time student programme Best for support, structure, and clear logistics. ### 6. African Impact Big 5 — best social Big 5 conservation experience Best for students wanting Greater Kruger, Big 5 exposure, and a polished international volunteer environment. ### 7. Siyafunda — best rugged field alternative Potentially excellent for a more independent student who wants practical reserve monitoring and is comfortable doing direct due diligence. ### 8. DAKTARI — best conservation education option Best for students who want children, environmental education, community impact, and animal care in one placement. --- ## Programmes to approach with extra caution These are not automatic “no” options, but students should ask more questions. | Programme type | Why caution is needed | | --------------------------------------------------------------- | --------------------------------------------------------------------------------------------------- | | Captive big-cat centres | South Africa has a history of cub-petting, predator interaction, and canned-hunting-linked tourism. | | “Cheetah encounter” programmes | Public contact can conflict with strict no-interaction ethics. | | Sanctuaries with lots of baby animals | Ask whether animals are genuinely orphaned/injured and whether breeding occurs. | | Programmes that promise “hands-on” predator contact | Usually a red flag. | | Programmes that do not disclose the reserve/site before booking | Hard to verify ethics, safety, and impact. | | Programmes with only Instagram evidence | Social media is not enough. Ask for impact reports, partners, or references. | --- ## Packing notes for students For bush/reserve programmes: - neutral-coloured field clothes; - warm layers for winter mornings; - hat, sunscreen, sunglasses; - headlamp; - sturdy closed shoes; - reusable water bottle; - notebook and pen; - binoculars if possible; - camera with extra batteries; - insect repellent; - basic first-aid kit; - offline maps; - power bank. For marine programmes: - windbreaker; - seasickness tablets; - waterproof bag; - warm layers; - quick-dry clothing; - reef-safe sunscreen; - polarised sunglasses. For rehab/sanctuary work: - clothes that can get dirty; - closed shoes or boots; - gloves if requested; - old towels if requested; - patience for repetitive tasks. --- ## Final decision matrix Choose **Wildlife ACT** if the student wants real endangered-species monitoring. Choose **SANCCOB** if the student wants hands-on seabird rehab. Choose **Oceans Research** if the student wants marine research. Choose **Kariega** if the student wants conservation plus community. Choose **GVI** if the student wants structure and support. Choose **African Impact** if the student wants social volunteer life and Greater Kruger. Choose **Siyafunda** if the student wants a direct, rugged wildlife-monitoring feel. Choose **DAKTARI** if the student wants conservation education with children and animals. Choose **VMF or C.A.R.E.** if the student wants primates. Choose **Panthera Africa** only if the student wants sanctuary care and accepts that captive big-cat work is not the same as wild conservation. --- ## Source list ### Official programme and organisation pages - Wildlife ACT: https://www.wildlifeact.com/ - Wildlife ACT endangered species volunteer programme: https://www.wildlifeact.com/volunteer/program/endangered-species-conservation-south-africa - Wildlife ACT why volunteer: https://www.wildlifeact.com/why-volunteer - GVI South Africa wildlife research: https://www.gvi.ie/programs/wildlife-research-south-africa-expedition/ - GVI South Africa overview: https://www.gvi.ie/volunteer-in-south-africa/ - African Impact Big 5 Wildlife Conservation: https://africanimpact.com/volunteer-projects/wildlife-conservation/african-big-5-wildlife-conservation/ - African Impact Shark, Penguin & Marine Conservation: https://africanimpact.com/volunteer-projects/marine-conservation/shark-conservation-volunteering-in-south-africa/ - Siyafunda: https://www.siyafundaconservation.com/ - Shamwari Conservation Experience PDF: https://www.shamwari.com/wp-content/uploads/pdf/nAE8Hr307Ti6oAK7BlIGjJZRf81PLA1RqmCGNjqrE1M8Ijw9.pdf - Kariega Foundation volunteer programme: https://kariegafoundation.com/volunteer - Kariega Foundation: https://kariegafoundation.com/ - SANCCOB volunteer page: https://sanccob.co.za/volunteer/ - DAKTARI: https://daktaribushschool.org/ - Moholoholo volunteer programme: https://www.moholoholo.co.za/facility/student-program/ - Moholoholo: https://www.moholoholo.co.za/ - Marine Dynamics Academy: https://marinedynamics.org/academy/ - Marine Dynamics marine volunteers: https://marinedynamics.org/academy/training/marine-volunteers/ - Oceans Research: https://www.oceans-research.com/ - Vervet Monkey Foundation volunteering: https://vervet.za.org/volunteering/ - Vervet Monkey Foundation short-term stays: https://vervet.za.org/short-term-volunteers-and-tours/ - Panthera Africa volunteering: https://pantheraafrica.com/volunteering/ - Cheetah Outreach volunteering: https://cheetah.co.za/become-a-volunteer/ - HESC: https://hesc.co.za/ - Tenikwa: https://tenikwa.com/ - Cheetah Experience: https://www.cheetahexperience.com/ - African Conservation Experience: https://www.conservationafrica.net/ - ACE animal welfare standards: https://www.conservationafrica.net/information/animal-welfare ### Review and comparison pages - Wildlife ACT Go Overseas: https://www.gooverseas.com/volunteer-abroad/south-africa/wildlife-act/21946 - GVI Go Overseas: https://www.gooverseas.com/volunteer-abroad/south-africa/global-vision-international/52250 - African Impact GoAbroad: https://www.goabroad.com/providers/african-impact/programs/african-big-5-wildlife-conservation-project-south-africa-108826 - Shamwari WorkingAbroad reviews: https://www.workingabroad.com/projects/shamwari-game-reserve-conservation-volunteer-programme-south-africa/reviews/ - Kariega WorkingAbroad reviews: https://www.workingabroad.com/projects/kariega-big-five-game-reserve-volunteer-programme-south-africa/reviews/ - DAKTARI Go Overseas: https://www.gooverseas.com/organization/daktari-bush-school-wildlife-orphanage-reviews - DAKTARI VolunteerWorld: https://www.volunteerworld.com/en/review/daktari-bush-school - Oceans Research VolunteerWorld: https://www.volunteerworld.com/en/review/oceans-research-institute - Panthera Africa VolunteerWorld: https://www.volunteerworld.com/en/review/panthera-africa - C.A.R.E. VolunteerWorld: https://www.volunteerworld.com/en/review/c.a.r.e. - Vervet Monkey Foundation Go Overseas: https://www.gooverseas.com/volunteer-abroad/south-africa/program/165495 - International Primate Rescue VolunteerWorld: https://www.volunteerworld.com/en/review/international-primate-rescue - APSS TripAdvisor: https://www.tripadvisor.com/Attraction_Review-g472522-d8722241-Reviews-African_Penguin_Seabird_Sanctuary-Gansbaai_Overstrand_Overberg_District_Western_C.html - Tenikwa TripAdvisor review: https://www.tripadvisor.com/ShowUserReviews-g1382495-d1025869-r574189747-Tenikwa_Wildlife_Awareness_Centre-The_Crags_Plettenberg_Bay_Western_Cape.html - Marine Dynamics Go Overseas: https://www.gooverseas.com/organization/marine-dynamics-reviews - VolunteerWorld South Africa animal volunteering overview: https://www.volunteerworld.com/en/volunteer-abroad/animal-south-africa - GoAbroad South Africa volunteer guide: https://www.goabroad.com/volunteer-abroad/search/south-africa/volunteer-abroad-1 ### Ethics and visa sources - SATSA Captive Wildlife Guide: https://www.satsa.com/sites/default/files/2023-10/SATSA%20Captive%20Wildlife%20Guide.pdf - SATSA decision tree: https://www.wildchoices.org/wp-content/uploads/2022/02/SATSA_Tool_for_Assessing_Captive_Wildlife_Attractions__Activities.pdf - South African Tourism animal interaction position: https://southafrica.net/zm/en/travel/article/say-no-to-animal-interaction - Wildlife ACT ethical volunteering article: https://www.wildlifeact.com/blog/why-ethical-wildlife-volunteering-is-crucial-for-conservation - South African government visa page: https://www.gov.za/services/temporary-residence/visa - South African High Commission volunteer permit guidance: https://dirco.gov.za/canberra/volunteer-permit/

The State of AI in Wildlife: A Global Field Guide to the Smartest Things Happening in Conservation Tech

Sun, 14 Jun 2026 00:00:00 GMT

import BlogImage from "../../components/blog/BlogImage.astro"; import wildlifeSavanna from "../../assets/blog/state-of-ai-in-wildlife/13129167_g_n.jpg"; import droneWildlife from "../../assets/blog/state-of-ai-in-wildlife/17471029_chris_clark.jpg";

Wildlife conservation is no longer short of sensors. Camera traps, acoustic recorders, satellite tags, drones, Earth observation satellites, electronic monitoring systems on fishing vessels, eDNA samplers, and citizen science apps are producing more biodiversity data than humans can review by hand. The bottleneck has shifted from{" "} collecting observations to{" "} turning observations into decisions.

That is where artificial intelligence is becoming genuinely useful. The best work in AI for wildlife is not about replacing ecologists, rangers, taxonomists, or local knowledge holders. It is about helping them move faster: filtering millions of empty camera-trap frames, detecting animals in real time, matching an individual whale or zebra across years of photographs, listening for rare birds in thousands of hours of audio, mapping habitats from satellites, and flagging vessel behaviour that would otherwise remain invisible.

This post is a global snapshot of the field in 2026: what is mature, what is experimental, what is genuinely exciting, and what still needs caution. The short version is simple: AI is becoming the connective tissue between wildlife sensors and conservation action.

Source note: This post was researched from official project documentation, peer-reviewed papers and preprints, conservation-technology organisations, Google Research, Microsoft AI for Good, Conservation X Labs/Wild Me, Cornell/BirdNET, Earth Species Project, Project CETI, Global Fishing Watch, EarthRanger, SMART, The Nature Conservancy, NASA/IBM, Google DeepMind, and recent AI-for-conservation literature. Sources are linked throughout and listed again at the end.

The Big Picture

The field is moving through three waves.

Wave	What changed	Typical tools	What it enables
Wave 1: automation	AI sorts and labels data humans already collected.	MegaDetector, SpeciesNet, BirdNET, Wildlife Insights, Wildbook.	Millions of images and recordings become usable without years of manual review.
Wave 2: real-time sensing	AI moves from the lab or cloud to the field edge.	TrailGuard AI, WPS cameras, SPARROW, cellular camera traps, on-board fisheries AI.	Rangers and managers can respond while the event is still happening.
Wave 3: foundation models	Large reusable models learn general representations of species, sounds, video, or landscapes.	BioCLIP, NatureLM-audio, AlphaEarth Foundations, Prithvi, SAM/SA-FARI.	Small local datasets can be connected to global models, making new projects cheaper to start.

The most important shift is that AI is becoming less of a one-off research model and more of an operational layer: a set of practical services that help conservation teams monitor populations, understand threats, and justify decisions with evidence.

G N on Pexels`} />

Where AI Is Already Useful

Some parts of the field are already operational. Others are still research-heavy. A realistic state-of-the-field map looks like this:

Area	Current maturity	What AI does well	Main caution
Camera-trap filtering	Operational	Finds animals, people, vehicles; removes blank images; speeds up review.	Species ID still needs local validation.
Camera-trap species classification	Operational but uneven	Labels common species across many regions.	Rare, cryptic, nocturnal, juvenile, and out-of-distribution species remain hard.
Individual animal re-identification	Operational for patterned species	Matches whales, sharks, zebras, giraffes, leopards, turtles and other visually distinctive individuals.	Requires good image quality, consistent matching workflows, and expert review.
Bioacoustic species detection	Operational for birds and some taxa	Detects calls in huge sound archives; helps build soundscape indicators.	Background noise, overlapping calls, geography, and behaviour context matter.
Animal communication AI	Early but exciting	Finds structure in whale, dolphin, bird, and other animal vocalisations.	“Translation” claims should be treated carefully.
Drones and aerial surveys	Growing	Counts large animals, detects thermal signatures, surveys dangerous or remote places.	Regulation, disturbance, battery life, and false detections are major constraints.
Anti-poaching and ranger support	Operational in selected sites	Predicts risk, routes patrols, sends alerts, integrates field observations.	Data bias and ranger safety must be central.
Fisheries and ocean transparency	Operational and expanding	Classifies vessel behaviour, detects dark fleets, reviews electronic-monitoring footage.	AIS manipulation, jurisdiction, and enforcement capacity remain hard.
Satellite habitat modelling	Operational	Maps land cover, habitat condition, deforestation, fragmentation, water, fire, and change.	Remote sensing sees habitat proxies, not always animals directly.
Multimodal biodiversity intelligence	Emerging	Combines camera traps, audio, drones, GPS collars, satellites, eDNA, and field records.	Standards, data governance, privacy, and interoperability are the bottleneck.

1. Camera Traps: The Workhorse of Wildlife AI

Camera traps are probably the most mature area of AI in wildlife. They are cheap enough to deploy widely, passive enough to monitor elusive animals, and productive enough to create a data avalanche. The challenge is that a single deployment can produce hundreds of thousands or millions of images, many of them empty.

MegaDetector, built by Microsoft’s AI for Good Lab, is one of the most important practical tools in this space. It does not try to identify every species. It detects animals, people and vehicles in camera-trap imagery, draws bounding boxes, and assigns confidence scores. That simple job is incredibly useful: it removes blanks, flags human activity, and gives downstream species models a cleaner set of images to analyse.

Source:{" "} MegaDetector documentation .

SpeciesNet, developed by Google and originally served through Wildlife Insights, is the next layer up: a camera-trap species classifier. Google describes the open-source SpeciesNet model as classifying nearly 2,500 mammal, bird and reptile categories, trained from 65 million labelled camera-trap images contributed by conservation partners. It works alongside MegaDetector: first find the animal, then classify the animal.

Source:{" "} Google Research SpeciesNet post ;{" "} SpeciesNet GitHub repository .

Wildlife Insights is the platform story around this. It is a cloud platform for managing, analysing and sharing camera-trap data, with machine-learning tools built into the workflow. For many conservation teams, that matters more than the model itself. A great model is not useful if teams cannot upload data, manage metadata, review predictions, export records, and cite datasets.

Source:{" "} Wildlife Insights .

PyTorch-Wildlife is part of the open tooling layer. It wraps wildlife detection and classification models into a more accessible PyTorch framework and model zoo, with documented use cases including Amazon rainforest species recognition and invasive opossum recognition in the Galápagos. This is important because conservation AI should not only live inside closed platforms.

Source:{" "} PyTorch-Wildlife paper ;{" "} PyTorch-Wildlife documentation .

Why this is cool

Camera-trap AI turns years of image review into days or weeks. It lets conservation teams ask bigger questions: not only “what species did we see?” but “how are animals using this corridor?”, “has activity shifted after fire?”, “are predators avoiding people?”, “are invasive species increasing?”, and “where should the next patrol or restoration effort go?”.

What still goes wrong

Camera-trap AI is not magic. Models can fail when the camera angle, region, vegetation, night illumination, animal age, animal posture, or species mix differs from the training data. Recent research has shown that training data quality and size can affect downstream ecological metrics, not just prediction accuracy. In other words: a model can look good in machine-learning terms and still bias an ecological conclusion if the validation is weak.

Source:{" "} training data quality and ecological metrics paper .

2. From Global Models to Local Wildlife Intelligence

The most interesting pattern in camera-trap AI is the movement from one global model to many locally adapted workflows. A global model like SpeciesNet or MegaDetector gives a project a strong starting point. But local teams still need to tune thresholds, add missing species, review uncertain predictions, and keep a human-in-the-loop pipeline.

This is where the “smart people” are doing some of their best work: not just building bigger models, but designing practical review systems, regional models, active-learning loops, and workflows for non-technical ecologists.

Approach	Example	Why it matters
Global detector	MegaDetector	Useful almost everywhere because animal/person/vehicle detection generalises better than species ID.
Global species classifier	SpeciesNet	Gives many projects a strong first pass across thousands of taxa.
Local or regional species model	UK mammal/bird models, African savanna models, Galápagos invasive-species models.	Improves performance where global models miss local fauna or confuse similar species.
Review UI and data platform	Wildlife Insights, Agouti, Camelot, TRAPPER, WildTrax, custom dashboards.	Turns predictions into auditable ecological records.
Active learning	Retrain on local mistakes and uncertain cases.	Reduces labelling burden while improving project-specific accuracy.

A good 2026 wildlife AI workflow is rarely “upload images and trust the model.” It is more like: detect animals, classify likely species, prioritise uncertain cases for review, export validated records, protect sensitive locations, and report uncertainty.

3. Video, Tracking and Segmentation: From Still Images to Behaviour

Still-image classification is useful, but ecology often needs movement and behaviour. Is that elephant feeding, walking, drinking, limping or interacting? Did the same leopard pass twice? Are multiple animals moving as a group? This is where video analysis, segmentation and multi-animal tracking become important.

The SA-FARI dataset, released by Conservation X Labs and Meta collaborators, is a major step in this direction. It stands for{" "} Segment Anything in Footage of Animals for Recognition and Identification . The dataset contains 11,609 camera-trap videos from 741 locations across four continents, spanning 99 species categories, with dense annotations for multi-animal tracking, segmentation masks, boxes and species labels.

Source:{" "} SA-FARI paper ;{" "} SA-FARI project page .

This connects to a broader shift in computer vision: promptable segmentation models such as Segment Anything and its successors are making it easier to outline animals, track them through frames, and build training data for species and behaviour tasks. The conservation value is not just a cleaner mask. It is the possibility of measuring body condition, group size, movement paths, interaction rates and habitat use at much finer temporal resolution.

What to watch

The next frontier is not only “which species is this?” It is “what is happening in the scene?” That means behaviour recognition, animal tracking through time, scene context, and eventually ecological summarisation. Some research is already combining object detectors with vision-language models to generate richer reports from camera-trap data, but this should be treated as decision support, not as ecological truth.

Source:{" "} context-rich automated biodiversity assessment paper .

4. Individual Animal ID: Turning Photos into Population Science

One of the most powerful uses of AI in wildlife is individual re-identification. Instead of only asking “how many zebra photos do we have?”, individual ID asks “which zebra is this, and where else has it been seen?” That unlocks mark-recapture analysis, survival estimates, social networks, migration routes and population trends.

Wildbook, from Wild Me / Conservation X Labs, is the flagship example. It is an open-source software framework for mark-recapture, molecular ecology and social ecology studies. Wildbook uses computer vision and machine learning to locate animals, identify species and suggest individual matches inside a database. Conservation X Labs states that Wildbook supports more than 250 species and more than 1.4 million sightings worldwide.

Source:{" "} Wildbook ;{" "} Wild Me Wildbook overview ;{" "} Conservation X Labs Wild Me Lab .

This is especially useful for animals with natural markings: whale flukes, whale shark spots, zebra stripes, giraffe coat patterns, leopard rosettes, manta ray markings, turtle faces, seal scars and more. The AI does not “know” an animal personally. It compares visual features and proposes candidate matches. Human experts still review those matches.

Why this matters

Individual ID is where AI becomes population monitoring rather than just image labelling. It can turn tourist photos, researcher images and camera-trap records into long-term histories of living animals. That is scientifically powerful and emotionally powerful: every datapoint can become a known individual with a life history.

5. Bioacoustics: Listening to the Planet at Scale

If camera traps are the eyes of AI conservation, bioacoustics is its hearing. Audio recorders can run for months, operate at night, cover dense forests where cameras see little, and detect species that are vocal but rarely photographed. The challenge is that one small recorder can produce thousands of hours of audio.

BirdNET, from the Cornell Lab of Ornithology and Chemnitz University of Technology, is one of the most visible success stories. It uses AI to identify bird species from sound and has become both a public app and a scientific tool. The official BirdNET project frames it as AI-powered sound ID for citizen scientists and bioacoustic research.

Source:{" "} BirdNET .

Arbimon, developed by Rainforest Connection, is another important platform. It helps researchers upload, store and analyse large acoustic datasets, including species-specific and multi-species models. This kind of infrastructure matters because bioacoustics needs more than a recogniser: it needs long-term storage, site management, time-of-day analysis, soundscape metrics and repeatable workflows.

Source:{" "} Arbimon ;{" "} Rainforest Connection ecoacoustics .

NatureLM-audio, from Earth Species Project, points to the next phase: foundation models for bioacoustics. Rather than training a narrow detector for one species, NatureLM-audio is designed for flexible bioacoustic tasks such as species classification, detection and captioning. Earth Species Project describes it as the first large audio-language model tailored specifically for animal sounds.

Source:{" "} Earth Species Project NatureLM-audio announcement ;{" "} NatureLM-audio on Hugging Face .

Why this is cool

Bioacoustics can turn forests, reefs, wetlands and farms into measurable soundscapes. It can reveal when birds return after restoration, whether frogs are breeding, whether gunshots or chainsaws are present, and how human noise changes animal behaviour.

What still goes wrong

Audio is messy. Wind, rain, insects, vehicles, overlapping calls and unknown species can fool models. Bioacoustic AI should usually be treated as a detection-assistance system unless the project has local validation data.

6. Animal Communication: The Most Sci-Fi Part, but Also the Easiest to Overhype

Some of the most fascinating work is happening at the boundary between bioacoustics, linguistics, machine learning and animal behaviour. The question is no longer only “which species made this sound?” but “what structure exists in the communication system?”

Project CETI is applying machine learning, robotics and field biology to sperm whale communication, especially in Dominica. Their goal is to listen to, model and eventually better understand sperm whale codas: patterned sequences of clicks used in social communication.

Source:{" "} Project CETI .

DolphinGemma, announced by Google in collaboration with the Wild Dolphin Project and Georgia Tech, is a large language model trained on dolphin audio to help scientists study dolphin communication. Google describes it as a model that can help learn the structure of dolphin vocalisations and generate dolphin-like sound sequences for research.

Source:{" "} DolphinGemma model page ;{" "} Google DolphinGemma blog .

This work is genuinely exciting. But it needs careful language. AI can find statistical structure, cluster call types, model sequences and suggest hypotheses. That is not the same as “translating whale language into English.” The best scientists in this space are careful because meaning requires behavioural context, social context, field experiments and ethics.

7. Drones, Aircraft and Aerial AI

Aerial surveys are expensive, dangerous and slow when done manually. AI is helping turn drone, aircraft and thermal imagery into animal counts, nest detections, carcass detections, anti-poaching intelligence and habitat assessments.

SCOUT, from Wild Me / Conservation X Labs, is an open hardware and open-source software solution for aerial surveys of wildlife and forests. It is designed to support analysis of large volumes of imagery from aerial surveys, making wildlife population assessments cheaper, safer and more accurate.

Source:{" "} SCOUT .

A 2025 review of drones and AI-driven solutions for wildlife monitoring describes applications across species identification, animal tracking, movement analysis, anti-poaching, population estimation and habitat assessment. The pattern is clear: drones expand the observation area, and AI reduces the image-analysis bottleneck.

Source:{" "} Drones and AI-driven wildlife monitoring review .

Where it shines

Drones are especially useful for open landscapes, wetlands, seabird colonies, large mammals, thermal detection, marine megafauna, nest monitoring and inaccessible terrain.

Where it struggles

Dense canopy, weather, flight permissions, battery life, animal disturbance and small species detection remain practical constraints. Aerial AI also needs careful ground-truthing: a false count can mislead management.

chris clark on Pexels`} />

8. Edge AI: Moving Intelligence into the Field

Most early wildlife AI worked after the field campaign: collect SD cards, drive back, upload data, run models. Edge AI changes the timing. It puts inference on or near the sensor, so a camera or recorder can decide what matters before sending data.

SPARROW, from Microsoft’s AI for Good Lab, stands for Solar-Powered Acoustic and Remote Recording Observation Watch. Microsoft describes it as an AI-powered edge computing solution designed to work autonomously in remote places, collecting camera-trap, acoustic and environmental data and processing it with wildlife AI models on power-efficient edge GPUs.

Source:{" "} Microsoft SPARROW announcement ;{" "} SPARROW GitHub repository .

TrailGuard AI and{" "} Wildlife Protection Solutions show the anti-poaching and real-time alerting side. TrailGuard uses small concealed cameras with onboard AI to detect humans, wildlife or vehicles and transmit relevant images. WPS describes AI-enabled camera systems for continuous real-time monitoring of protected areas.

Sources:{" "} RESOLVE TrailGuard AI ;{" "} Wildlife Protection Solutions .

A 2026 review of edge AI in biodiversity monitoring found growing research across acoustic, vision, tracking and multimodal systems, while also highlighting fragmented adoption and trade-offs among power consumption, compute, communication and ecological inference.

Source:{" "} Future of Edge AI in biodiversity monitoring review .

Why edge AI matters

In conservation, timing matters. A poacher alert three weeks later is an archive. A cattle-lion conflict alert the next morning may be too late. Edge AI is exciting because it can move wildlife monitoring from retrospective reporting toward response.

9. Protected Area Platforms: From Data Points to Decisions

AI models are not enough. Rangers and conservation managers need systems that combine observations, patrols, incidents, GPS collars, camera traps, vehicles, alerts, maps and reports. This is where protected-area platforms become central.

EarthRanger, developed by Vulcan and now associated with Allen Institute for AI/Ai2 support, is a software platform for real-time conservation operations, ecosystem monitoring and protected area management. It helps managers integrate data streams and make operational decisions.

Source:{" "} EarthRanger ;{" "} EarthRanger Methods in Ecology and Evolution paper .

SMART, the Spatial Monitoring and Reporting Tool, is an open-source, non-proprietary conservation tool suite for collecting, storing, communicating and evaluating data on wildlife and conservation areas. It is widely used by rangers and protected area teams.

Source:{" "} SMART Conservation Tools .

PAWS, the Protection Assistant for Wildlife Security, is one of the best-known examples of AI for patrol planning. The research uses machine learning, uncertainty modelling and game-theoretic ideas to identify poaching risk and plan patrol routes under limited resources. Field-test papers report improved detection of snares in some settings, but the most important lesson is that AI must be embedded in ranger knowledge, field constraints and safety.

Source:{" "} PAWS field-test paper ;{" "} PAWS remote-sensing enhancement paper .

10. Ocean AI: Fishing, Whales and the Invisible Sea

The ocean is hard to monitor because it is vast, mobile, legally complex and often hidden from public view. AI is starting to make it more visible.

Global Fishing Watch uses vessel tracking, satellite data and machine learning to classify vessel behaviour and make fishing activity more transparent. Their technology page describes models for vessel classification and fishing activity, while a 2024 study used machine learning and satellite imagery to reveal large amounts of industrial vessel traffic and offshore infrastructure not visible in public monitoring systems.

Source:{" "} Global Fishing Watch technology ;{" "} 2024 global ocean activity study release .

The Nature Conservancy is working on edge AI for electronic monitoring in fisheries. Instead of reviewing months of vessel video footage after a fishing trip, onboard AI can flag fishing events, target catch and bycatch for faster human verification.

Source:{" "} The Nature Conservancy AI monitoring in industrial fishing .

Whale Safe, from the Benioff Ocean Science Laboratory at UC Santa Barbara, is an AI-powered whale detection and ship-risk system. It integrates acoustic detections, visual sightings, habitat model predictions and ship data to provide near-real-time whale and ship information in busy shipping regions.

Source:{" "} Whale Safe .

Why this matters

The ocean has historically been under-observed. AI does not solve governance, but it changes the evidence base. It makes illegal, unreported and unregulated fishing harder to hide. It makes whale presence more visible to shipping operators. It makes bycatch monitoring less dependent on slow manual review.

11. Satellite AI: Habitat, Pressure and Planetary Context

Wildlife does not exist without habitat. That is why Earth observation AI belongs in the wildlife conversation. Satellites do not usually identify individual animals, but they can map the habitat conditions that determine whether species can survive: forest loss, grassland condition, water availability, fire scars, agriculture, roads, settlements, coastlines and climate stress.

AlphaEarth Foundations, announced by Google DeepMind in July 2025, is a geospatial foundation model that integrates large volumes of Earth observation data into reusable representations. The practical output in Earth Engine is the{" "} Google Satellite Embedding dataset: annual 10 m, 64-dimensional embeddings designed for clustering, classification, similarity search and change detection.

Source:{" "} Google DeepMind AlphaEarth Foundations ;{" "} Google Satellite Embedding Earth Engine dataset .

Prithvi, the NASA-IBM geospatial foundation model family, is another key example. NASA describes Prithvi as an open-source geospatial foundation model developed with IBM Research and the Jülich Supercomputing Centre, with applications across Earth observation tasks.

Source:{" "} NASA Prithvi applications ;{" "} IBM-NASA geospatial models on Hugging Face .

For wildlife teams, geospatial foundation models matter because they reduce the need to start from raw satellite bands every time. A conservation scientist with a small number of field plots, animal locations or habitat labels can use pretrained embeddings to build maps faster.

The key caution

Satellite AI maps habitat, not certainty. A “suitable habitat” map is a hypothesis. It needs field data, species knowledge and local validation.

12. Citizen Science: Millions of People as Sensors

AI in wildlife is not only for researchers with grants. Citizen science platforms are turning public observations into massive biodiversity datasets, and AI is making participation easier.

iNaturalist is the obvious example. Its computer-vision system suggests identifications from user-uploaded photos, while the community helps confirm observations. iNaturalist’s help documentation explains how taxa are included in computer-vision training, and a 2025 update noted that the model grows as new taxa meet photo and observation thresholds.

Source:{" "} iNaturalist computer vision taxa guidance ;{" "} iNaturalist 2025 model update .

Citizen science data also trains broader models. The original iNaturalist computer-vision dataset helped expose the difficulty of fine-grained species classification: natural-world datasets are imbalanced, long-tailed and filled with visually similar species.

Source:{" "} iNaturalist Species Classification and Detection Dataset paper .

The broader lesson is that AI and citizen science are a powerful loop: people produce observations, AI lowers the barrier to participation, experts and communities validate, and validated records improve future models.

13. Biological Foundation Models: AI That Understands the Tree of Life

A major shift is the rise of models trained not for one species or one reserve, but for broad biological representation.

BioCLIP is a vision foundation model for the tree of life. It was trained using biological images and taxonomic structure to perform general organismal image tasks across animals, plants and fungi. Microsoft Research describes BioCLIP as leveraging the TreeOfLife-10M dataset, while the paper reports strong zero-shot and few-shot performance improvements over general-purpose CLIP-style baselines.

Source:{" "} Microsoft Research BioCLIP ;{" "} BioCLIP project page .

BioCLIP 2 pushes this further with a much larger TreeOfLife-200M dataset and explores emergent biological structure in the embedding space, including ecological and trait-related information.

Source:{" "} BioCLIP 2 paper .

These models matter because wildlife science is long-tailed. There are many rare species, many local contexts and many datasets with only a few labels. Foundation models can make few-shot and transfer-learning workflows more practical.

What they do not solve

A foundation model is not a taxonomist. It may know visual similarity but not local field context, subspecies boundaries, hybridisation, age/sex classes, cryptic species, or the political sensitivity of a location record.

14. eDNA and AI: The Invisible Biodiversity Layer

Environmental DNA, or eDNA, detects genetic material that organisms shed into water, soil, air or sediment. It is not always framed as “wildlife AI,” but it increasingly belongs in the same monitoring stack. Machine learning can help interpret high-dimensional sequence data, identify species, detect invasive or rare taxa, and combine eDNA with GIS and remote sensing.

Recent reviews describe eDNA as a sensitive, efficient and non-invasive biodiversity-monitoring tool, while newer AI-GIS-eDNA frameworks are being proposed for real-time freshwater health assessment and broader ecosystem monitoring.

Sources:{" "} 2025 eDNA biodiversity monitoring review ;{" "} AI-GIS-eDNA framework review .

The exciting idea is that AI could eventually combine visible evidence from cameras, audible evidence from recorders, molecular evidence from eDNA, and landscape evidence from satellites. That is a much richer picture of biodiversity than any one sensor can provide.

15. Movement, Collars and Behaviour Recognition

Animal-borne sensors are becoming another AI frontier. GPS collars, satellite tags, accelerometers and other biologgers can record where animals move, how fast they move, and sometimes what they are doing.

Movebank, hosted by the Max Planck Institute of Animal Behavior, is a major data infrastructure for animal tracking. It helps researchers manage, share, protect, analyse and archive animal tracking and sensor data.

Source:{" "} Movebank .

Machine learning is increasingly used to classify animal behaviour from accelerometer data. The Bio-logger Ethogram Benchmark, for example, provides datasets and evaluation tasks for computational analysis of animal behaviour using animal-borne tags.

Source:{" "} Bio-logger Ethogram Benchmark paper .

For wildlife managers, this can mean detecting migration changes, identifying risky conflict zones, monitoring energy expenditure, or understanding when animals feed, rest, flee or reproduce. But as with camera traps, behaviour labels need careful validation.

A Global Map of What Is Happening

This is not a complete list of every project, but it shows the breadth of the field.

Region / domain	Examples	What is happening
Africa	EarthRanger, PAWS, Wildbook, SpeciesNet, MegaDetector, Wildlife Insights, TrailGuard AI, CXL/Wild Me.	Large-mammal monitoring, protected-area operations, endangered species ID, poaching risk, real-time alerts, aerial surveys.
Asia	SMART, PAWS, camera-trap AI, tiger/leopard conflict systems, acoustic and satellite monitoring.	Protected area management, conflict alerts, tiger monitoring, patrol planning, illegal activity detection.
Europe and UK	BirdNET, UK mammal models, Conservation AI / Trap Tracker, bioacoustic monitoring, eDNA surveys.	Open regional models, farmland and woodland monitoring, acoustic bird surveys, restoration tracking.
North America	iNaturalist, BioCLIP, Wildlife Insights, Whale Safe, Project CETI collaborators, Movebank, AI for Good Lab.	Citizen science, foundation models, whale-ship collision risk, animal communication, tracking-data infrastructure.
Latin America	Rainforest Connection / Arbimon, SpeciesNet, Wildlife Insights, AlphaEarth, Global Fishing Watch.	Rainforest soundscapes, camera-trap networks, forest and habitat mapping, illegal fishing transparency.
Oceania	Wildlife.ai, WildObs, drone thermal surveys, SpeciesNet local adaptation.	Marine reserve monitoring, invasive species, rare species ID, thermal drone detection.
Open ocean	Global Fishing Watch, TNC edge monitoring, Whale Safe, DolphinGemma, Project CETI.	Fishing transparency, bycatch monitoring, whale detections, animal communication, ship-strike mitigation.
Planetary habitat layer	AlphaEarth Foundations, Prithvi, Google Earth Engine, Dynamic World, Global Forest Watch.	Habitat mapping, change detection, restoration monitoring, climate and land-use context for species data.

The Emerging Wildlife AI Stack

The future will not be one magical model. It will be a stack of tools that work together.

Layer	Examples	Role
Sensors	Camera traps, microphones, drones, collars, satellites, eDNA samplers, vessel cameras.	Collect observations.
Edge intelligence	SPARROW, TrailGuard AI, WPS, onboard fisheries AI.	Filter and alert in the field.
Foundation models	BioCLIP, NatureLM-audio, AlphaEarth, Prithvi, SAM-style segmentation models.	Provide reusable features across tasks.
Task models	MegaDetector, SpeciesNet, BirdNET, local YOLO models, Wildbook matchers.	Detect, classify, match, count or segment.
Data platforms	Wildlife Insights, Wildbook, Arbimon, EarthRanger, SMART, Movebank, GBIF.	Manage records, metadata, review and sharing.
Human validation	Field rangers, ecologists, taxonomists, local experts, Indigenous and community knowledge holders.	Check predictions and interpret context.
Decision layer	Protected-area plans, patrol routes, restoration priorities, species recovery actions, policy evidence.	Turn observations into conservation action.

The best teams are not asking “which AI model should we use?” They are asking: “What decision are we trying to improve, what data do we need, what uncertainty is acceptable, who reviews the output, and how will it change action on the ground?”

The Hard Parts: What AI Still Cannot Fix

AI makes wildlife monitoring faster, but the hard conservation problems remain social, political, ecological and ethical.

1. Bias in biodiversity data

Biodiversity datasets are deeply uneven. Charismatic mammals, accessible places, wealthy institutions and well-studied regions are overrepresented. Small animals, nocturnal species, insects, cryptic taxa and underfunded regions are underrepresented. Models inherit this imbalance.

2. False confidence

A labelled spreadsheet can look official even when the model is wrong. Field teams need confidence thresholds, review queues, uncertainty flags and validation samples.

3. Sensitive species and location risk

AI can make it easier to find animals. That is good for conservation, but dangerous for species targeted by poachers, collectors or harassment. Sensitive location data should be protected, generalised or embargoed where needed.

4. Data colonialism

Conservation AI should not extract data from local landscapes, train models elsewhere, and return little value to the people protecting those landscapes. Local ownership, benefit sharing and governance matter.

5. Ecological interpretation

AI can count detections. It does not automatically estimate abundance, occupancy, survival, habitat preference or population health. Those require ecological models and assumptions.

6. Energy and infrastructure

Foundation models and large cloud workflows have environmental and financial costs. Edge AI may reduce data transfer, but it introduces hardware, maintenance and repair burdens.

7. The enforcement gap

A model can identify illegal fishing, a poaching hotspot or a deforestation alert. It cannot create political will, fund ranger salaries, repair justice systems or protect whistleblowers.

What Good AI Wildlife Projects Look Like

The strongest projects tend to share the same habits.

They start with a conservation question, not a model demo.
They have field partners from day one, including the people who will use the results.
They keep raw data and metadata organised, because bad metadata ruins good AI.
They validate locally, especially before using outputs for management decisions.
They protect sensitive locations and people.
They report uncertainty, not just confident labels.
They publish reusable tools or datasets where safe, so others can build on them.
They respect local and Indigenous knowledge, rather than treating AI as the only way of knowing a landscape.

Who to Watch

A non-exhaustive watchlist for anyone following the field:

Organisation / project	Why they matter	Link
Microsoft AI for Good Lab	MegaDetector, PyTorch-Wildlife, SPARROW and applied biodiversity tooling.	Tools
Google Research / Google Earth AI	SpeciesNet, Wildlife Insights support, Earth AI, AlphaEarth ecosystem.	SpeciesNet
Conservation X Labs / Wild Me	Wildbook, SCOUT, SA-FARI, Sentinel, FINFINDER and open conservation tools.	Wild Me Lab
Wildlife Insights	Camera-trap data infrastructure and AI-assisted image management.	Website
Cornell Lab / BirdNET	AI bird sound identification and bioacoustic public engagement.	BirdNET
Rainforest Connection / Arbimon	Ecoacoustic monitoring and analysis platform.	Arbimon
Earth Species Project	Foundation models for animal sounds and communication research.	Website
Project CETI	Machine learning, robotics and field biology for sperm whale communication.	Website
Global Fishing Watch	Machine learning and satellite data for ocean transparency.	Technology
EarthRanger	Real-time protected-area operations and ecosystem monitoring.	Website
SMART Conservation Tools	Open protected-area monitoring and reporting suite used by ranger teams.	Website
Imageomics / BioCLIP	Biological vision foundation models for the tree of life.	BioCLIP
NASA / IBM Prithvi	Open geospatial foundation models for Earth observation.	Models
Google DeepMind AlphaEarth	Satellite embeddings and geospatial foundation modelling for global mapping.	Announcement
WILDLABS	Community hub where field conservationists, engineers and researchers exchange tools and lessons.	Website

Where the Field Is Going

The next few years are likely to be shaped by six changes.

1. More multimodal monitoring

Camera traps, microphones, drones, collars, eDNA and satellites will be analysed together. A species record will not only be a photo; it may be a linked bundle of image, audio, movement, habitat and genetic evidence.

2. More local adaptation

Global models will remain useful starting points, but local fine-tuning, active learning and review tools will become standard for serious projects.

3. More edge systems

Remote conservation sites need low-power, low-bandwidth systems that make decisions locally. Edge AI will grow because connectivity is still a major constraint.

4. More foundation models

The field is moving toward reusable representations: visual embeddings for organisms, audio embeddings for soundscapes, geospatial embeddings for habitat, and multimodal embeddings that connect them.

5. More focus on verification

As AI outputs become easier to generate, audit trails will matter more. Conservation decisions need defensible methods, not black-box magic.

6. More pressure to make AI equitable

The people closest to wildlife need access to the tools, not only the institutions with cloud budgets. Open models, offline workflows, shared standards and training will be central.

The Takeaway

The state of AI in wildlife is hopeful, practical and still messy. The best tools already save enormous amounts of human time. They make hidden animals visible, quiet forests measurable, ocean activity traceable and long-term individuals recognisable. But the best teams are not using AI as a replacement for ecology. They are using it as a multiplier for field knowledge.

The future of wildlife AI is not one giant model that “solves conservation.” It is a global network of careful people using better sensors, open tools, foundation models, local validation and ethical data governance to understand life on Earth before more of it disappears.

Selected Sources and Further Reading

Technology in the Field: A Practical Guide to the Tools, Platforms, and Infrastructure That Power Modern Conservation

Sun, 14 Jun 2026 00:00:00 GMT

import BlogImage from "../../components/blog/BlogImage.astro"; import fieldRangerTech from "../../assets/blog/technology-in-the-field/4959212_magda_ehlers.jpg"; import gpsCollar from "../../assets/blog/technology-in-the-field/17242812_peoplebyowen.jpg";

Conservation field teams face a problem that most technology companies never think about. How do you run a real-time operations centre when the nearest mobile tower is 80 kilometres away and the nearest mains power is further? How do you collect standardised survey data when the temperature is 38°C, the sun is blinding the screen, and the baboons have just made off with your lunch? How do you track an elephant across a transboundary landscape when the collar battery dies after 18 months and the replacement costs more than a ranger's annual salary?

These are not hypotheticals. They are the daily reality for conservation teams working in protected areas, wildlife reserves, community conservancies, and research stations across every continent. And the answer is not one magic device. It is a layered stack of tools, each solving part of the problem, wired together with field-tested workflows and a healthy tolerance for failure.

This article maps the technology stack that modern conservation runs on. It covers the hardware, the software, the connectivity layer, and the operations platforms that turn raw field data into decisions. It is written for conservation practitioners, field operations managers, reserve owners, wildlife researchers, ranger units, and students building technology for field deployment.

Source note: This post draws on EarthRanger documentation, SMART Conservation Software resources, the SERCA alliance, WILDLABS community knowledge, published peer-reviewed literature on conservation technology, LoRa Alliance specifications, GSMA connectivity data, and field reports from African Parks, Panthera, WCS, and The Field Company's own field testing in the Cederberg Wilderness Area. Source links are listed at the end.

The Conservation Technology Stack

Think of field technology as five layers, each one depending on the layer beneath it. The layers are not rigid — a single device might span two layers — but the mental model helps when planning a deployment or diagnosing a failure.

Layer	What It Does	Examples
Power	Keeps everything running	Solar panels, battery banks, vehicle alternators, power management
Connectivity	Moves data between devices and people	VHF/UHF radio, LoRa, Meshtastic, satellite, Starlink, cellular
Sensors	Collect observations from the environment	Camera traps, GPS collars, acoustic recorders, weather stations, eDNA
Field Software	Turns observations into structured data	EarthRanger Mobile, SMART, CyberTracker, ODK, QField
Operations Platform	Aggregates data, visualises operations, supports decisions	EarthRanger, SMART Desktop, ArcGIS, custom dashboards

The layers are interdependent. A satellite collar is useless without a connectivity plan. A camera trap is useless without someone to retrieve the SD card. A field app is useless if the phone battery dies by 11:00. Successful field deployments treat the stack as a system, not a shopping list.

The Shift: From Paper to Platform

For most of conservation history, the technology in a ranger's hands was a notebook, a pencil, and a handheld GPS if the budget allowed. Data collection was slow, error-prone, and difficult to aggregate. A poaching incident recorded in a paper logbook might reach headquarters weeks later, if at all. A wildlife sighting noted on a map might never be digitised. Camera trap images sat on SD cards in desk drawers for months before anyone reviewed them.

That world is changing fast. The drivers are not subtle: smartphones got cheap enough to deploy in the field. Satellite connectivity got reliable enough to trust. Open-source platforms like SMART and EarthRanger made operations software free for conservation use. Machine learning models like MegaDetector and SpeciesNet made it possible to process millions of camera trap images automatically. And the conservation community built the training networks and field support systems that make the technology actually usable in practice.

The shift is not complete. Many teams still rely on paper. Many protected areas still lack connectivity. Many rangers still carry consumer-grade phones that overheat and run out of battery. But the direction is clear: conservation technology is moving from something a few well-funded projects use to something every protected area can access. The SERCA alliance — the merger of SMART and EarthRanger, announced in 2025 with backing from WWF, WCS, Panthera, Frankfurt Zoological Society, ZSL, North Carolina Zoo, Wildlife Protection Solutions, and Re:wild — is perhaps the clearest signal that this is now a mainstream movement, not an edge case.

Magda Ehlers on Pexels`} />

Layer 1: Power

Everything else fails without power. A field deployment plan that does not start with power is a plan to fail.

The most common power sources in conservation field work are solar panels (ranging from 5 W portable panels to multi-kilowatt fixed arrays at operations centres), sealed lead-acid or lithium-iron-phosphate batteries, and vehicle alternators for mobile teams. Solar is the default for stationary equipment — camera traps, LoRa gateways, weather stations, repeater sites — because it requires no fuel and minimal maintenance. Good solar deployments use panels sized for the worst month of the year, not the average, and battery banks sized for at least three days of zero sun.

For mobile teams, power is harder. A ranger on patrol carries at minimum a phone, a radio, and often a GPS unit or satellite messenger. Each needs charging. The standard field kit includes a power bank (20,000 mAh or larger) and a small solar panel. Vehicle-based teams can charge from the vehicle, but this introduces a dependency: the vehicle must be running and fuelled. Generator power is common at remote camps but introduces fuel logistics and maintenance overhead.

The rule of thumb: plan for every device to run for 72 hours without a recharge opportunity. Assume batteries degrade in heat. Assume solar panels get dusty. Assume someone forgets to plug something in overnight.

Layer 3: Sensors

Sensors are the eyes and ears of the conservation technology stack. They collect data from the environment that humans cannot observe directly or continuously.

Camera traps are the most widely deployed conservation sensor. Estimates suggest over 10 million camera trap images are now collected globally each year across research, monitoring, and anti-poaching programmes. Modern camera traps run for months on AA batteries, trigger on heat and motion, and store thousands of images on SD cards. Cellular camera traps add real-time transmission but at higher cost and power consumption. The bottleneck is no longer image capture — it is image processing and data management, which AI models like MegaDetector and SpeciesNet are now addressing.

GPS collars and tags track animal movement across landscapes. As of 2026, EarthRanger alone tracks over{" "} 23,000 animals via GPS. Collars range from simple VHF-only models that require manual triangulation to satellite-linked units that transmit positions via Iridium or Argos. Costs span from a few hundred dollars for a VHF collar to several thousand for a satellite collar with accelerometer. Collar weight must stay below 3—5% of the animal's body weight, limiting what species can be tracked.

Acoustic recorders — both terrestrial and underwater — are growing fast. BirdNET and Arbimon process terabytes of acoustic data for species identification. Passive acoustic monitoring (PAM) for marine mammals is now standard practice in offshore development impact assessments.

Environmental DNA (eDNA) samplers collect water, soil, or air samples that are later analysed in a lab to identify species present in an area. This is still more research than operations, but the cost per sample is dropping rapidly.

PeopleByOwen on Pexels`} />

Layer 5: Operations Platforms

The operations platform is the top of the stack. It aggregates data from all layers below — GPS positions, camera trap detections, ranger sightings, acoustic alerts, weather data — and presents it on a map that the operations team can use to make decisions in real time.

EarthRanger is the dominant platform in this space. Free for conservation use, it is deployed at over{" "} 900 sites in more than 80 countries. It tracks wildlife, vehicles, aircraft, and personnel in real time. It generates alerts on movement patterns — immobility, speed changes, boundary crossings — and sends them via SMS, WhatsApp, or email. It integrates with over 150 third-party tools including camera traps, animal collars, vehicle trackers, IoT sensors, radios, and satellite imagery. EarthRanger Mobile (now part of the Ecoscope app) enables offline field data collection with automatic sync when connectivity is restored.

SMART (Spatial Monitoring and Reporting Tool) is the other major open-source platform. Developed by a consortium of eight conservation organisations, it is used in over{" "} 1,000 sites globally. SMART is particularly strong in law enforcement monitoring — tracking patrol routes, recording incidents, and generating reports for protected area managers. Its strength is structured data collection with configurable data models.

In 2025, EarthRanger and SMART merged under the SERCA{" "} (SMART-EarthRanger Conservation Alliance) umbrella, creating a unified technology and training organisation backed by WWF, WCS, Panthera, Frankfurt Zoological Society, ZSL, North Carolina Zoo, Wildlife Protection Solutions, and Re:wild. The goal is to close the gap between field data collection and real-time operations — and to provide the training and support that makes the technology usable for protected area teams worldwide.

Other platforms include ArcGIS (commercial, dominant in government agencies and large NGOs), QGIS (open-source desktop GIS), and Google Earth Engine (cloud-based satellite imagery analysis). Most conservation operations use a combination: EarthRanger or SMART for daily operations, ArcGIS or QGIS for spatial analysis, and Earth Engine for regional-scale remote sensing.

What Actually Happens in the Field

The gap between a technology specification sheet and its performance in the field is wide. Devices that work perfectly on a test bench at headquarters can fail in the first week of deployment. The most common failure modes are not technical — they are environmental and human.

Failure Mode	What Happens	Mitigation
Heat	Phones shut down. Batteries degrade faster. Solar panels lose efficiency.	Shade devices. Use ruggedised hardware. Oversize batteries.
Dust and sand	Charging ports clog. Connectors corrode. Screens become unreadable.	Port covers. Wireless charging where possible. Compressed air cleaning.
Water	Short circuits. Fungus on optics. Paper records destroyed.	IP67 or better rating. Dry bags. Silica gel in equipment cases.
Connectivity loss	Data cannot sync. Alerts do not fire. Teams cannot coordinate.	Offline-first apps. Store-and-forward protocols. Redundant comms layers.
Battery drain	Device dies mid-patrol. GPS logging stops. Emergency comms lost.	Power banks. Solar chargers. Battery discipline training.
User error	Wrong data model selected. Device not charged. Antenna not extended.	Training. Checklists. Simplified interfaces. In-field support.

The most reliable field deployments are not the ones with the most expensive equipment. They are the ones where the team has tested everything in real conditions, built redundancy into every layer, trained every user, and accepted that things will fail — and planned accordingly.

The Open-Source Advantage

Conservation technology has an unusual characteristic compared to most technology sectors: many of the most important tools are open-source and free for conservation use. This is not accidental. It is the result of deliberate decisions by conservation organisations, technology companies, and philanthropic foundations to ensure that cost is not the barrier to effective protection.

EarthRanger is free for qualifying conservation organisations. SMART is open-source and freely available. CyberTracker, one of the oldest field data collection tools (created in 1996 by Louis Liebenberg and Justin Steventon), is free. ODK (Open Data Kit) is open-source and used by conservation organisations worldwide for custom survey data collection. QField, the mobile companion to QGIS, is open-source and free. MegaDetector, the machine learning model for camera trap image processing, is open-source. PyTorch-Wildlife, Microsoft's conservation AI framework, is open-source.

The open-source model has practical advantages for conservation. Teams can inspect the code to understand exactly what it does. They can modify it for local conditions — adding a language translation, a new data field, a custom report. They are not locked into vendor contracts or subject to licence fee increases. And the community of users and developers provides peer support that is often faster and more relevant than commercial support channels.

The trade-off is that open-source tools require local capacity to deploy and maintain. A team that downloads SMART or EarthRanger still needs someone who can configure the data model, set up the server, train the users, and troubleshoot problems. The SERCA alliance is explicitly addressing this gap with a combined training and support organisation.

The Training Gap

The most under-discussed challenge in conservation technology is not hardware or software — it is training. A 2023 survey by WILDLABS found that lack of technical capacity was cited as a barrier to technology adoption more often than cost. Protected area budgets often allocate money for equipment but not for the training, support, and maintenance that make the equipment useful.

The pattern is common: a grant buys camera traps, tablets, and GPS units. The equipment arrives. Someone installs the apps. The rangers go out. Six months later, half the tablets have cracked screens. The camera trap SD cards are full and no one has offloaded the images. The GPS units have dead batteries. The project report says "technology deployment was less successful than anticipated" and the cycle repeats with the next grant.

The organisations that avoid this trap invest in training before equipment. They bring rangers into the process early — testing devices, giving feedback on the data collection interface, flagging problems before the full deployment. They budget for replacement batteries, spare cables, screen protectors, and annual refresher training. They designate a "tech champion" on the team — someone who can answer questions, troubleshoot common problems, and serve as the bridge between the field team and the IT support.

What Is Coming

Several trends are converging that will change what field technology looks like in the next five to ten years.

AI at the edge. Machine learning models that run on the device — not in the cloud — are becoming practical for field deployment. A phone or a camera trap that can classify species, detect gunshots, or identify a vehicle without any internet connection changes what is possible in disconnected environments. PyTorch-Wildlife, Google's SpeciesNet, and TinyML frameworks are making this accessible.

Satellite connectivity at scale. Starlink, Project Kuiper, and the next generation of low-Earth-orbit satellites are bringing broadband to remote areas. Direct-to-device satellite services (AST SpaceMobile, Starlink Direct to Cell) promise to connect standard smartphones to satellites without any additional hardware. If these services deliver, the connectivity layer of the field technology stack becomes dramatically simpler.

Unified platforms. The SERCA alliance between SMART and EarthRanger signals a consolidation trend. Instead of every protected area running its own cobbled-together stack of tools, the sector is moving toward shared, open-source platforms with professional support organisations behind them. This reduces duplication, improves interoperability, and makes training more transferable between sites.

Cheaper, better sensors. The cost of camera traps, GPS collars, acoustic recorders, and environmental sensors continues to fall. A camera trap that cost $400 in 2015 costs under $100 today with better image quality and longer battery life. eDNA analysis that cost hundreds of dollars per sample five years ago is approaching tens of dollars. As sensor costs fall, the density and coverage of monitoring networks increases.

The limiting factor, as always, will not be the technology. It will be the capacity to deploy it, maintain it, train people to use it, and act on the data it produces. Technology is a tool. The people in the field — rangers, researchers, community scouts, protected area managers — are the ones who make it work.

Deep Dives

This post is an overview. The companion articles below go deeper into specific layers of the field technology stack:

Sources

Research drawn from EarthRanger platform documentation, SMART Conservation Software resources, the SERCA alliance, WILDLABS community knowledge, and The Field Company's own field testing in the Cederberg Wilderness Area, 2026.

The Field Co

Open-Source Conservation Technology

How to Process Wildlife Images: A Practical Guide for Junior Field Rangers

Sun, 14 Jun 2026 00:00:00 GMT

import { Image } from "astro:assets"; import BlogImage from "../../components/blog/BlogImage.astro"; import leopardsSavanna from "../../assets/blog/wildlife-image-processing-guide/27288744_frans_van_heerden.jpg"; import leopardNight from "../../assets/blog/wildlife-image-processing-guide/6465808_gill_heward.jpg";

Wildlife images are evidence. A photo of a leopard crossing a road, a hyena at a waterhole, a vehicle on a boundary fence, or an empty night frame is not just a picture. It is a record of what happened at a place and time.

Your job as a field ranger is to help turn those pictures into information that other people can trust. That means protecting the original files, keeping good notes, checking dates and camera locations, sorting out empty images, identifying animals carefully, and reporting results in a way that is useful for patrol planning, ecological monitoring, and conservation decisions.

This guide is written for a junior field ranger. It does not assume you are a data scientist or programmer. The goal is to give you a safe, repeatable workflow you can follow every time you collect camera-trap images, ranger patrol photos, or other wildlife images from the field.

Source note: This guide was prepared from current camera-trap data guidance from GBIF, Wildlife Insights, Camtrap DP, Microsoft MegaDetector and PyTorch-Wildlife documentation, and sensitive species data guidance accessed on 14 June 2026. Local reserve protocols, national law, permit conditions, and instructions from your ecologist or conservation manager always take priority.

What “Processing Wildlife Images” Means

Processing wildlife images means taking raw photos from a camera or phone and turning them into clean records. A clean record should answer four simple questions:

Question	Example answer	Why it matters
What was seen?	Leopard, human, vehicle, elephant, bird, blank image	This tells the team what activity happened.
Where was it seen?	Camera station KNP-North-014 or waterhole W03	This helps map animal movement, patrol risk, and habitat use.
When was it seen?	2026-06-14 at 21:34 local time	This helps understand activity times and match events to patrols or other cameras.
How sure are we?	Confirmed by ranger, uncertain, AI suggestion only	This stops weak identifications from being treated as facts.

Good image processing is not about speed only. Speed helps, especially when a camera produces thousands of images, but accuracy and traceability matter more. Someone should be able to look at your record later and understand where the file came from, who reviewed it, what was identified, and whether anything still needs expert checking.

The Five Golden Rules

Keep these rules in mind before you touch any image files:

Do not edit or delete original files. Keep the raw images exactly as they came from the camera.
Copy first, work later. Always copy files to the project storage before sorting or reviewing.
Keep images grouped by deployment. A deployment means one camera at one place for one time period.
Record uncertainty honestly. “Unknown antelope” is better than a wrong species name.
Protect sensitive information. People, vehicles, camera locations, rhino sightings, pangolin records, nests, dens, and rare species may need restricted access.

These rules are simple, but they prevent most serious data problems: lost files, mixed-up camera stations, wrong dates, accidental exposure of sensitive locations, and false species records.

The Whole Workflow in One View

Wildlife image processing follows the same path every time:

Step	What you do	Main output
1. Collect	Bring in memory cards or field photos with deployment notes.	Raw images plus field sheet.
2. Copy	Copy files to the correct project folder.	Safe working copy.
3. Backup	Make a second copy before reviewing.	Protected original dataset.
4. Organise	Group files by project, camera, location, and deployment.	Clean folder structure.
5. Check metadata	Check date, time, camera ID, and location notes.	Reliable deployment record.
6. Run AI or first sort	Separate animals, people, vehicles, and blanks.	Shorter review queue.
7. Human review	Confirm species, count, behaviour, and uncertainty.	Verified observations.
8. Quality check	Look for mistakes, missing files, duplicates, and sensitive records.	Clean dataset.
9. Export/report	Send summary and cleaned data to the senior ranger, ecologist, or data manager.	Report, spreadsheet, dashboard, or archive.

Step 1: Collect Images and Field Notes Properly

Good processing starts in the field. If the camera card arrives without proper notes, the images may be difficult or impossible to use later.

For every camera check, record at least:

Project name: the monitoring project or reserve programme.
Camera ID: the unique code written on the camera body.
Location ID: the code for the camera station or site.
Date and time collected: when the card was removed.
Deployment start and end: when the camera was active.
GPS or site reference: use the approved reserve system, not a public location description for sensitive sites.
Camera condition: working, damaged, stolen, low battery, full card, lens blocked, moved by animal, wrong time, wrong angle.
Card number: useful if several cards are collected in one patrol.
Ranger name or team: who collected or checked the unit.

Do not rely on memory. At the end of a long patrol, several cameras and cards can look the same. Write the information down at the camera site or immediately when you return to the vehicle.

Step 2: Copy and Backup Before You Review

When you return from the field, treat every memory card like evidence. The first job is to protect the files.

Put the memory card into the computer or card reader.
Open the project storage folder.
Create the correct folder for that deployment.
Copy all images and videos from the card into that folder.
Check that the number of files copied matches what is on the card.
Make a second copy to backup storage if your team has one.
Only format or reuse the memory card after your supervisor confirms the copy and backup are safe.

A good folder structure is simple and predictable. For example:

        
          project_name/ raw/ camera_id/ location_id/
          deployment_2026-06-01_to_2026-06-14/ DCIM/ IMG_0001.JPG IMG_0002.JPG
          processed/ exports/ reports/

Keep the original images in the raw folder. If you crop, rotate, rename, resize, or annotate anything, save those edited files somewhere else, such as processed. This keeps the original evidence safe.

Step 3: Organise Files Without Damaging the Raw Data

Most camera-trap systems already give files names like{" "} IMG_0001.JPG or PICT0001.JPG. Do not rename raw files unless your project manager has a clear rule for doing it. It is safer to keep original file names and organise by folder.

Use one folder per deployment. Do not mix two cameras in one folder. Do not mix two time periods in one folder. Do not place files from an unknown camera into a known camera folder just because you think it is probably correct.

Good habit	Bad habit
Keep one folder for one camera deployment.	Dumping all cards from the patrol into one folder.
Keep original camera file names.	Renaming files with species names before review.
Use folder names that sort by date.	Using names like “new photos”, “more pics”, or “camera stuff”.
Use simple characters: letters, numbers, hyphens, underscores.	Using spaces, apostrophes, slashes, emojis, or long notes in file names.

If your team uses a formal data platform, follow that platform’s upload rules. If your team uses folders and spreadsheets, keep them tidy enough that another ranger can understand them without phoning you.

Step 4: Check the Important Metadata

Metadata means information about the image. For wildlife monitoring, the most important metadata are usually date, time, camera, place, and deployment details.

Check these before identifying species:

Camera time: is the camera clock correct? If it is wrong, record the correction.
Time zone: does your project use local time or UTC? Use one system consistently.
Camera ID: does the folder match the camera body or field sheet?
Location ID: does the deployment match the correct site?
Start and end date: do the images fall inside the deployment period?
Camera problems: lens blocked, camera tilted, false triggers, low battery, water damage, wrong height, or wrong direction.

If you find a problem, do not hide it. Add a note. A dataset with honest warnings is more useful than one that looks clean but contains hidden errors.

Step 5: Do a First Pass Sort

The first pass is not final species identification. It is a rough sort that helps you reduce the workload.

Sort images into basic categories:

Category	Meaning	What to do
Blank	No animal, person, or vehicle visible.	Keep the record, but do not spend much time on it unless investigating camera faults.
Animal	Wildlife or livestock visible.	Send to species review.
Human	Person visible.	Restrict access and follow privacy/security protocol.
Vehicle	Vehicle, motorbike, bicycle, or aircraft visible.	Restrict access if it relates to security or patrol investigations.
Unknown	Something is visible but unclear.	Flag for review by a senior ranger or ecologist.

If your team uses an AI tool, this is usually where it helps most. The AI can remove many blank images and find images with animals, people, or vehicles. You still need a human to review important records.

Step 6: Use AI as an Assistant, Not as the Final Authority

AI tools can save a lot of time, but they are not magic. A model may miss small animals, distant animals, animals partly hidden behind grass, night images, unusual camera angles, local species it has not seen before, or confusing backgrounds.

A common workflow is:

Run a detector to find animals, people, vehicles, and blanks.
Review high-risk or uncertain results first.
Use a species classifier only if your project has one that is suitable for your region.
Confirm important sightings manually.
Record whether the classification was made by AI, a ranger, an ecologist, or a combination.

MegaDetector is an example of a widely used open-source camera-trap detector. It detects animals, people, and vehicles, but it is not a species classifier. PyTorch-Wildlife is the broader framework that can load MegaDetector and other conservation AI models. Platforms such as Wildlife Insights also support camera-trap data management and AI-assisted annotation.

For a junior ranger, the most important rule is this:{" "} AI suggestions must be checked before they become official records. {" "} This is especially important for rare species, conflict species, endangered species, poaching-related evidence, and records that will be used in management decisions.

Gill Heward on Pexels`} />

Step 7: Review Animals Carefully

When reviewing animals, work from easy facts to harder ones.

Is there an animal? If yes, continue. If not, mark blank.
What broad group is it? Mammal, bird, reptile, livestock, unknown animal.
Can you identify the species? Only use species level if you are confident.
How many individuals? Count visible animals, not guessed animals.
Is age or sex obvious? Record only if you can see it clearly.
Is behaviour important? Drinking, feeding, mating, fighting, carrying prey, fence crossing, using road, inspecting camera.
Is it sensitive? Rare species, threatened species, rhino, pangolin, nest/den, carcass, snare, human, vehicle, firearm, or camera tampering.

When you are not sure, choose a higher-level label. For example, use “antelope”, “small carnivore”, “bird”, or “unknown animal” rather than guessing a species. A wrong species can damage an analysis. An honest unknown can still be useful.

Situation	Better label	Why
Only legs visible at night.	Unknown mammal	Not enough evidence for species.
Small antelope partly hidden by grass.	Unknown antelope	Species may be unclear.
Clear elephant herd at waterhole.	African elephant, count visible individuals	Species and count are clear.
Leopard-like spotted cat but blurred.	Possible leopard, needs expert review	Important record should be confirmed.

Frans van Heerden on Pexels`} />

Step 8: Group Images into Events When Needed

Many cameras take several photos of the same animal in one visit. For example, a hyena may trigger ten images in two minutes. Counting those as ten separate hyenas would be wrong.

Your project may ask you to group images into events or{" "} sequences. An event is a set of images that likely show the same visit by the same animal or group. The exact rule depends on your project, but teams often use a time gap such as 30 minutes between independent records.

Always follow your local project rule. If no rule has been given, ask the data manager or ecologist before making independent counts.

Record type	What it means
Image-level record	One row per image.
Event-level record	One row per animal visit or sequence.
Deployment-level summary	One summary for a whole camera deployment.

For patrol and management use, event-level records are often easier to understand than image-level records. For detailed research, the data team may need both.

A Simple Spreadsheet Template

If your team does not use a dedicated platform, a simple spreadsheet can work. Keep the columns consistent. Do not change column names halfway through a project.

Column	Example	Notes
project_id	north_boundary_2026	One project name or code.
camera_id	CAM_014	Must match the camera label.
location_id	NB_014	Use the reserve-approved site code.
deployment_id	CAM_014_20260601_20260614	One camera, one site, one time period.
file_name	IMG_0042.JPG	Original file name.
file_path	raw/CAM_014/NB_014/...	Where the image is stored.
timestamp	2026-06-14T21:34:00+02:00	Use a consistent date-time format.
observation_type	animal	animal, human, vehicle, blank, unknown.
common_name	Leopard	Use accepted reserve names.
scientific_name	Panthera pardus	Use only if your team requires it and you know it.
count	1	Number visible, not guessed.
confidence	confirmed	confirmed, likely, possible, unknown.
classified_by	ranger_jabu	Person, AI model, or both.
classification_method	human_review	human_review, ai_suggestion, ai_plus_human.
sensitive	yes	yes/no. Follow reserve protocol.
notes	Blurred; possible cub nearby	Keep notes short and professional.

This is a practical field template, not a full scientific data standard. If your project will publish data to GBIF or exchange it with other systems, the data manager may convert it to Camtrap DP or Darwin Core later.

Step 9: Protect Sensitive Data

Some images must not be shared widely. This is not about hiding science; it is about protecting animals, people, staff, cameras, and the reserve.

Treat the following as sensitive unless your supervisor says otherwise:

Images of people, staff, community members, tourists, suspects, or patrol activity.
Vehicles, number plates, firearms, snares, carcasses, fence breaks, or signs of illegal activity.
Exact locations of rhino, pangolin, rare orchids, nests, dens, breeding sites, or other vulnerable species.
Exact camera locations, especially active cameras.
Notes that reveal patrol routes, informant information, security weaknesses, or private property details.

Do not post camera-trap images on social media without permission. Do not share GPS coordinates in WhatsApp groups unless the group is approved for that purpose. Do not include exact sensitive locations in public reports. If a record is important but sensitive, mark it clearly and send it through the correct internal channel.

When data are shared outside the reserve, sensitive locations may need to be generalized. This means giving a wider area rather than the exact camera point. The correct level of generalization depends on the species, the threat, and the rules of your organization.

Step 10: Do a Quality Check Before Reporting

A quality check is a final look for mistakes before data are used. Use this checklist before sending your file to a senior ranger, ecologist, data manager, or reserve manager.

Check	Question to ask	Action if wrong
File count	Did all files copy from the card?	Re-copy from card or report missing files.
Folder	Are the files in the correct deployment folder?	Move to correct folder before processing.
Timestamp	Does the camera time look correct?	Record time error and correction.
Species	Are rare or difficult species confirmed?	Flag for expert review.
Blanks	Are blank images really blank?	Spot-check AI blank results.
Duplicates	Are the same files copied twice?	Remove duplicates from processed data, not from raw backup.
Sensitive records	Are people, vehicles, rare species, and security images marked?	Restrict, anonymize, or generalize before sharing.
Names	Are common and scientific names consistent?	Use the reserve species list or ask the ecologist.

Step 11: Report What Matters

Your report should be useful to the people who make decisions. Do not just send a folder of images unless that is what was requested.

A good short report includes:

Project and date range: what was processed.
Number of cameras or deployments: how much effort was covered.
Total media processed: number of images/videos.
Summary by category: animals, humans, vehicles, blanks, unknowns.
Species list: confirmed species detected.
Important sightings: rare species, breeding signs, unusual behaviour, conflict species.
Security flags: humans, vehicles, snares, firearms, fence damage, camera tampering.
Data warnings: wrong camera time, missing cards, failed camera, unclear images, heavy false triggers.
Next actions: expert review needed, patrol follow-up, camera maintenance, redeployment, or data upload.

Keep sensitive details out of general reports. Use a separate restricted report for exact locations, suspected illegal activity, or highly vulnerable species.

Example Summary for a Senior Ranger

This is the kind of short summary a junior ranger could send after processing a batch:

Camera-trap image processing summary — North Boundary, 1–14 June 2026
Processed 4 deployments from cameras CAM_011 to CAM_014. Total media reviewed: 8,420 images. First-pass sorting found 5,930 blanks, 2,210 animal images, 184 human images, 72 vehicle images, and 24 unknown/unclear images. Confirmed species include elephant, kudu, impala, spotted hyena, leopard, porcupine, vervet monkey, and side-striped jackal. One possible pangolin record from CAM_014 at NB_014 needs expert confirmation and has been marked sensitive. CAM_012 time appears to be 2 hours behind local time; correction noted in the spreadsheet. CAM_013 produced many blank triggers due to grass movement and should be trimmed or repositioned. Human and vehicle records were separated into the restricted folder for security review.

Notice that the summary gives numbers, important sightings, problems, and actions. It does not expose sensitive coordinates in the general text.

Common Mistakes to Avoid

Mistake	Why it is a problem	Better practice
Deleting blank images too early.	Blank rate can show camera problems or effort.	Keep raw files; mark blanks in processed data.
Renaming files with species names.	Names can be wrong and break links to records.	Keep species names in the spreadsheet or data platform.
Guessing species when unsure.	Wrong records can mislead conservation decisions.	Use “unknown” or “possible” and flag for review.
Mixing cards from different cameras.	Location and effort become unreliable.	Copy one card at a time into the correct deployment folder.
Ignoring wrong camera time.	Activity patterns and event matching become wrong.	Record the time error and correction.
Sharing rare species images publicly.	Can increase risk to animals and camera sites.	Follow sensitive data rules and get approval.
Trusting AI without checking.	AI can miss animals or misread local species.	Human-review important, uncertain, and sensitive records.

How to Treat AI Confidence Scores

Some AI tools give a confidence score. A high score means the model is more confident. It does not mean the answer is definitely correct.

AI result	What it means	What you should do
High-confidence animal detection	The model is confident there is an animal.	Review for species, count, and sensitivity.
Low-confidence animal detection	The model saw something but is unsure.	Check manually; it may be a small, distant, or hidden animal.
High-confidence blank	The model thinks nothing important is visible.	Spot-check a sample, especially at new sites or with new cameras.
Species classifier suggestion	The model suggests a species.	Accept only after human review, especially for important records.

If you see repeated AI mistakes from the same camera, tell the data manager. The problem may be vegetation, camera angle, infrared glare, dust, water on the lens, or a local species the model does not handle well.

Quick Checklists

Before reusing a memory card

Files copied to project storage.
File count checked.
Backup completed.
Deployment folder confirmed.
Supervisor or project rule allows formatting.

Before submitting processed data

Blanks, animals, humans, vehicles, and unknowns separated or labelled.
Important species confirmed or flagged.
Sensitive records marked.
Wrong camera time noted.
Missing or damaged data reported.
Summary report prepared.

When the Data Team Talks About Standards

You may hear terms like Camtrap DP,{" "} Darwin Core, GBIF, or{" "} FAIR data. You do not need to master these on day one, but it helps to know what they mean.

Term	Plain-language meaning
Camtrap DP	A standard way to package camera-trap data using tables for deployments, media, and observations.
Darwin Core	A widely used biodiversity data standard for sharing species occurrence records.
GBIF	A global network and data platform where biodiversity records can be published and cited.
FAIR data	Data that are findable, accessible, interoperable, and reusable.

Your clean field notes and careful image processing make these standards possible. If the field data are messy, no standard can fix everything later.

Training Exercise for a Junior Ranger

Use this exercise to practise before processing real high-value data.

Take a small folder of 100 mixed images from a training camera.
Create a deployment folder with the correct project, camera, location, and date range.
Copy the images into the raw folder.
Create a spreadsheet using the template in this guide.
Sort the images into blank, animal, human, vehicle, and unknown.
Identify animals only as far as you are confident.
Mark any sensitive images.
Write a five-line report for your supervisor.
Ask a senior ranger or ecologist to review ten random records and all uncertain records.
Discuss any mistakes and update your notes.

The goal is not to be perfect on the first try. The goal is to build a habit of careful, repeatable work.

Final Advice

A good ranger is not just someone who sees wildlife in the field. A good ranger helps turn sightings into reliable knowledge. Every clean folder, corrected timestamp, honest unknown, confirmed species record, and protected sensitive image improves the reserve’s ability to understand and protect wildlife.

Work carefully. Ask when unsure. Protect the originals. Respect sensitive information. Let AI help, but do not let it replace field judgement. Your image records may later support patrol planning, anti-poaching investigations, ecological research, species recovery work, habitat management, and conservation funding.

Sources and Further Reading

Ingwe Research Program — 200 Leopards, One Full-Time Person, and a Model Worth Watching

Sat, 13 Jun 2026 00:00:00 GMT

import { Image } from "astro:assets"; import BlogImage from "../../components/blog/BlogImage.astro"; import IngweWildl2 from "../../assets/blog/ingwe-wildlife-trust/29415794_lucie_burlet.jpg"; import IngweWildl1 from "../../assets/blog/ingwe-wildlife-trust/28612999_chané_timmerman.jpg"; import leopardSighting from "../../assets/blog/ingwe-wildlife-trust/1642068_satria_bagaskara.jpg"; import leopardKruger from "../../assets/blog/ingwe-wildlife-trust/1319504_magda_ehlers.jpg";

Marine Servonnat is the Managing Executive of the Ingwe Research Program. She wrote the Annual Impact Report, designed it, and published it. She manages five active research projects, 247 meetings, and everything else that comes with building a conservation organisation. There is no deputy. There is no admin assistant. Ingwe runs on one full-time salary and a network of reserves, guides, volunteers, and citizen scientists who believe the leopards are worth it.

Despite this — or because of the clarity it forces — Ingwe has individually identified over{" "} 200 leopards across 31 reserves in the Greater Kruger. It has processed 11,000+ unique sighting encounters submitted by 400 citizen scientists. It runs five research projects, launched a K9 scent detection pilot, and published a Framework for Research Ethics approved by its Board. In June 2026, it launched a mobile app built by two volunteers in 14 months.

This is not a story about an app. It is a story about a conservation model that works because it was designed around what field researchers actually need, not what an organisation chart says they should have.

What Ingwe Actually Is

Ingwe Research Program is a nonprofit company registered in South Africa (NPC K2024401955, PBO 930087476), but the research itself has been running for much longer — the organisation traces its lineage back over two decades of leopard monitoring in the Lowveld. It was formally incorporated in 2024 with headquarters in Hoedspruit and a second office in Cape Town.

The name is deliberate. Ingwe is the Zulu word for leopard. Zulu does not have a separate word for "leopardess." Female leopards are called ingwe, same as males — a linguistic nod to the species equality that the research itself reflects: Ingwe studies the population, not just the big males that tourists want to photograph.

The mission, in their words: to effectively conserve leopards in South Africa by empowering stakeholders through practical scientific recommendations, enabled by formal research, citizen science, novel technologies, and proactive collaboration.

What "practical scientific recommendations" means on the ground: a reserve manager gets a home range map showing which leopard uses which portion of their property. A provincial roads agency gets a culvert usage report with photographs of seven individual leopards using specific underpasses, along with mortality data from the 47km stretch of road those culverts sit beneath. A lodge guide gets an identification kit that tells a guest the leopard they just photographed is named Bokamoso, she had a cub last season, and she was last sighted three kilometres east of here in April.

The Five Research Projects

Ingwe runs five active research projects simultaneously. On one salary. Here is what each one does.

1. Citizen Science Leopard Monitoring

The backbone. Guides, rangers, lodge managers, and guests across 31 reserves submit leopard sightings — photos, video, GPS location — that feed into a central population database. Each leopard's unique spot pattern is matched using AI-assisted identification. The dataset contains 11,000+ unique encounters and has identified over 200 individual leopards. The network includes 45+ partner entities and over 400 citizen scientists. This is the largest carnivore-focused citizen science project in South Africa.

2. Road Ecology on the R40

The R40 is a 47km stretch of provincial road running through the heart of leopard country between the Kruger National Park and the Blyde River Canyon. In 84 survey days, Ingwe documented 198 wildlife mortalities across 38 species — including seven individually named leopards. Extrapolated across a full year, the estimate approaches 860 animals lost on this single road. Africa Geographic is producing a documentary about this project, Spots on the Line.

3. Wildlife Use of Culverts

Camera traps positioned inside road culverts reveal which underpasses wildlife actually uses. The project's emblematic image: Bokamoso, a resident female, emerging from a culvert with her cub. She had taught her cub to use it. This data will inform infrastructure decisions — which culverts need maintenance, which need modification, and where new crossing structures should be built.

4. K9 Scent Detection Pilot

A first-of-its-kind project exploring whether trained detection dogs can locate leopard scat for genetic sampling. Partnered with Canines for African Nature. If successful, this provides a non-invasive method for population genetics that does not require collaring or darting.

5. Human-Leopard Coexistence

Led by researcher Eleanor Salisbury in partnership with Transfrontier Africa, this project investigates perceptions and behaviours of communities living alongside leopards, with the goal of developing evidence-based coexistence methods.

Magda Ehlers on Pexels`} /> Satria Bagaskara on Pexels`} /> Chané Timmerman on Pexels`} />

What the Numbers Say

Metric	Value
Individually identified leopards	200+
Unique sighting encounters	11,000+
Partner reserves	31
Partner entities enrolled	45+
Citizen scientists	400+
Wildlife mortalities documented (R40, 84 days)	198
Leopards killed on R40 (documented)	7
Active research projects	5
Full-time staff	1
Newsletter subscribers	600
Months of reserves remaining	7

The last two rows are connected. Seven months of reserves on one salary is not a financial crisis — it is better than most small conservation NGOs manage. But it is not sustainable. The organisation's FY26 Annual Impact Report makes this explicit: the infrastructure is proven, the data is rigorous, the citizen science network is working. What is missing is the funding to hire staff so the person running five projects can focus on research instead of admin.

Lucie Burlet on Pexels`} />

The People Who Make It Work

Marine Servonnat is the Managing Executive and the organisation's sole full-time employee. She manages the research, the partnerships, the communications, the fundraising, the volunteers, and the board. She wrote and designed the Annual Impact Report. She gets up at sunrise and drives the R40.

The Board: Simon Hartley, Tom Lautenbach, and trustees Sophie Gandet, Paul ffolkes Davis, and Mark Dumbleton provide governance, strategic direction, and donor stewardship. They are not figureheads — Simon Hartley publicly describes Marine's work as carrying the programme forward "every single day with such a potent blend of passion and clearheadedness."

The volunteers: Liam Hoffmann, a software engineer who read the newsletter and emailed offering to build an app. Fourteen months later, the Ingwe app launched on Google Play. Jemma Jeffery, a UX designer pivoting into conservation technology, designed the entire interface from scratch. Eleanor Salisbury leads the human dimensions research. Rachael Leeman contributed to the road ecology work. Dozens of guides, rangers, and lodge staff submit sightings daily. Hundreds of guests photograph leopards and submit those images to a database that turns tourism into science.

Why This Model Matters

Most conservation organisations follow a predictable trajectory: a passionate founder starts small, secures a grant, hires staff, builds infrastructure, and gradually becomes an institution. The model works when the funding holds. When it does not — when the grant ends and the staff scatter — the data, the relationships, and the institutional knowledge often scatter with them.

Ingwe has inverted this. The research network — 31 reserves, 400 citizen scientists, 11,000 encounters — exists independently of the organisation's staffing level. The guides will keep submitting sightings whether Marine has an assistant or not. The leopards will keep walking through culverts whether the K9 pilot gets funded or not. The institutional value is in the dataset and the network, not in the org chart.

This is a genuinely different structure. Ingwe is thin at the centre and thick at the edges. The centre is one full-time researcher with a board. The edges are 400 people holding phones across an area the size of a small country. The technology — the app, the AI identification, the database — is what connects the edges to the centre. It is the infrastructure, not the product.

The Field Company exists to build tools for organisations like this. The thesis: the labour, the expertise, and the commitment are already out there. What has been missing is the infrastructure that lets the effort at the edges flow into structured, analysable, actionable data at the centre without requiring the centre to hire a developer.

Ingwe proved the thesis is correct. One full-time person. 200 leopards. Five research projects. Two volunteers built the app in 14 months because the organisation had a clear problem, a working network, and enough clarity that a software engineer could look at the situation and understand exactly what needed to be built.

The model is replicable. Not the specifics — leopards in the Greater Kruger — but the structure: a thin professional core, a thick network of contributors, technology that connects them, and the discipline to run on a shoestring until the data proves the case for more funding. That structure works for any species, any region, any research question where the observers are already out there and the infrastructure is the missing piece.

How to Support Ingwe

Ingwe is in an active fundraising phase. They have seven months of reserves and a proven track record. Here is what moves the needle:

Read the Annual Impact Report FY26. It is available on their website at ingweresearchprogram.org. It is the clearest window into what this organisation actually does, and it makes the case for professional conservation infrastructure better than any summary can.

Subscribe to the newsletter. 600 people already do. It goes out monthly via Substack and it is consistently the best-written conservation newsletter we have read — direct, honest, specific.

Donate. The GoFundMe is at{" "} gofundme.com/f/ingwe-research-program. Donations go directly into field operations — equipment, vehicle costs, data processing, the unglamorous infrastructure that produces the science.

Share. If your network includes conservation donors, grant-makers, or CSR decision-makers who fund wildlife research in Southern Africa, forward them the Impact Report. Ingwe needs sponsors more than it needs signal boosts, but signal boosts find sponsors.

If you work in the Greater Kruger. If your reserve, lodge, or organisation operates in the area and you want to understand how the citizen science network works, contact admin@ingweresearchprogram.org. The network grows by adding reserves, not by centralising everything into one office.

If you are a developer. Read our post on{" "} how to contribute to open source conservation technology. Liam Hoffmann did it in 14 months. You can too.

Ingwe Research Program NPC — NPO K2024401955, PBO 930087476. ingweresearchprogram.org . Annual Impact Report FY26 available on their website. Newsletter:{" "} ingweresearchprogram.substack.com . Donate:{" "} GoFundMe. Documentary: Spots on the Line by Africa Geographic. All figures from Ingwe's FY26 Annual Impact Report and June 2026 LinkedIn update. Leopards photographed by citizen scientists across 31 reserves in the Greater Kruger.

The Field Co

Open-Source Conservation Technology

What to Use When You Find Something in the Field

Mon, 01 Jun 2026 00:00:00 GMT

import cybertrackerLogo from "../../assets/blog/cybertracker_logo.svg"; import odkLogo from "../../assets/blog/odk.svg"; import smartLogo from "../../assets/blog/SMART-squarelogo-white.png"; import surveyLogo from "../../assets/blog/survey123.png"; import fulcrumLogo from "../../assets/blog/fulcrum-logo.svg"; import inatLogo from "../../assets/blog/inaturlist.svg"; import fieldlogLogo from "../../assets/images/fieldlog-logo.png"; import phoneField from "../../assets/blog/phone-use-outside.jpg"; import clipboardField from "../../assets/blog/clipboard-checklist.jpg"; import ipadField from "../../assets/blog/ipad-checklist.jpg";

You spend three weeks in the bush. You see a pangolin. You scribble in a notebook. You come home, transcribe the notes into a spreadsheet, manually geotag the photos, and email the file to someone who opens it six weeks later and asks: "What date was this?"

This is how most wildlife data still moves — through notebooks, memory, Excel, and hope. The tools to fix this exist. They are not new. But field scientists, rangers, and conservationists face a bewildering landscape of apps, each claiming to solve the problem. Most of them were not built for the people actually doing the fieldwork.

Here is every major option, compared honestly. Each section names the platform, links to it, and gives you the unvarnished truth about what it does well and where it falls short.

CyberTracker {" "} is the original. Developed in 1996 by Louis Liebenberg and Justin Steventon to help illiterate San trackers in the Kalahari record wildlife observations using icons instead of text. It proved that field data collection can work for anyone, anywhere — and became the template for an entire category of software.

What it does well: Icon-based interface eliminates literacy barriers. Extremely efficient for rapid species logging while tracking on foot. Works offline. Free for conservation use. Deeply respected in the African conservation community. The CyberTracker protocols for tracking and observation are a methodological standard.

Where it shows its age: The desktop version requires Windows and looks like it. Mobile app interface has not been significantly modernized in a decade. Data export workflows are clunky — you can get data out, but it takes work. No real-time team sync. No native cloud dashboard. Limited to pre-defined species lists; customizing the observation form requires digging through the sequence editor, which is powerful but arcane. Documentation is scattered.

Best for: Lone trackers and small teams in Southern Africa who need an icon-based, low-battery, offline-first logger. CyberTracker's spiritual DNA runs through every app that followed, but in 2026 it feels like working inside a well-preserved fossil.

ODK {" "} (Open Data Kit) is the standard for form-based field data collection in global health, development, and environmental research. Over 2 million users send 250 million submissions annually through ODK Collect.

What it does well: Form builder (XLSForm) is the industry standard — define a form in a spreadsheet, upload it, deploy it. Supports skip logic, calculations, multiple languages, GPS, photos, and barcode scanning. Works fully offline with automatic sync. Open-source and auditable. Massive community — 17,000 forum members, comprehensive docs. Free if you self-host (requires technical skill). ODK Cloud starts at $199/month for 10K monthly submissions. API access for integrating with R, Python, Power BI.

Where it falls short: ODK is a form engine, not a wildlife tool. There are no built-in species databases, no tracking protocols, no field-guide integration. You build everything from scratch. The mobile interface is functional but utilitarian — optimized for enumerators administering surveys, not for a ranger in the rain with one hand free. Entities (for linking observations to individual animals) only arrived recently and require the Professional tier ($499/month). No team coordination features — each device is an island until sync. Analysing data requires external tools.

Best for: Large-scale structured surveys where you need a bulletproof form engine with rigorous data validation. WHO uses it. The Red Cross uses it. It is excellent at what it does — but what it does is forms, not fieldwork.

SMART {" "} (Spatial Monitoring and Reporting Tool) is the dominant platform for conservation law enforcement. Developed by a coalition of WWF, WCS, ZSL, Panthera, and others, it is purpose-built for ranger patrols, anti-poaching operations, and protected area management.

What it does well: End-to-end patrol management — plan patrols, collect observations, analyze threats, generate reports. Purpose-built for the use case: illegal activity logging, snare detection, carcass reporting, arrest documentation. Extensive training materials and a global community of practice. Desktop application for analysis, querying, and mapping. Free and open-source. Now aligned with EarthRanger through the SERCA alliance for real-time situational awareness.

Where it falls short: Heavy. SMART is a full desktop application plus mobile app plus server component. Setup and configuration require training — this is not something you install and start using today. The interface is conservation-law-enforcement-first; general wildlife observation feels like an afterthought. Data entry is form-heavy and structured around incident reporting protocols. Mobile app performance on older devices can be slow. Not designed for informal observation logging or ad-hoc field notes. The analysis output is powerful but prescriptive — it tells you patrol effort and threat distribution, not "what birds did we see this week?"

Best for: Protected area authorities running formal ranger patrols with defined patrol sectors, threat monitoring, and law enforcement mandates. If you manage a national park with armed rangers and need court-admissible evidence trails, SMART is the tool.

ArcGIS Survey123 {" "} is Esri's form-centric field data collection app, deeply integrated with the ArcGIS ecosystem. Widely used in government, utilities, and environmental consulting.

What it does well: Beautiful form designer with drag-and-drop interface. Deep ArcGIS integration — survey responses appear as feature layers on your maps instantly. Smart forms with conditional logic, repeats, calculations, and rich media capture. Works offline with automatic sync. Enterprise-grade security, single sign-on, and compliance. Survey123 Connect (desktop) for advanced XLSForm-based authoring. Native integration with ArcGIS Dashboards for real-time monitoring.

Where it falls short: ArcGIS dependency is absolute — if your organization is not already an Esri shop, the licensing cost is prohibitive (ArcGIS Online subscription required). The free tier is severely limited. The form-first paradigm means every observation is a survey response, which becomes awkward for continuous tracking or rapid-fire species logging. No built-in wildlife-specific features — no taxonomies, no protocols, no field guides. Customization beyond forms requires ArcGIS developer skills. The mobile app is polished but heavy; startup time can be slow on older devices. Data ownership lives inside the Esri ecosystem.

Best for: Organizations already invested in ArcGIS who need to add field data collection to their existing GIS workflows. Environmental consultancies doing structured site assessments. Not designed for the rhythm of wildlife tracking.

Fulcrum {" "} is a commercial field data collection platform used by nearly 3,000 companies, recently acquired Wildnote for environmental compliance. Strong in utilities, engineering, and environmental consulting.

What it does well: Polished mobile and web experience. Drag-and-drop app builder with no coding required. AI-powered FastFill for voice dictation and photo-based data entry. Strong GIS integration with Esri feature layers. Real-time dashboards and reporting. SOC 2 compliant, SCIM provisioning. Developer API for custom integrations.

Where it falls short: Built for industrial field operations — pole inspections, site assessments, compliance audits. The workflow is optimized for structured inspections at known locations, not for opportunistic wildlife observation across unknown terrain. Pricing starts at enterprise tiers (contact sales). No wildlife-specific features. The AI features are impressive but tuned for infrastructure classification, not species identification. No offline-first team coordination — devices operate independently during collection. The platform is powerful but the use case mismatch with conservation fieldwork is significant.

Best for: Engineering firms, utilities, and environmental consultants doing structured inspections at known sites. If you need to inspect 10,000 utility poles with AI-assisted photo classification, Fulcrum excels. If you need to log animal sightings while tracking through the bush, it feels like driving a semi-truck down a footpath.

iNaturalist {" "} is the world's largest biodiversity observation platform. A joint initiative of the California Academy of Sciences and the National Geographic Society, with over 200 million observations from citizen scientists worldwide.

What it does well: Beautiful, intuitive mobile app. Computer vision species identification — point your camera at a plant or animal and get real-time suggestions. Massive community of expert identifiers who review and confirm observations. Data feeds into GBIF (Global Biodiversity Information Facility) for scientific research. Completely free. Builds public engagement and conservation literacy at planetary scale. The gamification (observation streaks, species counts) drives sustained participation.

Where it falls short: Citizen science, not field science. No structured survey protocols — observation effort is uncontrolled and biased toward trails, roads, and charismatic species. No offline mode (requires data connection for species suggestions). No team coordination features. Data quality depends on community verification, which is excellent for birds and butterflies in North America, and sparse for everything else everywhere else. Observations are public by default (you can obscure locations for sensitive species, but the default is open). Cannot build custom observation forms — the data model is fixed. Not suitable for systematic monitoring, population surveys, or any study requiring standardized sampling effort.

Best for: Engaging the public, building species distribution maps at continental scale, and having a delightful experience identifying that weird beetle in your backyard. A triumph of citizen science — but citizen science is not the same thing as field science.

Field Log {" "} is built by The Field Company for the people who are already out there. Rangers, field technicians, ecologists, Indigenous land stewards — anyone whose office is a landscape and whose data matters. It is a field-first mobile platform for recording wildlife observations, coordinating teams, and syncing data — even when you have no signal.

What it does well: Designed for the actual rhythm of fieldwork — rapid logging while moving, structured forms when stationary, notes when things get interesting. Works fully offline; data syncs when you get back to signal. Real-time team coordination — everyone on a project sees what has been recorded, eliminating duplicate effort and keeping the team aligned. Supports photo attachments, GPS waypoints, individual animal IDs, custom observation forms, and habitat condition logs. Export to CSV, GIS, and standard biodiversity database formats. Runs on any modern smartphone. The data is yours — no vendor lock-in, no proprietary formats, no surprise pricing changes.

What distinguishes Field Log is that it was not designed by developers who read about fieldwork. It was designed with the understanding that field science has its own cadence — bursts of observation followed by long silences, the need to capture data with cold fingers, the reality that batteries die and signal disappears and conditions are never ideal. The app gets out of your way and lets you do your job.

Where it falls short: Field Log is young. It does not have ODK's library of pre-built form templates or iNaturalist's computer vision pipeline. It does not offer the enterprise compliance certifications of Fulcrum or the ArcGIS integration depth of Survey123. It is not trying to be everything to everyone — it is trying to be the best tool for the specific people who do the specific work of watching wild things and recording what they see.

Best for: Conservation field teams, ecological monitoring projects, ranger patrols, and anyone who needs their field observations to end up in a structured, analyzable dataset without fighting the tool to get there.

Side-by-Side

Photo by{" "} RDNE Stock project on Pexels

| | CyberTracker | ODK | SMART | Survey123 | Fulcrum | iNaturalist | Field Log | | -------------------------- | ------------ | --------- | ------- | --------- | ------- | ----------- | ----------- | | **Offline-first** | Yes | Yes | Yes | Yes | Yes | No | **Yes** | | **Team sync** | No | No | No | No | No | No | **Yes** | | **Icon-based interface** | Yes | No | No | No | No | No | No | | **Wildlife-specific** | Yes | No | Yes | No | No | **Yes** | **Yes** | | **Custom forms** | Limited | **Yes** | Limited | **Yes** | **Yes** | No | **Yes** | | **Free tier** | Full | Self-host | Full | Limited | No | **Full** | **Yes** | | **Species identification** | No | No | No | No | No | **Yes** | No | | **Law enforcement ready** | No | No | **Yes** | No | No | No | No | | **Data ownership** | You | You | You | Esri | Fulcrum | You | **You** | | **Learning curve** | Week | Day | Week | Day | Day | Minutes | **Minutes** | | **Built for field rhythm** | Yes | No | Yes | No | No | **Yes** | **Yes** |

Which One Should You Use?

If you are managing a national park with armed ranger patrols, use{" "} SMART — it was built for you and nothing else comes close.

If you are running a large-scale household survey for a global health study, use ODK — its form engine is the standard for a reason.

If you are an environmental consultancy already paying for ArcGIS, use{" "} Survey123 — the integration will save you hours.

If you are inspecting infrastructure assets at known locations, use{" "} Fulcrum — its AI-assisted workflows are genuinely impressive.

If you are a citizen scientist or want to engage the public, use{" "} iNaturalist — it is one of the best things the internet has ever produced.

If you respect tradition and work in the Southern African bush, use{" "} CyberTracker — it earned its place in conservation history.

But if you are a field scientist, ecologist, ranger, or land steward who needs to record what you see, coordinate with your team, and trust that your data will be clean, accessible, and yours — use Field Log. It is the tool built for the work you actually do. Get started at fieldlog.thefieldco.com, or visit thefieldco.com to learn more about The Field Company.

What We Learned Building Offline-First Software for the Field

Thu, 07 May 2026 00:00:00 GMT

import { Image } from "astro:assets"; import BlogImage from "../../components/blog/BlogImage.astro"; import OfflineFir2 from "../../assets/blog/offline-first/10825089_vietnam_photographer.jpg"; import OfflineFir1 from "../../assets/blog/offline-first/10366330_ron_lach.jpg"; import remotePhoneUse from "../../assets/blog/offline-first/7065478_alexander_kondibko.jpg"; import syncState from "../../assets/blog/offline-first/sync-state-machine.svg"; import fieldTesting from "../../assets/blog/offline-first/5293197_hong_son.jpg"; import offlineArch from "../../assets/blog/offline-first/offline-architecture.svg";

"Works offline" are the hardest three words in software. Not "scales to millions." Not "real-time collaboration." Offline. Three words that sound like a feature but are actually a complete architectural commitment that touches every layer of your stack — and every layer can fail in ways you will not anticipate.

We learned this the hard way. Over two years of building Field Log, a conservation data collection app that runs on $50 Android phones in places with no cell signal, no Wi-Fi, and no IT support, we broke nearly everything there was to break. Duplicate records. Clock skew that corrupted temporal ordering. A sync that took four hours. A phone that ran out of storage mid-expedition. Safari deleting an entire PWA's data because the user hadn't opened it in a week.

This is not a product announcement. This is an engineering post-mortem. Here is what broke, what held up, and what we would do differently.

The Offline-First Spectrum

Not all "offline" is the same. There is a spectrum, and where you land on it determines everything about your architecture.

Offline-Capable

Firebase Firestore is the canonical example. It caches recently-read data and queues writes when the network drops, then syncs when connectivity returns. This is not a database — it is a cache. The SDK decides what stays and what gets evicted. Pending writes are dropped after roughly 30 days. You cannot query data that has not been fetched before going offline. "Offline-capable" means "survives brief disconnections." A field team offline for two weeks is not a brief disconnection.

Offline-First

ODK Collect is the standard here. The app stores a complete working dataset in a real embedded database (SQLite). Offline is the default state. Sync is a background operation, not a prerequisite for the app to function. Forms are downloaded ahead of time. Submissions are queued locally. When you get back to signal, you upload. Simple, battle-tested, and limited: each device is an island. No team coordination. No partial records. All-or-nothing form submission.

Local-First

Ink & Switch's 2019 paper defined seven ideals: fast (no spinners), multi-device sync, offline for weeks or months, collaboration, longevity (the app outlives the company), privacy and ownership, and user control. Martin Kleppmann later added the acid test: if the developer goes out of business and shuts down the servers, does the app still work? If the answer is no, it is not local-first.

Where Field Log Lands

We are offline-first with local-first aspirations. The app works fully offline with SQLite on device. Team sync happens when connectivity returns, using a server-authoritative protocol. Data is yours — export to CSV, SQLite dump, GeoJSON. We are not yet at Kleppmann's third test: the sync server is required for multi-device coordination. Our sync protocol is documented but not yet a standalone open standard. That is on the roadmap. For now: the app works without signal, and the data is never held hostage.

Alexander Kondibko on Pexels`} />

The Stack

Every architectural decision is a tradeoff. Here are ours, with the reasoning behind each one — and the downsides we accepted.

SQLite on Device

We did not choose SQLite. The constraints of field hardware chose it for us. On a $50 Android phone with 1GB of RAM and 8GB of storage, you cannot run Postgres. You cannot run WASM-based databases with acceptable performance. You cannot afford the abstraction tax of IndexedDB wrappers on underpowered JavaScript engines.

SQLite is a single file. Zero configuration. Zero administration. Full SQL with indexes, triggers, and views. Atomic transactions. It runs on every phone ever made. It competes with fopen(), not with Postgres — and in the field, that is exactly what you want. A database that treats itself as infrastructure, not as a service.

We use WAL mode for concurrent reads during writes, raw SQL with parameterized queries (no ORM — ORMs generate unpredictable query plans, and on low-end hardware that matters), and SQLCipher for encryption at rest. The local database has never corrupted. It has never lost a record. In two years of field use across hundreds of devices, SQLite has been the one component we never had to apologize for.

Server-Authoritative Sync

The server is the source of truth. Clients push changes up and pull changes down. The protocol is delta-based: only changed rows move across the wire, identified by a monotonically increasing version number assigned by the server. Sync is checkpoint-resumable: if the connection drops at 60%, the next sync resumes from the last acknowledged checkpoint. No starting over.

We chose this over CRDTs deliberately. CRDTs guarantee deterministic conflict-free merges without coordination, which sounds perfect for offline use. In practice, for structured field data — observation records, GPS waypoints, form submissions — CRDTs are overkill. They require per-field conflict tracking. They accumulate unbounded operation history. The WASM-based implementations (Automerge) hit memory ceilings on low-end devices. Cinapse, a production app that tried this path, reported 89% fewer support requests and 66% lower hosting costs after moving from CRDTs to server-authoritative sync.

We are not syncing collaborative text documents. We are syncing structured observations where each record has a single author and conflicts are rare by design. For this use case, a server-assigned version number and idempotent writes solve the problem with a fraction of the complexity.

What We Did Not Use

Firebase/Firestore. The offline cache is LRU-evicted. Pending writes are dropped after ~30 days. You cannot query data you have not previously fetched. This is not a debate — Firestore is not suitable for multi-week offline field work. It was not designed for it.

IndexedDB. On paper, it is the browser's offline database. In practice, Safari evicts IndexedDB data after 7 days of inactivity on iOS. Chrome and Firefox have variable quotas that depend on available disk space. The storage is "best effort" — the OS can delete it under pressure without warning. Building offline-first on IndexedDB is building on sand.

PouchDB/CouchDB. The revision tree model is elegant in theory. In practice, compaction bugs, poor performance on large datasets, and the NoSQL-only data model made it a worse fit than raw SQLite for structured conservation data. We wanted joins. We wanted migrations. We wanted to know exactly what was stored and why.

The sync protocol handshake: client sends last-known server version, server returns all changes since that version. Delta-based, resumable, idempotent.

What Broke

This is the part most engineering blog posts skip. Here is everything that failed, exactly how it failed, and what we did about it.

Duplicate Records

The client creates an observation offline and assigns it a UUID. The observation gets queued for sync. The sync request reaches the server, the server creates the record, the response is sent — and then the connection drops before the client receives the acknowledgment. The client, not knowing the record was created, retries. The server creates a second record with a different UUID but identical data.

We had a user with 47 observations that became 141. Every single one, triplicated. The deduplication script we wrote to clean it up was nearly as complex as the sync protocol itself.

The fix: idempotency keys. The client generates a UUID before sync begins. The server stores a mapping of client-UUID → server-ID. When a retry arrives with the same client UUID, the server returns the existing server ID instead of creating a new record. Every write operation is idempotent by construction. This is not a novel idea — Stripe's API has used idempotency keys for a decade — but we learned it the hard way.

Clock Skew: The Phone That Thought It Was 2014

A field worker's phone had its battery pulled and replaced. The system clock reset to the factory default: January 1, 2014. The worker then recorded two weeks of observations. Every record carried a 2014 timestamp. When the phone eventually connected to a network and the clock corrected itself, the temporal ordering of the entire dataset was corrupted. Observations from 2026 appeared to have been made two years before the app existed.

We were using wall-clock timestamps for ordering and conflict resolution. This was a mistake. You cannot trust the device clock. Ever. $50 phones in remote areas may never have connected to an NTP server. Battery pulls reset clocks to epoch. Some devices ship with the wrong timezone and never correct it.

The fix: the server assigns a monotonic version number to every record at sync time. The client uses the server's version, not the device clock, for ordering. Local timestamps are stored as metadata only — useful for the user, never used for logic. A hybrid logical clock would be the ideal solution, but server-assigned versions solved 95% of the problem with 5% of the complexity.

The Sync That Took Four Hours

A field team returned from a two-week expedition with 847 observations and roughly 1,200 photos. They connected to a weak 3G signal at the ranger station and hit sync. Four hours later, it was still running. The photos were 6-8MB each — full-resolution JPEGs from a phone camera. Total payload: over 7GB. Over a connection that averaged 80KB/s.

The problem was not just the bandwidth. The sync protocol was uploading records and photos in a single monolithic request. If the connection dropped at 90%, everything restarted from zero. The app became unusable during sync because the upload monopolized the network thread. Battery drained from 60% to zero during the attempt.

The fix was three parts. First, photo compression: we added a pipeline that resizes images to 1920px on the long edge and applies JPEG compression at quality 70 before upload. Average photo size dropped from 6.5MB to roughly 400KB — a 94% reduction with no visible quality loss on a phone screen. Second, chunked uploads: records and photos are uploaded in batches of 20. Each batch is an atomic checkpoint. If the connection drops, sync resumes from the last completed batch. Third, background sync: the upload runs on a background thread via Android's WorkManager, with constraints for unmetered network and adequate battery. The user can keep working while sync runs.

The $50 Phone That Ran Out of Storage

We tested on a Tecno Spark Go — 1GB RAM, 8GB internal storage, Android Go edition. After three months of daily use, the SQLite database hit 1.2GB. The phone had 400MB of free space remaining. Android started killing background processes. The camera app refused to take photos. The sync process crashed with an out-of-disk-space error that we were not handling.

The SQLite file had bloated for two reasons. First, we were storing full-resolution photos as BLOBs in the database instead of as files on disk with database references. Second, we were retaining soft-deleted records indefinitely instead of garbage-collecting them after sync confirmation.

The fix: photos live on the filesystem with SQLite storing only filepaths, thumbnails, and metadata. Soft-deleted records are purged from the local database after successful server sync. We added a storage monitor that warns the user when free space drops below 200MB. We added an explicit "Clear synced photos" option. And we added a worst-case handler: if a write fails with a disk-full error, the app shows a clear message — "Storage full. Free up space or your new observations will not be saved" — instead of silently failing.

Schema Migration on a Stale Client

We shipped an update that added a new required field to the observation schema: habitat_type. The server was updated. New clients downloaded the update. But one device was offline for six weeks — a researcher on an extended expedition in a remote valley. When they finally connected and tried to sync, the server rejected every observation because the client's schema was missing the required field. 312 observations failed to sync. The error message was an HTTP 422 with a JSON body. The user saw a spinner.

The fix: version negotiation in the sync protocol. The client sends its schema version in the sync handshake. The server knows which fields are required as of which version. For clients on older versions, the server accepts records with missing new fields and fills them with a default value (habitat_type: "unknown"). The client is notified that an update is available. Records are never rejected because of a schema mismatch. Backwards compatibility is not optional when your clients can be offline for longer than your release cycle.

Safari Deleted Everything

Before the native Android app, we experimented with a PWA. It used IndexedDB for local storage, registered a service worker for offline caching, and had a manifest for install-to-home-screen. It worked beautifully in testing. A field tester in northern Kenya installed it on an iPhone, used it daily for a week, then switched to other tasks for nine days. When they opened the PWA again, the IndexedDB database was gone. Every observation. Every photo. Every form.

Safari's storage policy on iOS evicts IndexedDB data after 7 days of app inactivity. This is documented behavior. It is not a bug — it is a policy designed to prevent websites from consuming storage indefinitely. It also makes PWAs fundamentally unsuitable for offline-first field work on iOS. You cannot build an app that says "your data is safe offline" when the operating system reserves the right to delete that data if you do not open the app every week.

We killed the PWA experiment. Data that only exists locally is data that is already lost. The native Android app, using SQLite in the app's protected storage directory, is not subject to browser eviction policies. This is the only safe choice for data that matters.

"I Thought It Synced"

A ranger recorded 89 observations over three days. The app showed a green checkmark next to each one — the local-save confirmation. The ranger assumed this meant "synced to server." It did not. The phone had no signal for the entire three days. The ranger returned to headquarters, handed in the phone, and the device was wiped for the next patrol. The data had never left the device.

This was not a technical failure. It was a UX failure. We had designed the sync indicator to be subtle — a small icon in the corner that turned from cloud-outline to cloud-with-checkmark. Invisible. Meaningless. We had violated the most important rule of offline-first design: the user must always know whether their data is only local or safely on the server.

The fix was a complete redesign of sync indicators. Every screen now shows a persistent sync status bar: "12 observations pending upload," "Last synced: 3 hours ago," "Sync requires 8MB — connect to Wi-Fi." The main screen has a large sync button with a badge count. When sync is pending, the badge is red. When sync completes, it turns green and shows a timestamp. You cannot miss it. We also added a warning dialog that appears if you try to log out or clear the app while records are pending sync: "You have 12 observations that have not been synced. If you continue, this data will be lost." Making sync state impossible to ignore is not annoying — it is honest.

Sync state machine — on device, syncing, synced, and failure states

Every failure mode we encountered, categorized. Duplicate records and clock skew accounted for 60% of data issues. Storage exhaustion and UX failures accounted for the rest.

Hong Son on Pexels`} />

What Worked

Not everything broke. Some decisions held up from day one. These are the things we got right, often because we were forced into them by the constraints of the hardware.

SQLite as the Local Database

We have never had a corrupted database. We have never had a lost record that was successfully written to SQLite. The database has survived force-quits mid-transaction, battery pulls, storage-full errors, and Android killing the process to reclaim memory. SQLite's atomic commit model means a write either completes or it does not — there is no partial state. For an app whose entire purpose is "do not lose the data," this property is everything.

We use WAL (Write-Ahead Logging) mode. This allows concurrent reads during writes, which matters when you are logging observations rapidly and the UI needs to refresh a list. We use PRAGMA synchronous=NORMAL — a calculated tradeoff. FULL synchronous would survive OS crashes but doubles write latency. On a $50 phone where every millisecond of UI blocking matters, NORMAL is the right call. In two years, we have not seen a single corruption from this choice.

Append-Only Observation Records

Observations are never updated. They are only appended. If a user needs to correct an observation, they create a new revision record that references the original. The original record remains immutable. This eliminates conflicts entirely — you cannot have a write-write conflict if nobody writes to the same record twice. It also gives us a full audit trail. Every change is traceable. For conservation data that may end up as evidence in court, immutability is not a nice-to-have.

The tradeoff is storage growth. An observation with ten revisions stores eleven rows instead of one. We accept this. Storage is cheaper than data loss, and storage is predictable — you can plan for it. Conflict resolution bugs are not predictable.

Delta Sync with Checkpoints

Our sync protocol works like Git. The client sends its last-known server version. The server returns all changes — inserts, updates, deletes — since that version. Each batch of 50 changes is a checkpoint. If the connection drops, the client resumes from the last acknowledged checkpoint. The server only sends the rows that changed, not the full dataset.

For a project with 50,000 observations, the initial sync might transfer ~15MB (compressed). Subsequent syncs transfer kilobytes — only what changed since the last connection. A ranger syncing after a day in the field typically uploads 200-500KB and downloads 50-100KB. On a 2G connection at 20KB/s, that is 15-30 seconds. Tolerable.

Optimistic Local Writes

When a user saves an observation, the UI updates immediately. The record is written to SQLite and appears in the observation list with no perceptible delay. Sync to the server happens in the background. The user never waits for a network round-trip. This is not just about performance — it is about trust. If the app feels fast and responsive, the user trusts it. If the user trusts it, they use it. If they use it, data gets collected. The entire chain starts with the UI not blocking.

Photo Compression Pipeline

We compress every photo before storage, not just before upload. On ingestion, the image is resized to 1920px on the long edge (sufficient for species identification and habitat documentation), compressed to JPEG quality 70, and stripped of EXIF data (GPS is stored separately in the observation record). A typical phone photo goes from 6-8MB to 300-500KB. We store the compressed version and discard the original. This saves storage, speeds up sync, and costs nothing in practical image quality for the use case. If a user needs the full-resolution original, they can export it before compression — but in two years, no user has asked for this.

Testing on Real $50 Phones

We bought a collection of the cheapest Android devices available in African markets: Tecno Spark Go, Infinix Smart, Itel A-series. We test every release on these devices. We test on 2G and EDGE network profiles. We test with airplane mode toggling every few minutes. We test with the system clock set wrong. We test by filling the storage to 95% and then trying to record observations.

Emulators do not tell you the truth. An Android emulator on a MacBook Pro has effectively unlimited RAM and storage, a fast CPU, and a simulated network that behaves nothing like a real 2G connection in a valley. Real hardware surfaces real problems: the Infinix Smart's camera takes 4 seconds to initialize. The Tecno Spark Go's GPS takes 45 seconds to get a first fix. The Itel A-series has a bug where System.currentTimeMillis() returns zero after a cold boot until the network time sync completes. You cannot find these on an emulator.

Ron Lach on Pexels`} />

Testing Without Signal

Testing offline-first software requires simulating a world where the network is unreliable, devices behave unpredictably, and the user has no one to call for help. Here is our testing toolkit.

Network Simulation

We use Linux tc (traffic control) to simulate real network conditions on our test server. We have profiles for EDGE (20KB/s down, 10KB/s up, 600ms latency), 3G (384KB/s, 150ms), and 4G (5MB/s, 50ms). We add packet loss (1-15%) and jitter. We simulate connections that drop for 30 seconds every 5 minutes — common in areas with patchy coverage. The test suite runs against all profiles.

Time-Travel Testing

We have an automated test that advances the system clock by 14 days, generates observations at various points along that timeline, then connects to the server and verifies that all records are ordered correctly by server-assigned version, not by local timestamp. Another test sets the clock to January 1, 1970 (Unix epoch) and verifies the app does not crash, does not produce negative timestamps, and syncs correctly. A third test sets the clock to January 19, 2038 (32-bit overflow). We run these on every commit.

Storage Pressure Tests

We fill the device storage to 95% using a script that writes large files to the filesystem. Then we run the app: create 500 observations with photos, attempt sync, verify that the app warns the user before storage exhaustion, and verify that no data is lost when writes begin to fail. The app must degrade gracefully — show a clear error, stop accepting new photos, preserve existing records — rather than crashing or silently losing data.

Conflict Simulation

Two instances of the app modify the same observation while both are offline. Instance A changes the species. Instance B changes the count. Both connect and sync. The test verifies that both changes are preserved (as separate revision records), neither is silently overwritten, and the merge result is deterministically correct. We run this with 2, 3, and 5 concurrent offline editors.

Sync Interruption Fuzzing

We start a sync of 1,000 records and randomly kill the network connection at various points (10%, 25%, 50%, 75%, 90% complete). We verify that the client can resume from the last checkpoint, no records are duplicated, no records are lost, and the app remains usable during and after the interruption. We also force-kill the app mid-sync and verify recovery on restart.

None of these tests are exotic. They are the bare minimum for software that must work when nothing else does. If you build offline-first software and you are not running these tests, you are shipping bugs you have not yet discovered.

Vietnam Photographer on Pexels`} />

Why Offline-First Matters for Conservation

Most software can afford to be online-only. Slack does not need to work when you are off the grid. Netflix does not need to stream without Wi-Fi. The consequences of a failed sync are an annoyed user, a lost message, a buffering spinner.

In conservation, the consequences are different.

A ranger records a poaching incident — GPS coordinates, photos of snares and carcasses, descriptions of the perpetrators. If that data does not survive until sync, it is not a UX bug. It is evidence destruction. Court cases depend on this data. The chain of custody must be unbroken from the moment of observation to the moment of prosecution.

A field team spends two weeks and $30,000 on an expedition to survey an endangered species population. They record every sighting, every track, every habitat measurement. If sync fails and the data is lost, the expedition's scientific output is zero. The funding is wasted. The population estimate has a two-week gap that cannot be recovered.

A community conservation program trains local people to monitor wildlife in their area. They use the cheapest phones available. They have no IT support. If the app crashes, loses data, or requires an internet connection to function, they stop using it. The monitoring stops. The data stops. The area goes dark.

The app has one job: do not lose the data. Not "provide a delightful user experience." Not "leverage AI to generate insights." Not "scale to millions of concurrent users." One job. Everything else is secondary. If you internalize that constraint — that data loss is not an annoyance but a failure of the entire purpose of the software — your architectural decisions change. You stop optimizing for speed and start optimizing for durability. You stop treating offline as an edge case and start treating it as the default. You stop shipping features and start removing failure modes.

What We'd Do Differently

If we started over tomorrow, here is what we would change.

Start with Idempotency

Every write operation should be idempotent from day one. Not added later when duplicates start appearing. Idempotency keys are not a feature — they are the foundation of a correct sync protocol. We would design the sync API so that every request can be safely retried without creating duplicates, without requiring the client to track which requests succeeded, and without the server needing to maintain per-client state beyond the idempotency key mapping.

Never Trust the Device Clock

We would use server-assigned monotonic version numbers for all ordering and conflict resolution from the start. Local timestamps would be stored as user-facing metadata only, never used for logic. If we needed causally-consistent ordering across devices without a server, we would use hybrid logical clocks. We would never, under any circumstances, use Date.now() to decide which record wins a conflict.

Make Sync State Impossible to Ignore

The default state of any offline-first app should be: the user can see, at a glance, exactly which records have been synced and which have not. The sync indicator should be the most prominent element on the screen when records are pending. The penalty for ignoring it should be explicit — a dialog that says "data will be lost" in plain language, not a subtle warning icon. We would design this before designing any other UI element.

Test with the Clock Set Wrong

We would add clock-skew tests to the CI pipeline on day one. Epoch (1970), year 2038 overflow, clock set 5 years in the past, clock set 3 days in the future, clock that jumps backward by 6 hours (battery pull simulation). Every sync test would run against every clock state. These tests cost nothing to write and would have caught our worst data corruption bug before it reached a user.

SQLite WAL Mode from the Start

We shipped the first version with the default rollback journal mode. Write operations blocked reads. When a user was rapidly logging observations, the list view would freeze for 100-300ms on each save. Switching to WAL mode eliminated the freezes entirely. It is a one-line change: PRAGMA journal_mode=WAL. There is no reason not to do it on day one.

Buy the Actual User Hardware

Do not test on a flagship phone. Do not test on an emulator. Buy a handful of the cheapest Android devices available in the markets where your users live. In our case: Tecno, Infinix, Itel. Test every feature on every device. The bugs you find on a Tecno Spark Go are not the bugs you find on a Pixel. The Tecno's GPS is slower. Its camera is slower. Its JavaScript engine is slower. Its storage is smaller. If your app works well on the Tecno, it will work well on everything else. The reverse is not true.

Open-Source the Sync Protocol

We built the sync protocol as an internal component. We should have built it as a documented, standalone protocol that anyone could implement against any backend. Kleppmann's third test of local-first — the app outlives the company — requires that the protocol be open. We are working on this. It is the single most important thing we can do for the long-term integrity of the data people trust us with.

The Tools We Use

A list, without commentary, of the tools that have survived two years of field use.

Component	Tool	Why
Local database	SQLite (WAL mode)	Never corrupted. Never lost data. Runs on every phone.
Query layer	Raw SQL, no ORM	Predictable query plans. No abstraction tax on low-end hardware.
Encryption	SQLCipher	Encrypt-at-rest. Full database encryption, transparent to application code.
Server database	PostgreSQL	Rock-solid. Row-level security for multi-tenant data isolation.
Sync protocol	Custom (delta-based, checkpoint-resumable)	Purpose-built for structured field data. Simpler than CRDTs. More reliable than generic sync engines.
Background sync	WorkManager (Android)	Handles Doze mode, network constraints, battery optimization. The only reliable way to run background work on modern Android.
Photo compression	libjpeg-turbo via native code	Resize to 1920px, JPEG quality 70. 8MB → 400KB average.
Network simulation	Linux tc (traffic control)	EDGE, 3G, 4G profiles. Packet loss. Jitter. Connection drops.
Test devices	Tecno Spark Go, Infinix Smart, Itel A-series	$50-70 each. The actual hardware our users carry.

Offline-first architecture — device, sync protocol, and server stack

Field Log is open source. We built it in public — the bugs, the fixes, the four-hour syncs, the phone that thought it was 2014. Everything we learned is in the code. If you build software for the field, for conservation, for anyone whose data matters and whose connection does not, steal what you need. The protocol docs are at{" "} github.com/thefieldcompany . The only thing we ask is that you tell us what you broke, so the next team does not have to.

Built with: SQLite, PostgreSQL. Evaluated:{" "} PowerSync, ElectricSQL, RxDB, Automerge. Inspired by:{" "} Ink & Switch local-first , Martin Kleppmann's work on CRDTs, ODK's offline architecture. Test devices: Tecno Spark Go and Infinix Smart series (Android 11 Go Edition).

The Field Co

Open-Source Conservation Technology

How to Set Up a Biodiversity Monitoring Protocol

Wed, 22 Apr 2026 00:00:00 GMT

import { Image } from "astro:assets"; import BlogImage from "../../components/blog/BlogImage.astro"; import Biodiversi2 from "../../assets/blog/biodiversity-monitoring/10365936_ron_lach.jpg"; import Biodiversi1 from "../../assets/blog/biodiversity-monitoring/10365934_ron_lach.jpg"; import fieldMonitoring from "../../assets/blog/biodiversity-monitoring/9544941_ron_lach.jpg"; import fieldResearcher from "../../assets/blog/biodiversity-monitoring/6033820_cottonbro_studio.jpg"; import decisionTree from "../../assets/blog/biodiversity-monitoring/monitoring-decision-tree.svg";

Bad monitoring is worse than no monitoring. It consumes budgets, burns out field teams, and produces datasets that cannot answer the questions they were designed to answer — or any questions at all. We have seen five-year programmes that collected exactly one statistically usable data point per year, and ten-year programmes that never defined what a decline would look like. This guide exists so you do not become either of those programmes.

Monitoring is not photography. It is not a species list. It is the systematic collection of data that can detect change over time with a known level of confidence. If you cannot put a number on the change you can detect and the confidence you can detect it with, you do not have a monitoring protocol — you have a field trip.

The good news: designing a protocol that actually works does not require a PhD in biostatistics. It requires clarity about what you are measuring, discipline about how you sample, and the humility to pilot before you commit. Here is how.

What to Monitor

Before you buy a single camera trap, ask yourself: what exactly are you trying to measure? Not "biodiversity" — that is a concept, not a variable. A population trend? A species list? A habitat condition score? The question drives everything that follows, and most programmes fail here, at the starting line, because they never answer it with enough specificity.

There are three defensible approaches. Each has a different purpose, a different cost profile, and a different statistical relationship to the truth.

Approach 1: Indicator Species

Pick two to five species and track them relentlessly. This is the pragmatic approach. You get high statistical power because you are sampling intensively for a small number of targets. You can afford detection probability modelling for each species individually. Your field team develops deep expertise. Your data actually means something.

The key is picking the right indicators. A good indicator species is detectable (you can reliably observe it with your chosen method), sensitive to the pressure you care about (logging, poaching, climate shift), and representative — its response correlates with other species in the community. NEON (the US National Ecological Observatory Network) uses sentinel taxa — ground beetles, mosquitoes, small mammals, ticks, and plants — chosen because each responds to a different axis of environmental change. You do not need NEON's budget to adopt the logic.

Pros: Statistically powerful. Affordable. Clean data. Cons: What if your indicators miss something? If the pressure you care about is poaching for bushmeat and you monitor only birds, you may detect nothing while mammal populations collapse.

Approach 2: Keystone Species and Process Monitoring

Monitor a species whose presence or behaviour tells you about an ecosystem process, not just a population count. Apex predators tell you about trophic integrity. Seed-dispersing frugivores tell you about forest regeneration. Dung beetles tell you about nutrient cycling.

This approach is harder to quantify than indicator species monitoring but can be more meaningful. The presence of a breeding pack of African wild dogs tells you something about prey density, habitat connectivity, and human pressure that a count of all 47 bird species in the area does not. The limit is that process interpretation requires expert knowledge and is difficult to automate or scale.

Approach 3: Full Taxonomic Surveys ("Everything")

This is the seductive one. Deploy every method. Count every taxon. Build the definitive species list. It sounds comprehensive. It is usually a mistake.

A full survey dilutes effort across too many targets to achieve statistical power for any one of them. You end up with presence data for 300 species and trend data for zero. If you cannot afford detection probability modelling for every species in your list — and almost nobody can — do not pretend you are monitoring all of them. Call it what it is: an inventory. An inventory tells you what is there. Monitoring tells you what is changing. These are different activities, and confusing them is the single most expensive mistake in conservation science.

Our advice: Start with indicator species. If you have the budget, add one keystone process. Leave the everything survey for your baseline year — do it once, well, and use it to select your indicators.

Cottonbro Studio on Pexels`} /> Ron Lach on Pexels`} />

How to Sample

You know what you are monitoring. Now: how do you actually collect the data? The method must match the question, and no method is perfect. Every sampling technique has a detection bias — a set of species it sees well, a set it sees poorly, and a set it misses entirely. Understanding these biases is more important than the method itself.

Monitoring method decision tree — match your monitoring question to the right sampling approach

Method	Best For	Pros	Cons
Line transects	Large mammals, birds in open habitat, primates	Well-established statistical framework (distance sampling). Repeatable. Low equipment cost.	Requires skilled observers. High labour cost. Detection falls with distance — the animals you miss are not random.
Point counts	Birds, amphibians (acoustic)	Simple to standardise. Good for community-level trends. Volunteer-friendly.	Detectability varies dramatically with vegetation density, time of day, and observer skill. Not suitable for cryptic or nocturnal species.
Camera trap grids	Medium to large terrestrial mammals, ground birds	Works 24 hours. Permanent record for verification. No observer bias. Density estimation via capture-recapture for marked species.	Equipment cost ($200–600/unit). Theft and elephant damage are real. Data processing is substantial — thousands of images to classify. Fails for arboreal and small-bodied species.
Acoustic recording	Birds, bats, amphibians, orthoptera, marine mammals	Low-cost hardware (AudioMoth ~$60). Permanent record. Multi-taxon from single device. Automated analysis improving rapidly.	Battery and storage management. Automated classifiers are good but not perfect — false positives and negatives require validation. Performance drops in high-wind and rain.
eDNA	Aquatic vertebrates, invertebrates, some terrestrial taxa via soil/air	Detects species that evade all other methods. One water sample can detect dozens of fish species. Rapidly maturing technology.	Contamination risk is real and consequential. Cannot estimate abundance (yet). Reference libraries incomplete for many regions. Lab costs are significant — $50–200/sample. Requires cold chain.

Camera Trap Grids in Detail

Camera traps have become the default tool for terrestrial mammal monitoring. They should not be the automatic choice — transects and acoustic methods are cheaper and more appropriate for many taxa — but when mammals are your target, camera traps are the best tool we have.

The gold standard is the TEAM (Tropical Ecology Assessment and Monitoring) protocol: 60 camera trap points per array, arranged in a systematic grid with 1 km spacing between points, deployed for a minimum of 30 consecutive days per season. This density and duration is not arbitrary — it is the product of power analysis that established the minimum effort required to detect a 5% annual change in occupancy for medium-to-large terrestrial mammals over a 5-year period. The TEAM network has run this protocol across 17 sites on three continents since 2009, and the protocol documentation is public, peer-reviewed, and free.

If 60 cameras sounds like a lot: it is. The TEAM protocol was designed for sites of 100–300 km². Most smaller programmes adapt it — 20–30 cameras, 30-day deployments, 1 km minimum spacing. The spacing matters more than the count. Cameras too close together are not independent sampling units; you are photographing the same animals and calling it replication. This is a statistical error, not a budget choice.

Spatial Design

Random is better than convenient. Always. Trails and roads are easy to walk but they are not random samples of the landscape — they follow topography, avoid thick vegetation, and attract different animal behaviour than off-trail habitat. A random or systematic grid will be harder to deploy and will produce data that you can actually analyse. The convenient sample will be easier and will produce data that statisticians will reject.

If you must use trails for access, offset your sampling points by a consistent distance — 50 to 100 metres perpendicular to the trail. This is not perfect (it still biases toward trail-adjacent habitat) but it is better than placing cameras on the trail itself.

Pilot Everything

Before you commit to a protocol, run one full pilot season. Deploy your chosen method at your chosen density and duration. Analyse the pilot data for detection rates, species accumulation curves, and coefficient of variation. Ask: can this design actually detect the magnitude of change I care about? If the answer is no — and it often is on the first attempt — adjust the density, the duration, or the target species, and pilot again. One pilot season costs less than five years of useless data.

How Often

Seasonality drives detectability. In the wet season, vegetation is dense, animals disperse, and your cameras photograph leaves moving in the wind. In the dry season, animals concentrate at water sources and your detectability spikes. Neither season is "better" — but you must sample consistently across seasons, or your trend line will reflect the weather, not the population.

Minimum Viable Design

For a programme that can detect a 30% decline over 10 years with reasonable statistical power (80%), you need:

30–50 independent sampling units (camera points, transect segments, acoustic stations). Fewer than 30 and your confidence intervals will be too wide to detect anything short of a population crash.
Annual surveys at minimum. Biannual (wet/dry) is better. Quarterly is a luxury. The trend you can detect is a function of the number of data points in time, not just space.
At least 3–5 years of data before your first trend analysis. Population time series are noisy. Two data points — year one and year two — tell you nothing about a trend. They tell you about the difference between two particular field seasons.

Occupancy Modelling

The most important paper in monitoring methodology is MacKenzie et al. (2002): Estimating site occupancy rates when detection probabilities are less than one. Before this paper, monitoring programmes treated "not detected" as "absent." That assumption is wrong. Detection probability (p) is almost never 1.0. If p = 0.5 and you visit a site once, you will miss a present species half the time. Five visits gets you to a 97% chance of detecting it if it is present. Occupancy models separate the biological process (occupancy ψ) from the observation process (detection p) using repeat visits. This is not an advanced option — it is the minimum standard for credible monitoring.

The practical implication: your protocol must include repeat surveys at each sampling unit within a season. Three to five repeat visits (or 3–5 independent sampling occasions within a 30-day deployment for cameras) is the typical minimum for occupancy modelling. If you visit each site only once per season, you cannot model detection probability, and you cannot separate true absence from failed detection. Your occupancy estimates will be biased low, and the bias will vary with conditions you are not measuring.

Document Every Protocol Change

You will change your protocol. New cameras will have different trigger speeds. A new observer will join the team. A road will wash out and you will shift a transect. Every change, documented, is a covariate in your analysis. Every change, undocumented, is a confounding variable that may invalidate your entire trend estimate. Keep a protocol log. Date every entry. It will save your dataset.

Ron Lach on Pexels`} />

What Tools to Use

Tools do not design your protocol — but the wrong tool will break it. A field app that crashes when you have no signal, a GPS that runs out of battery at noon, a camera trap whose trigger speed misses every animal that walks past — these are not inconveniences. They are data loss events. Here is an honest assessment of the hardware and software that field programmes actually use, with no vendor enthusiasm.

Data Collection Apps

CyberTracker — The original. Icon-based interface built for illiterate San trackers in the Kalahari in 1996. Still works offline. Still free for conservation. Still respected. The desktop component requires Windows and looks like it. The mobile interface has not been significantly modernized in a decade. Data export is clunky. If you work in Southern Africa and your field team is comfortable with the icon paradigm, CyberTracker is a proven choice. If you are starting from scratch elsewhere, there are smoother options.

ODK (Open Data Kit) — The standard for form-based data collection. Build forms in XLSForm (Excel), deploy to Android via ODK Collect. Free if you self-host. ODK Cloud starts at $199/month for 10K submissions. Massive community, rigorous data validation, offline sync. The trade-off: ODK is a form engine, not a wildlife tool. There are no built-in species lists, no tracking protocols, no field guides. You build everything. If you have the technical capacity to design your own forms and you need bulletproof data validation, ODK is excellent.

SMART Mobile — Purpose-built for ranger patrols and conservation law enforcement. Developed by WWF, WCS, ZSL, and others. End-to-end patrol management: plan routes, log observations, document illegal activity, generate reports. Free and open-source. Heavy to set up — SMART is a full desktop application plus mobile app plus server component. Training required. If you are managing a protected area with formal patrol mandates, SMART is the tool. If you are doing general ecological monitoring, it will feel like driving a fire engine to the grocery store.

Field Log — Built by The Field Company for field-first observation logging. Offline-first. Real-time team sync when you have signal. Structured forms and rapid logging modes in the same app. Export to CSV and GIS formats. Designed for the rhythm of fieldwork rather than the requirements of a server room. Free tier available. Does not yet have ODK's template library or iNaturalist's computer vision. Best for conservation field teams who need their observations to end up in a usable dataset without fighting the tool.

Camera Trap Hardware

Brand	Price (USD)	Trigger Speed	Notes
Browning (Strike Force, Spec Ops)	$150–250	~0.3s	Best value for most programmes. Fast trigger, good battery life, reliable in humidity. The Spec Ops Elite HP5 is the current workhorse recommendation.
Reconyx (HyperFire 2)	$500–650	~0.2s	Industry gold standard. Built for decade-scale deployments. No-glow IR is genuinely invisible. If you need data continuity across a 10-year programme, Reconyx is the benchmark.
Bushnell (Core, Prime, CelluCore)	$100–300	~0.3–0.6s	Widely available. Cellular models (CelluCore) enable near-real-time data retrieval. Build quality is variable — some units last years, some fail in months. Test your batch.
Cuddeback (CuddeLink system)	$200–400	~0.25s	Mesh networking between cameras — one cellular uplink serves up to 24 cameras over radio. Reduces cellular costs dramatically. Proprietary network; you are locked in.

Acoustic Recorders

AudioMoth (~$60/unit) is the breakthrough device. Developed by Open Acoustic Devices, it is tiny, cheap, and capable. You can deploy 50 AudioMoths for the price of one professional recorder. The trade-offs: waterproofing requires a third-party case, battery life tops out around 2–4 weeks depending on duty cycle, and the onboard clock drifts enough to matter for synchronisation. For most field programmes, these trade-offs are worth it.

Wildlife Acoustics Song Meter ($500–1,000/unit) is the professional option. Weatherproof housing, scheduled recording, programmable gain, GPS-synced clock, months of deployment on AA batteries. If you need acoustic data from a remote ridgeline during monsoon season and you cannot afford a failed deployment, this is the tool. The SM4 and Mini are field-proven across every continent.

GPS Collars and Telemetry

GPS collars are not monitoring — they are a specialised subset of it. They tell you about movement, home range, and survival of individual animals. They cost $800–5,000 per collar plus satellite or cellular data fees. They require animal capture, which requires veterinary expertise, ethical approval, and significant risk. They are essential for some questions — corridor use, human-wildlife conflict, mortality causes — and an expensive distraction for others. Do not put a collar on an animal unless the collar answers a question that no other method can answer.

Ron Lach on Pexels`} />

What to Do With the Data

You have spent three field seasons collecting data. Now what?

Most monitoring programmes treat data management as something that happens after fieldwork. This is a mistake. Your analysis pipeline should exist before your data does. Write the analysis script first, with simulated data. If you cannot write the script, you do not understand the question well enough to design the protocol. This sounds backwards but it is the single highest-leverage thing you can do before entering the field.

Metadata Is Not Optional

A dataset without metadata is a collection of numbers with no meaning. Two years from now, someone — possibly you — will open a CSV called camera_data_final_v3_corrected.csv and have no idea what the columns represent, which protocols were used, or whether the dates are in DD/MM or MM/DD format. Metadata prevents this.

Use Darwin Core terms for biodiversity data. Darwin Core is the standard vocabulary maintained by Biodiversity Information Standards (TDWG). It defines exactly what goes in fields like occurrenceID, eventDate, decimalLatitude, scientificName, and basisOfRecord. Using these terms means your data is interoperable with every major biodiversity database on Earth from day one. It also means someone else's parser will understand your data without a phone call.

For dataset-level description, use EML (Ecological Metadata Language). An EML file describes the who, what, when, where, and why of your dataset: project abstract, methodology, geographic coverage, taxonomic coverage, temporal coverage, personnel, and data table structure. EML is the standard that GBIF, DataONE, and most ecological data repositories require. Writing an EML document takes an afternoon. Not writing it costs years of data reusability.

The Analysis Pipeline

Your analysis pipeline should answer the specific question your protocol was designed around. For most monitoring programmes, the pipeline looks roughly like this:

Data validation — check for missing values, duplicate records, coordinate outliers, date inconsistencies. Automate this. Manual validation is error-prone and nobody does it consistently after month three.
Detection histories — construct the matrix of detections/non-detections per species per site per sampling occasion. This is the input to occupancy models.
Occupancy or abundance estimation — estimate occupancy (ψ) with detection probability (p), or density (D) via distance sampling or spatial capture-recapture. The R package unmarked handles occupancy, Distance handles distance sampling, secr handles spatial capture-recapture.
Trend analysis — model the change in occupancy, abundance, or species richness over time, accounting for covariates (season, observer, habitat). The R package rtrim is purpose-built for wildlife trend analysis with missing data.
Reporting — produce estimates with confidence intervals, not just point estimates. A trend of -3% per year with a 95% CI of [-8%, +2%] is telling you something different from a trend of -3% with a CI of [-4%, -2%]. The first is noise. The second is evidence.

GBIF and the Kunming-Montreal Framework

Publishing your data to GBIF (Global Biodiversity Information Facility) is the standard for making biodiversity data discoverable and citable. GBIF requires Darwin Core formatting, a data licence (CC0, CC BY, or CC BY-NC), and an EML metadata document. Publication gives your dataset a DOI — it becomes a citable scientific output, not just an internal report.

The Kunming-Montreal Global Biodiversity Framework (2022) includes 23 targets and a monitoring framework with headline indicators. Target 4 requires parties to "halt human-induced extinction of known threatened species." Target 21 requires "the best available data, information and knowledge" to be accessible to decision-makers. If your monitoring programme produces Darwin Core-compliant data published to GBIF, you are contributing directly to the global indicators that will be reported against in 2030 and 2050.

Common Mistakes

Every one of these we have made, seen made, or been asked to rescue a programme from.

1. No Pilot Season

A programme in northern Mozambique deployed 40 camera traps across 200 km² before testing a single unit. The trigger speed of the cameras they bought was 1.2 seconds — slow enough that an elephant walking at normal pace passed through the detection zone without triggering a single frame. They discovered this after 90 days of deployment. Total data: eleven photographs of grass blowing in the wind. A single camera, placed in the backyard for a week, would have revealed the problem. Cost of the pilot: $200 and seven days. Cost of the error: $12,000 and a lost field season.

2. Inconsistent Protocols

A mammal monitoring programme in East Africa changed camera models between year two and year three. The new cameras had a wider detection zone and a faster trigger. Detection rates increased 40% across all species. The programme reported a "dramatic recovery" of the mammal community to its donors. The actual change in animal abundance was statistically indistinguishable from zero — the cameras were simply better at detecting them. Protocol consistency is not optional. If you must change equipment, run a paired calibration study so you can statistically correct for the difference.

3. Ignoring Detectability

A wetland bird survey in Southeast Asia counted birds seen during a single dawn visit to each site, once per year, for seven years. The raw counts declined 35%, and the programme reported a crisis. When occupancy modelling was applied retroactively — using multiple observers and weather covariates as proxies for detection — the decline disappeared. Bird numbers were stable. Visibility had declined because reed beds were growing denser due to reduced grazing pressure. The birds were still there; the observers simply could not see them. This is not a rare edge case. This is what happens when you treat "not detected" as "absent."

4. Collect Now, Figure Out Later

A biodiversity survey in West Africa collected 14,000 camera trap images across 80 sites. The images sat on an external hard drive for four years because no one had budgeted for the time required to classify them. When a graduate student finally took on the work, the original field team had dispersed, the field notes were lost, and 30% of the images could not be geolocated because the camera metadata format had changed between firmware versions. Classifying camera trap images takes roughly 30–60 seconds per image when done properly. At 14,000 images, that is 120–240 hours of labour — three to six weeks of full-time work. Budget for it before you deploy the cameras, or do not deploy the cameras.

5. No Decision Pathway

Monitoring exists to inform decisions. If your protocol detects a 30% decline in your indicator species, what happens next? Who is notified? What action is triggered? Within what timeframe? If you cannot answer these questions, you are not monitoring — you are documenting decline.

Write the decision pathway before you start. At minimum: define the trigger (e.g. 20% decline in occupancy over three years, p < 0.05), the response (e.g. increase anti-poaching patrols, close area to grazing, commission targeted study), the responsible party (named person, not "the management committee"), and the timeline (e.g. response initiated within 30 days of analysis completion). A monitoring programme without a decision pathway is a report that sits on a shelf while the thing it was monitoring disappears.

What This Costs

Real numbers. No grant-writing optimism. All figures in USD at 2026 prices, exclusive of international travel and institutional overhead.

Tier 1: $500 — The Bare Minimum

What you get: One pilot season of point count or transect surveys for a single taxon (birds or large mammals). One or two field technicians. Clipboard and datasheets — no electronic equipment. Analysis in R or Python on a personal laptop. One site, one season, one question.

Equipment: Binoculars ($100), GPS unit or smartphone ($0–150), field guides ($50), notebooks ($20), local transport ($180).

What you can detect: Large changes (50%+ decline) in common species at a single site. Species list for the site. Baseline for a future proper programme.

Limitations: No detection probability modelling. No camera traps. No occupancy analysis. No trend detection for anything subtle. This is a pilot, not a monitoring programme. Do not call it one.

Tier 2: $5,000 — A Working Programme

What you get: One full field season of camera trap or acoustic monitoring at one site, with occupancy modelling and a written report. Two to four field technicians for deployment and retrieval. Basic analysis pipeline in R. One repeat season (wet/dry) if acoustic.

Equipment: 10–15 Browning camera traps ($1,500–3,750) or 15–20 AudioMoths ($900–1,200), SD cards ($50–100), rechargeable AA batteries and charger ($150), Pelican cases or dry bags for transport ($200), GPS unit ($150), smartphone for data collection with Field Log or ODK (free), local transport and field rations ($800–1,500), basic data processing labour ($500).

What you can detect: 30% occupancy change over 3–5 years for medium-to-large mammals or vocalising birds. Species accumulation curves. Activity patterns. First occupancy estimates with confidence intervals.

Limitations: Single site means no inference beyond that site. Limited statistical power for rare species. Image classification is a time sink — budget for it. Equipment failure and theft will claim 10–20% of your units, and $5,000 does not include a replacement buffer.

Tier 3: $50,000 — A Multi-Year Programme

What you get: Two to three field seasons across two or three sites, with multiple methods (camera traps + acoustics + eDNA pilot). Occupancy and trend modelling. Dedicated data manager. GBIF publication. Decision pathway integrated with management.

Equipment: 30–60 camera traps — mix of Browning and Reconyx ($8,000–25,000), 20–40 AudioMoths ($1,200–2,400), 2–4 Wildlife Acoustics Song Meters for acoustic standardisation ($1,000–4,000), GPS units and satellite communicator ($500–800), ruggedised tablets or phones for data collection ($500–1,000), eDNA sampling kits and lab costs for 200 samples ($10,000–15,000), weather station for site-level covariates ($500), battery chargers, SD cards, spare parts, protective cases ($1,500).

Personnel: Field team leader (3–6 months/year), 2–4 technicians, 1 data manager (part-time), 1 statistician or ecologist for analysis design (consulting). Local field assistants are essential and should be paid, not volunteered.

What you can detect: 10% annual occupancy change over 5 years. Multi-species trends. Relationships between species occurrence and habitat covariates. Community-level change. All with publishable confidence intervals.

Hidden Costs

Things that appear in nobody's budget and consume everybody's:

Equipment replacement: Budget 15–20% of your hardware cost annually. Cameras fail. Ants colonise them. Elephants investigate them. Floods take them. This is not exceptional — it is the average.
Batteries: A 30-camera grid running 30-day deployments consumes roughly 360 AA batteries per deployment. At $0.50 per lithium AA, that is $180 per season just for batteries. Rechargeable AAs cost more upfront and less over time but require charging infrastructure and spares. The logistical overhead of managing this at a remote field site is real.
Data processing labour: Image classification is the silent budget killer. If you collect 50,000 images per season and each takes 30 seconds to classify, you need 416 hours — roughly 10 weeks — of focused work. That is a person. Budget for them or use automated classifiers (MegaDetector, Wildlife Insights) and accept that you will still need manual validation on 5–10% of detections.
Data storage and backup: A single camera trap deployment can produce 100–500 GB of images and video. Across three seasons, two sites, and raw + processed copies, you are looking at 2–10 TB. Cloud storage costs $5–25/TB/month. Hard drives fail. Budget for redundant, geographically separated backups.
Permits and community agreements: Research permits, park access fees, community consent processes — these are not free and the lead time is measured in months, not weeks. Start the permit process before you buy equipment.

Start Here

A checklist, because monitoring is complicated enough without having to remember what comes next.

Define your question. Not "monitor biodiversity." A specific, quantifiable question: "Is the occupancy of duiker declining by more than 5% per year in the northern block?" Write it down. If you cannot write it in one sentence, the question is not clear enough to design a protocol around.
Choose your indicators. Two to five species. For each, verify that your chosen sampling method can reliably detect it, that it is sensitive to the pressure you care about, and that its response is representative of the community. If you are not sure, spend a pilot season finding out.
Select your method. Match method to question and taxa. Camera traps for terrestrial mammals. Acoustic recorders for bats and birds. Line transects for large mammals in open habitat. eDNA for aquatic systems. Do not default to camera traps because everyone else uses them.
Run a power analysis. With your estimated detection probability and expected effect size, calculate whether your proposed sampling intensity can detect the change you care about. The R package dsop does this for distance sampling. unmarked includes simulation tools for occupancy. If the power analysis says no, increase your sampling intensity or adjust your question before you go to the field.
Pilot for one full season. Test your method, equipment, data pipeline, and team logistics under real field conditions. Analyse the pilot data. If detection rates or species accumulation curves reveal gaps, iterate and pilot again. One pilot season costs less than five years of useless data.
Build the analysis pipeline before data collection starts. Write the scripts. Simulate the data. Confirm that the pipeline produces sensible estimates with known inputs. If you hit a statistical problem during the pilot, you can fix the protocol. If you hit it during year four, you have four years of data you cannot analyse.
Document everything. Write the protocol in a document that someone who was not at the planning meeting can follow. Include: sampling design rationale, equipment specifications and settings, field data collection forms, data management plan, analysis pipeline description, decision pathway with triggers and responsibilities. This document is your programme's memory.
Collect data consistently. Same season, same duration, same equipment configuration, same observer training standards, every year. If something changes, document it in the protocol log on the day it changes. A documented change is a covariate. An undocumented change is a confounding variable that may invalidate your trend.
Analyse and report annually. Do not wait five years to look at the data. Annual analysis catches protocol drift, failing equipment, and unexpected trends while there is still time to respond. An annual report also keeps donors and stakeholders engaged and your team accountable to the data, not just the fieldwork.
Publish to GBIF. Format your data with Darwin Core terms. Write an EML metadata document. Choose a licence. Publish. Your dataset gets a DOI. It becomes part of the global biodiversity evidence base. Someone you will never meet will use your data to answer a question you never thought to ask. This is how science works.

Monitoring is not glamorous. It is repetitive, logistically demanding, and statistically humbling. It will not trend on social media. It will not win awards. But it is the only way to know whether anything we are doing — the protected areas, the restoration projects, the policy interventions, the billions of dollars spent on conservation every year — is actually working. If we are not measuring, we are guessing. And guessing is not good enough for the species we are trying to keep on this planet.

Field Log is a field-first mobile platform built by The Field Company for the work described in this guide. Offline-first. Team sync. Structured forms and rapid logging. It is free to start and your data stays yours.{" "} Get started at fieldlog.thefieldco.com. For more on building conservation technology that respects the people doing the fieldwork, visit{" "} A Letter to Humanity .

Key references: MacKenzie, D.I. et al. (2002, 2017) Occupancy Estimation and Modeling. Buckland, S.T. et al. (2001) Introduction to Distance Sampling. Rovero, F. & Zimmermann, F. (2016) Camera Trapping for Wildlife Research. Ruppert et al. (2019) eDNA metabarcoding review in Global Ecology and Conservation. Field, S.A. et al. (2007) "Making monitoring meaningful" Austral Ecology. Protocols and platforms: SMART, NEON, GBIF, Darwin Core, CyberTracker, ODK.

The Field Co

Open-Source Conservation Technology

Camera Trap Software Compared

Sun, 15 Mar 2026 00:00:00 GMT

import { Image } from "astro:assets"; import BlogImage from "../../components/blog/BlogImage.astro"; import CameraTrap2 from "../../assets/blog/camera-trap/18288499_lindsey_willard.jpg"; import CameraTrap1 from "../../assets/blog/camera-trap/16778606_david_rodrigues.jpg"; import wildlifeTrapShot from "../../assets/blog/camera-trap/28495105_sascha_weber.jpg"; import cameraPipeline from "../../assets/blog/camera-trap/camera-trap-pipeline.svg"; import cameraTrapDeployed from "../../assets/blog/camera-trap/33310474_ali_kazal.jpg"; import camelotLogo from "../../assets/blog/camelot-logo.jpg"; import wildlifeInsightsLogo from "../../assets/blog/wildlife-insights.png"; import megaDetectorLogo from "../../assets/blog/megadetector.png"; import timelapseLogo from "../../assets/blog/TimelapseLogo.png"; import speciesNetLogo from "../../assets/blog/cameratrapai.png";

Camera trap pipeline — from deployment to analysis

Your camera traps just came back from three months in the field. You have{" "} 87,000 images across 24 cameras. Based on the first hundred you spot-checked, maybe 15 contain animals. The rest are wind-triggered grass, heat shimmer, and one very determined spider that built a web across the sensor.

You could spend the next two weeks clicking through blanks. Or you could pick the right tool and be analyzing occupancy models by Thursday.

Here is every major option, compared honestly. Each section names the tool, links to it, and gives you the unvarnished truth about what it does well and where it falls short. No hype. No marketing. Just what works.

Wildlife Insights

Wildlife Insights {" "} is the all-in-one cloud platform. Upload images, get AI species predictions across ~2,000 species (powered by Google's SpeciesNet ensemble), review in-browser, download results. Built by a consortium including WCS, Google, Smithsonian, WWF, and the Zoological Society of London.

What it does well: Best-in-class species coverage — no other tool identifies as many species out of the box. Zero setup. Create an account, upload, and the AI starts working. The geofencing feature actually improves accuracy by ruling out species that do not exist in your region (no, that is not a kangaroo sliding down a waterfall in England). Good sharing and collaboration features for multi-institution projects.

Where it falls short: Cloud only. If you are working from a field station with satellite internet, or your data contains rhino coordinates you legally cannot upload to a US server, this is a non-starter. The free tier is generous for academics, but government agencies and companies pay — pricing is opaque and determined by negotiation. The review interface is functional but slow compared to a native desktop app. You cannot customize the AI model for your ecosystem. No offline mode.

Best for: Organizations that want a zero-install, managed solution. Projects where data sharing is the goal. Teams that do not have sensitive species locations or data sovereignty constraints.

MegaDetector

MegaDetector {" "} is not a platform. It is an AI model built by Microsoft's AI for Good Lab. It detects animals, people, and vehicles in camera trap images. It does not identify species. It is a blank-filtering engine — and it is the best one available. V6 was released in May 2026.

What it does well: Removes{" "} 70–95% of your empty frames before a human looks at anything. V6 is fast — the compact variant is 2.3M parameters and runs on a laptop CPU (~2–5 images/sec). The larger variants hit 83% animal recall and run at 100–200 images/sec on a GPU. Free, open source (MIT), runs offline, and integrates with almost every tool in this list. Three ways to run: Python API (three lines), CLI (one command), or GUI via AddaxAI. Used by 80+ organizations worldwide.

Where it falls short: Detection only — no species ID, no review interface, no data management. Using it directly requires Python or CLI comfort (AddaxAI solves this for non-programmers). Accuracy varies by ecosystem: it sees large mammals in open habitat better than small, distant, or camouflaged animals. Test it on your own data before trusting the numbers.

Best for: The first step of any pipeline. Run MegaDetector before anything else. It compresses your dataset so everything downstream moves faster. If you adopt only one tool from this article, make it this one.

Timelapse (Timelapse2)

Timelapse {" "} is the desktop review workhorse. A Windows application built by Saul Greenberg, professor emeritus of Computer Science at the University of Calgary, over 15+ years. Think of it as a specialized image viewer designed for the specific task of turning camera trap images into structured data. Free and open source.

What it does well: Extreme flexibility. You define your own data fields — species, count, age class, behavior, weather, whatever your protocol demands. Imports MegaDetector JSON output natively: bounding boxes, confidence scores, and species labels appear in the review interface automatically. Excellent efficiency features: persistent zoom and pan across images, bulk data entry, keyboard navigation, time-based grouping of burst sequences. Fully offline. Free. Backed by peer-reviewed HCI research published in Ecology & Evolution (2019) and demonstrated to produce measurable efficiency gains in controlled studies.

Where it falls short: Windows only. This is a significant limitation — no Mac, no Linux. The UI looks its age. Single-user desktop app with no built-in collaboration. No AI inference of its own — you must run MegaDetector separately first. Template setup requires upfront work, though the defaults handle common use cases.

Best for: Researchers who need complete control over their data schema. Offline field work. Teams that want human-in-the-loop review with AI pre-filtering. Anyone who processes tens of thousands of images and needs efficiency.

Camelot

Camelot {" "} takes the database approach. Open source (Eclipse Public License), built by Chris Mann and Heidi Hendry in Sydney. Camelot is a desktop application with a web-based interface that structures your camera trap operation as a proper database: cameras, sites, surveys, sightings — all related. Recommended by WWF in their camera trapping guidelines.

What it does well: Multi-user support out of the box — run it on a local network and multiple people can classify simultaneously. Strong data integrity with metadata validation and relational structure. Reports export directly to PRESENCE and camtrapR formats. Free and open source. 1,842 commits on GitLab — this is not abandonware.

Where it falls short: The web interface adds friction compared to a native app. AI integration is limited — it can import MegaDetector output but the workflow is not as seamless as Timelapse. Development pace appears slow; the last release was version 1.6.16 with no clear public roadmap. Smaller community than Timelapse.

Best for: Team-based projects that need structured, relational camera trap data. Groups already using PRESENCE or camtrapR for analysis. Organizations that want open-source, local-first database management with multi-user support.

David Rodrigues on Pexels`} />

AddaxAI

AddaxAI {" "} (formerly EcoAssist) is AI without code. A desktop GUI that wraps MegaDetector and 20+ regional species classifiers into a point-and-click workflow. Built by Peter van Lunteren at Addax Data Science. Runs on Windows, Mac, and Linux. Free and open source.

What it does well: Zero programming required. Install, point at a folder, and it runs detection and classification in one go. GPU acceleration works on both NVIDIA and Apple Silicon. The regional model library is genuinely impressive: 135-class Australian model, 328-class Sub-Saharan drylands model, 68-class US Southwest model, 84-taxa Neotropical model, and many more. Exports to Timelapse format. Fully offline after installation.

Where it falls short: Regional model accuracy is uneven — some were trained on narrow datasets and will not generalize to your site. You need to test each model on your data before trusting it. The species models are static; you cannot fine-tune them through the GUI. As a wrapper, it depends on the underlying models staying current.

Best for: Field biologists who want AI assistance without learning Python. Projects in regions with a matching species model. Anyone who wants MegaDetector and species ID in one install with a GUI.

Lindsey Willard on Pexels`} />

SpeciesNet

SpeciesNet {" "} is the species classifier behind Wildlife Insights — available standalone. An ensemble of MegaDetector (detection) and Google's SpeciesNet classifier covering 2,000+ species labels. You run it locally via Python ( pip install speciesnet). Apache 2.0 license.

What it does well: Most comprehensive species coverage of any classifier. Geofencing using ISO 3166-1 country codes improves accuracy by constraining predictions to species that actually occur where the camera was deployed. Taxonomic rollup means it predicts at genus or family level when species-level confidence is low — honest and practical. Trained on 65M+ images.

Where it falls short: Command-line and Python only for local use — no GUI. The default ensemble script predicts one species per image (the highest-confidence detection), which is a problem for multi-species images. The alternative workflow (run_md_and_speciesnet) handles this but is separate.

Best for: Users who want the same classifier Wildlife Insights uses but need to run it offline or in their own pipeline. Works as the classifier half of a MegaDetector + SpeciesNet two-stage pipeline.

TrapTagger

TrapTagger was developed by Panthera for camera trap data management with a focus on individual animal identification — stripe and spot pattern matching for tigers and other wildcats. As of June 2026, the website is unreachable and public access appears restricted. It is listed as a MegaDetector adopter in the Microsoft Biodiversity ecosystem but does not appear to be a publicly available tool.

Best for: Panthera collaborators. Not a general-purpose option at this time.

Ali Kazal on Pexels`} />

Side-by-Side

Tool	Offline	AI Species ID	AI Blank Filter	Pricing	Learning Curve	Best For
Wildlife Insights	No	Yes (~2,000 spp)	Yes	Free (academic) / Paid (gov, companies)	Low	Managed cloud platform, data sharing
MegaDetector	Yes	No (detection only)	Yes	Free (MIT)	Medium (Python/CLI)	Blank filtering, pipeline component
Timelapse	Yes	No (imports MD)	Via MD import	Free	Medium	Flexible human review, offline work
Camelot	Yes	No	Limited	Free	Medium	Database-driven team projects
AddaxAI	Yes	Yes (20+ regional)	Yes	Free	Low	GUI AI for non-programmers
SpeciesNet	Yes	Yes (~2,000 spp)	Yes	Free (Apache 2.0)	Medium (Python)	Local species classification
TrapTagger	Unknown	Individual ID	Yes	Unknown	Unknown	Panthera collaborators only

Sascha Weber on Pexels`} />

Which One Should You Use?

Start here: Run MegaDetector on your images first. It is free, works offline, and removes 70–95%{" "} of blanks before you spend a single minute reviewing. You can run it via AddaxAI (GUI, no code), the command line (one command), or Python (three lines). Do this before anything else.

Then, pick your path:

Path A — "I want it simple, offline, and free":

Install AddaxAI{" "} (Windows, Mac, Linux)
Run detection and classification on your images
Export to Timelapse format
Review in Timelapse (Windows only — if Mac or Linux, review in AddaxAI itself)
Export CSV and analyze in R ( camtrapR) or PRESENCE

Path B — "I want maximum species coverage, can work online":

Upload to Wildlife Insights
Let AI classify across 2,000+ species
Review and annotate in-browser
Download data and analyze

Path C — "I have a team and need structured data":

Run MegaDetector on all images
Set up Camelot with your survey, camera, and species structure
Import images and MegaDetector results
Multiple reviewers classify simultaneously
Export to camtrapR or PRESENCE for analysis

Path D — "I need maximum control and flexibility":

Run MegaDetector via Python or CLI
Pair with SpeciesNet for species classification
Review in Timelapse with custom data templates
Export CSV and run custom R or Python analysis

The honest truth: Most camera trap software is free, built by scientists, and maintained on shoestring budgets. The tools work, but none of them feel like polished commercial software — and that is fine. The gap is not between tools. The gap is between what AI promises and what it delivers in your specific ecosystem. Test on your own data. Keep a human in the loop. And run MegaDetector first.

This comparison was researched and written June 2026. Tools change. Check project websites for current status. If something here is wrong or outdated, email us.

Tools referenced:{" "} Wildlife Insights, MegaDetector, Timelapse, Camelot, AddaxAI, SpeciesNet, TrapTagger. Hardware prices from TrailCamPro, June 2026. AI model specs from Microsoft AI for Good Lab and Google Research.

The Field Co

Open-Source Conservation Technology

A Field Guide to GPS Accuracy

Sat, 07 Feb 2026 00:00:00 GMT

import { Image } from "astro:assets"; import BlogImage from "../../components/blog/BlogImage.astro"; import GpsAccurac2 from "../../assets/blog/gps-accuracy/24245275_asad_photo_maldives.jpg"; import GpsAccurac1 from "../../assets/blog/gps-accuracy/13124396_break_media.jpg"; import gpsNavigation from "../../assets/blog/gps-accuracy/7009835_tima_miroshnichenko.jpg"; import gpsHandheld from "../../assets/blog/gps-accuracy/10807595_maël__balland.jpg";

You log a point on your phone. The screen says{" "} accuracy: 3 meters. You drop a pin, label it "CT-47," and move on.

Next field season you are back. Pin CT-47 is{" "} 40 meters from where you are standing. The camera trap is not at the pin. The camera trap is somewhere in dense bush, invisible until you are right on top of it. You spend 25 minutes searching.

This is not a bug. This is how GPS works — and more importantly, how it does not work — in field conditions. The accuracy number your phone displays is an estimate, not a measurement. It assumes clear sky, good satellite geometry, no reflections, and a stationary receiver held flat toward the sky. If any of those assumptions are wrong — and in the field, all of them usually are — your real accuracy can be ten times worse than what the screen says.

This guide explains why, and what you can do about it. No equations. No GIS jargon. Written for someone with mud on their boots.

How GPS Actually Works

GPS satellites orbit about 20,000 kilometers above the Earth. Each satellite broadcasts two things continuously: its exact position and the time the signal was sent, according to its onboard atomic clock.

Your receiver picks up signals from at least four satellites. It knows the signal travels at the speed of light, so it multiplies travel time by the speed of light to get distance. Four distances, four spheres — the point where those spheres intersect is your position. That is trilateration. The fourth satellite is needed to solve for your receiver's clock error, because your phone does not have an atomic clock.

Everything that follows — every source of error, every trick for improvement — comes down to how accurately those distances are measured. The satellites know their position to within a few meters. The clocks are precise to nanoseconds. But the signal has to travel through 20,000 km of space, through the ionosphere and troposphere, and sometimes bounce off things before reaching your receiver. Every one of those steps introduces error.

When everything goes right — clear sky, direct signals, good satellite spread, modern receiver — civilian GPS can place you within about{" "} 3 to 5 meters. When things go wrong — canopy, canyon walls, bad geometry, multipath — errors of 15 to 50 meters are normal.

Why Your Phone Lies About Accuracy

When your phone says "accuracy: 3m," it is not telling you a measured error. It is telling you a statistical estimate based on satellite geometry. The chipset knows where the satellites are, knows how they are spread across the sky, and computes a number called DOP — Dilution of Precision. Low DOP means the satellites are well spread. High DOP means they are clustered together, and small timing errors get magnified into large position errors.

The "3m" estimate assumes every signal arrived via a clean, direct line-of-sight path.{" "} It has no way to know if a signal bounced off a cliff face before reaching you. {" "} Multipath is invisible to the DOP calculation. Your phone can be receiving a perfect geometry score while every single signal is contaminated by reflections — and the accuracy number stays stubbornly at 3m.

It also assumes you are holding the receiver flat, with an unobstructed view of the sky from horizon to horizon. If you are holding your phone vertically (like most people do), the internal antenna is edge-on to the satellites. You are attenuating the signal before it even reaches the chip. And your body — about 70% water — absorbs radio waves at GPS frequencies.

That 3m number? It is a 68% confidence interval under ideal conditions. At 95% confidence, it doubles. Add canopy, add multipath, hold the phone wrong, and real accuracy can be 10–20 meters on a good day. On a bad day, 40 meters.

Mael Balland on Pexels`} />

Dedicated GPS Units vs. Phones

If you spend serious time in the field, a dedicated GPS unit is the single best investment you can make for spatial data quality. The difference is in the hardware, not the satellites — both are listening to the same signals.

	Smartphone	Garmin GPSMAP 66sr	Bad Elf GPS Pro+	Garmin eTrex 32x
Real accuracy (clear sky)	5–15m	2–4m	1–3m (with SBAS)	3–5m
Constellations	GPS + GLONASS + Galileo + BeiDou	GPS + GLONASS + Galileo	GPS + GLONASS + Galileo + BeiDou	GPS + GLONASS
SBAS	Uncertain (model dependent)	WAAS / EGNOS	WAAS / EGNOS / MSAS	WAAS / EGNOS
Battery life (GPS active)	4–6 hours	16+ hours	10+ hours	25 hours (2× AA)
Durability	Fragile, not waterproof	IPX7, rugged	Water resistant	IPX7, rugged
Waypoint averaging	App-dependent	Built in	App-dependent (host device)	Built in
Barometric altimeter	Yes (but GPS vertical is poor)	Yes	No	Yes
Price	$0 (already have one)	{"~"}$450	{"~"}$240	{"~"}$250

The Bad Elf is a Bluetooth GPS receiver that pairs with your phone — it has a professional-grade antenna and chipset but uses your phone for display. It updates at 10Hz (ten times per second) instead of the phone's typical 1Hz, which matters if you are logging tracks.

For most field research, a Garmin eTrex or GPSMAP is worth the money. The accuracy improvement is real, the battery lasts all day, and you can replace AA batteries in the eTrex anywhere in the world. If you can only afford one upgrade for your field kit, make it a dedicated GPS receiver.

Tima Miroshnichenko on Pexels`} />

What Makes GPS Wrong

Multipath — Signals That Bounce

The single biggest source of error you cannot control. GPS signals are radio waves. When they hit a canyon wall, a building, a rock face, or a dense tree trunk, they reflect. The reflection takes a longer path to your receiver than the direct line. Your receiver calculates distance from travel time — it does not know the signal bounced. A reflected signal says "this satellite is farther away than it really is," and your position shifts toward the reflector.

Counterintuitive: Multipath is worse when you are stationary than when you are moving. When you walk, the false solutions from reflected signals fail to converge — only the direct signals produce stable measurements. This is why your track log might look smooth but your stationary waypoint drifts. It is also why the urban canyon effect is so severe: you are usually standing still on a sidewalk, surrounded by reflective surfaces, with half the sky blocked.

Forest Canopy

Leaves and branches attenuate GPS signals. Under dense canopy, the signal can drop below the receiver's noise floor — it simply cannot hear the satellite. You need line-of-sight to at least four satellites for a fix. Under canopy, you might only track two or three. If your receiver locks four but one is marginal, your position might be computed but with terrible geometry. Rainforest and dense evergreen canopy are the worst. Deciduous forest in winter is barely an obstacle. Open savanna is ideal.

Multi-constellation helps here. If GPS satellites are blocked by the canopy above you, GLONASS satellites (higher orbital inclination) or Galileo satellites at different positions may still be visible through gaps. A modern phone tracking four constellations can see 20+ satellites where a GPS-only unit sees 6.

Cold Start — The Airplane Mode Problem

GPS works in airplane mode. GPS is receive-only — your phone does not transmit anything. It just listens. The myth that GPS fails in airplane mode comes from how long it takes to get the first fix.

To calculate your position, your receiver needs to know which satellites are overhead and where they are in their orbits. This data — the almanac and ephemeris — is transmitted by the satellites themselves at a glacial 50 bits per second. Downloading the full almanac from the satellites takes 12.5 minutes.

Your phone normally cheats. When it has a cellular connection, it downloads the almanac and ephemeris from a server over the internet — this is A-GPS, Assisted GPS. It takes one second instead of twelve minutes. That is why your phone gets a GPS lock in seconds when online.

When you have been in airplane mode for days, your phone has no almanac. It must download it from the satellites. 2 to 4 minutes is normal. Up to 12.5 minutes is possible. This feels broken — the GPS appears to do nothing while a progress bar sits there — but it is working. It is just working at 1980s modem speeds.

Field tip: If you know you will need GPS soon, turn off airplane mode briefly to let A-GPS update the almanac, then switch back. Many phones cache almanac data for up to 7 days. Some dedicated GPS units let you preload it. If you are offline and stuck, turn on your GPS receiver 15 minutes before you actually need it and leave it with a clear sky view.

Satellite Geometry

The position of satellites in the sky changes continuously. Sometimes they are well spread — one overhead, one near each horizon — giving a low DOP and good accuracy. Sometimes they are clustered in one part of the sky, giving a high DOP and poor accuracy. GPS satellites repeat their ground tracks roughly every 12 hours. If you got a bad fix at 10am, try again at 10pm — the geometry will be completely different.

Break Media on Pexels`} />

Coordinate Systems Matter

GPS satellites use WGS84 — the World Geodetic System 1984. This is the coordinate system that Google Maps, OpenStreetMap, and virtually all modern GPS receivers use by default. It is defined by the US Department of Defense and continuously refined to be accurate worldwide.

Unless you have a specific reason to do otherwise,{" "} always record your field data in WGS84.

The problem comes when mixing data. Many countries have their own local datums, developed from ground surveys long before GPS existed. The shift between WGS84 and a local datum can be 50 to 200 meters. If you are working in East Africa, your government's protected area boundaries might be in Arc 1960. If you overlay your WGS84 field points on an Arc 1960 boundary without transforming, a point that appears well inside the park could actually be outside it. A report about poaching inside park boundaries becomes incorrect.

This is one of the most common and most invisible errors in field GIS. {" "} It looks right on screen. The numbers are close. Everything aligns — but the underlying coordinate systems are different, and the shift is uniform across all points. You will never notice it unless someone physically verifies a point.

In your metadata, always record:

The coordinate system used (WGS84 is the safe default)
Whether coordinates are in decimal degrees or degrees-minutes-seconds
At minimum 5 decimal places (0.00001° ≈ 1.1 meters at the equator)
The source datum of any external data you are combining with

Asad Photo Maldives on Pexels`} />

How to Get Better Accuracy

Before You Record

Get a clear sky view. Step out from under canopy if possible. You do not need a full horizon, but at least 30° above the horizon should be clear in all directions. The more sky you can see, the more satellites you will track and the better your DOP.

Wait for a good lock. Do not record immediately after the first fix. Watch the accuracy estimate settle. If you can see HDOP (horizontal dilution of precision), aim for below 2. Below 5 is acceptable. Above 5, consider waiting or moving. Smartphone apps that show satellite data include GPS Status (Android) and Gaia GPS.

Hold the receiver flat. For phones, this means screen facing the sky, horizontal. Not vertical. Not at an angle. The internal GPS antenna is designed to receive best when the phone is flat. For dedicated GPS units, the antenna is typically in the top portion — hold it flat or at most 45° from horizontal.

Check your satellite count. More than 8 is good, more than 12 is excellent. Less than 6 is marginal. Many data collection apps show this.

While Recording

Use waypoint averaging. This is the single most impactful thing you can do. Take 30 seconds of readings. Because GPS errors are largely random, averaging cuts the error by the square root of the number of readings. One reading might be 10m off. Thirty readings averaged is about 2m off. Most dedicated GPS units have a built-in averaging function. For phones, some apps support it (Gaia GPS, Avenza Maps). If your app does not, take 10–20 points manually and average them later.

Record metadata. At minimum: HDOP, number of satellites, fix quality, and whether you used averaging. If your app does not record these automatically, write them down. A point without metadata is a guess. A point with HDOP of 8 and 4 satellites tells you "this is approximate, plan to search a larger area."

Take a second point. Walk 10 meters away and record again. The vector between the two points should roughly match your walking direction. If it does not, you have multipath.

What Not to Do

Do not hold the phone against your body. Your body is mostly water and absorbs GPS frequencies. Hold it away from you, at arm's length.

Do not trust a single-point reading taken immediately after unlocking the phone. Give it time to settle.

Do not assume the accuracy number means what it says. If conditions are not ideal — canopy, buildings, canyon walls — assume your actual accuracy is 2–5× worse than the displayed estimate.

SBAS: Free Accuracy from Space

SBAS — Satellite-Based Augmentation System — is a free correction service broadcast from geostationary satellites. Ground stations at known locations measure GPS errors and broadcast corrections. A WAAS-enabled receiver applies these corrections and can improve accuracy from ~5m to{" "} 1–2 meters.

The catch: coverage is regional.

WAAS: United States, Canada, Mexico
EGNOS: Europe, North Africa (Morocco to Egypt), Middle East
MSAS: Japan
GAGAN: India
SouthPAN: Australia, New Zealand (operational since 2022)
A-SBAS: Africa and Indian Ocean —{" "} under development, not yet available

If you are working in sub-Saharan Africa, there is currently no operational SBAS coverage. You are on autonomous GPS. This is important to know because it means your accuracy ceiling is fundamentally higher than colleagues working in North America or Europe. A point recorded in Kenya with a phone is inherently less accurate than a point recorded in Montana with the same phone, all other factors being equal.

Most dedicated GPS receivers (Garmin, Bad Elf) support WAAS and EGNOS. Phone support varies — check your model. Even when supported, SBAS requires a clear line-of-sight to the geostationary satellite, which sits low on the southern horizon in the northern hemisphere and vice versa. If you are in a valley or under canopy, you may not receive the corrections even when they are available.

When You Need Sub-Meter Accuracy

For camera trap placement and most field waypoints,{" "} 2–5 meter accuracy is sufficient. You can find the trap again. Sub-meter accuracy is expensive and complex, and you should only pursue it when your research question demands it.

DGPS (Differential GPS) uses a ground reference station at a known location. The station broadcasts corrections. Accuracy: 1–3 meters. Requires a correction source within a few hundred kilometers. Cost: $500–2,000 for a DGPS-capable receiver. Many marine DGPS networks still operate in coastal areas.

RTK (Real-Time Kinematic) takes carrier-phase measurements of the GPS signal — far more precise than the code-phase timing used by consumer receivers. Accuracy: 1–3 centimeters. Requires a base station (known location) within ~20 km and a rover receiver, plus a radio link between them. Entry-level RTK systems start at $3,000–10,000. Setup complexity is high. Used by surveyors and precision agriculture. Overkill for most field ecology.

PPP (Precise Point Positioning) uses precise satellite orbit and clock corrections from global networks instead of a local base station. Can achieve centimeter-level accuracy, but typically requires 20–60 minutes of continuous observation to converge. Galileo's High Accuracy Service (HAS) provides free PPP corrections via satellite.

Accuracy needed	Equipment	Approximate cost	When to use
5–15m	Smartphone	$0	Reconnaissance, approximate waypoints
2–5m	Dedicated GPS + SBAS	$250–600	Camera traps, vegetation plots, transect markers
1–3m	DGPS receiver	$500–2,000	Permanent plot markers, precise habitat mapping
1–3 cm	RTK system	$3,000–15,000	Drone ground control, sub-centimeter surveys

Field Accuracy Checklist

Get clear sky — at least 30° above horizon all around
Hold the receiver flat, not vertical, not against your body
Wait for the fix to settle — watch HDOP drop below 2 if possible
Check satellite count — more than 8 is good, less than 6 is marginal
Use waypoint averaging — 30 seconds minimum
Record metadata — HDOP, satellite count, fix quality, averaging used
Record in WGS84, decimal degrees, 5+ decimal places
Take a backup point 10m away to check for multipath
Write down conditions — under canopy, in gully, clear sky
If accuracy matters, use a dedicated GPS receiver with SBAS
If you are in sub-Saharan Africa, SBAS is not available — plan accordingly
If you have been in airplane mode, turn off airplane mode briefly to update A-GPS
Assume real accuracy is 2–5× worse than the displayed accuracy estimate
Never trust GPS vertical accuracy from a phone

The accuracy number on your screen is a promise the hardware cannot keep. Treat it as an optimistic estimate at best. When your point matters — when you need to find this location again next season, when the analysis depends on it, when someone else will navigate to your coordinates — invest the 30 seconds to average, check your metadata, and record conditions. Your future self, standing in the bush holding a phone that says "CT-47 is 3 meters ahead," will thank you.

GPS technical reference: GPS.gov, USCG NAVCEN GPS Accuracy . SBAS/WAAS coverage:{" "} FAA WAAS , EGNOS (European Geostationary Navigation Overlay Service) . Device specs:{" "} Garmin GPSMAP 66 series, Bad Elf GPS Pro+. Phone GPS accuracy: van Diggelen (2009) A-GPS: Assisted GPS, GNSS, and SBAS. Multi-constellation: GPS + GLONASS + Galileo + BeiDou.

The Field Co

Open-Source Conservation Technology

A Letter to Humanity

Sun, 01 Feb 2026 00:00:00 GMT

import { TemperatureChart, CO2Chart, OceanHeatChart, EmissionsChart, ArcticIceChart, AntarcticChart, CoralChart, AquifersChart, LPIChart, LPISystemChart, FishChart, ForestChart, NitrogenChart, CompaniesChart, PlasticChart, EwasteChart, InequalityChart, InequalityTimeChart, SolarChart, BatteryChart, ElectricityChart, LitigationChart, } from "@components/blog/charts";

This is not a warning. Warnings are what we received in 1992, when 1,700 scientists signed the{" "} World Scientists’ Warning to Humanity. This is a damage report. A field assessment of what we have done to the only planet we will ever have.

In 2024, our planet crossed{" "} 1.55°C above pre-industrial temperatures {" "} — breaching the threshold we promised we would not cross. It was the warmest year in the entire 175-year observational record.

Six of nine planetary boundaries are now breached. The systems that regulate our climate, our water cycles, our nitrogen flows, our biodiversity — all operating outside the safe zone that sustained civilisation for 10,000 years.

I. The Heat

Our oceans absorbed{" "} {" "} 23 zettajoules{" "} {" "} of excess heat in 2025 alone — equivalent to 39 times all the energy humanity produced that year.

Our wildfires released{" "} {" "} 8 billion tonnes of CO₂ {" "} {" "} in the 2024–2025 season. Canada alone burned 15 million hectares in 2023 — nearly four times more carbon than global aviation.

II. The Ice

An ice-free Arctic summer could arrive as early as{" "} {" "} 2027{" "} {" "} . It is locked in regardless of what we do next.

Antarctic ice loss has{" "} {" "} quadrupled in three decades{" "} {" "} — from 48 billion tonnes per year in the 1980s to 202 billion tonnes per year in the 2010s.

Beneath the permafrost lies 1,500 gigatonnes of carbon{" "} — twice what is currently in the entire atmosphere. Four core climate tipping points trigger at 1.5°C. We are already at 1.55°C.

III. The Water

The fourth global coral bleaching event in 2023–2024 impacted{" "} 84% of the world’s coral reefs {" "} . 25% of all marine species depend on coral reefs for survival.

71% of global aquifers are in decline {" "} . We are drawing down 324 billion cubic metres of freshwater more than is replenished every year.

We have created{" "} {" "} 405 dead zones{" "} {" "} in our oceans.

IV. The Living World

Since 1970, monitored wildlife populations have collapsed by{" "} {" "} 73%{" "} {" "} . Not a decline. A collapse.

37.7% of fish stocks are now overfished {" "} — up from 10% in 1974.

In 2024, we lost another{" "} {" "} 6.7 million hectares{" "} {" "} of pristine rainforest.

V. The Soil Beneath Our Feet

{" "} 75%{" "} {" "} of Earth’s land surface became permanently drier in the last three decades. We are losing 24 to 75 billion tonnes of soil annually.

VI. How We Are Doing This

Fossil Fuels

Just{" "} {" "} 32 companies{" "} {" "} produced over 50% of global fossil CO₂ emissions in 2024.

The Things We Throw Away

The Great Pacific Garbage Patch contains{" "} {" "} 3.6 trillion pieces of plastic {" "} {" "} .

We generated{" "} {" "} 62 million tonnes{" "} {" "} of electronic waste in 2022. Only 22.3% is recycled.

VII. Who Is Responsible

The richest 1% of humanity exhausted their fair share of the 1.5°C carbon budget in{" "} {" "} 10.2 days of 2026{" "} {" "} .

The poorest 50% — responsible for 12% of emissions — are exposed to 74% of income losses from climate change. This is architecture.

VIII. What Is Still Possible

This letter is not a eulogy. Not yet.

Solar energy is now{" "} {" "} 41% cheaper{" "} {" "} than fossil alternatives. Battery storage costs have fallen{" "} {" "} 93%{" "} {" "} in a decade.

There are{" "} {" "} 2,967 climate cases{" "} {" "} filed across 55 jurisdictions.

The technology exists. The legal frameworks are emerging. Two trillion dollars was invested in clean energy in 2024. The solutions are deployed, proven, and cheaper.

And yet — current commitments target 52 to 58 gigatonnes in 2030. Keeping 1.5°C requires 25 to 30. We are aiming for double what the planet can absorb.

IX. The Question

We are the first generation to fully understand what we are doing to our planet. If we do not act, the data says we will be the last to have had the chance.

So the question is not whether we can afford to act. It is whether we can afford not to.

This letter was compiled from 24 research packets containing data from the World Meteorological Organization, NASA, NOAA, the IPCC, UNEP, the FAO, the WWF, Nature, Science, and dozens of peer-reviewed sources. Every figure is cited. Every trend is verifiable.

The Field Co

Open-Source Conservation Technology

Cape Town, South Africa · February 2026