fruitpunch

Guide · AI & data

AI for Bear Conservation: Identifying Bears With Machine Learning

By the FruitPunch research deskUpdated November 15, 20269 min read

In short

Machine learning can now tell individual brown bears apart from photos: the open source BearID model reached 83.9% identification accuracy on 4,674 images of 132 bears (Ecology and Evolution, 2020), and the 2026 PoseSwin model was trained on over 72,000 photos of 109 bears. The same tools feed camera trap alert systems built to cut human bear conflict, which hurt 238 people in Japan in fiscal 2025 and killed 14 people in Romania over five years.

This address once hosted FruitPunch AI, an Eindhoven AI for good community. FruitPunch is a new, independent publication and is not that organization. This guide was written from public sources; the history page explains the change.

AI helps bear conservation in two practical ways: it recognizes individual bears in photos, which bears' lack of spots or stripes made almost impossible to automate, and it spots bears in camera trap images fast enough to warn people nearby. The best known example is BearID, an open source face recognition tool for brown bears. In its 2020 paper in Ecology and Evolution, the team trained and tested it on 4,674 images of 132 known bears from British Columbia and Alaska and reached a face detection average precision of 0.98 and an individual identification accuracy of 83.9%. Since then, researchers have pushed the idea toward small field computers and real time conflict alerts.

Key facts

  • BearID identified individual brown bears with 83.9% accuracy on 4,674 images of 132 bears (Clapham et al., Ecology and Evolution 2020).
  • The 2026 PoseSwin model was built from over 72,000 photos of 109 brown bears taken between 2017 and 2022 (EPFL).
  • On an NXP i.MX93 neural processing unit, bear models ran 5 to 19 times faster than on the CPU (Arm).
  • A bear repeller built for the Romanian Carpathians draws 0.06 watts on standby (Q42 Engineering).
  • Japan recorded 238 people injured or killed by bears in fiscal 2025, including 13 deaths, a provisional figure (Ministry of the Environment, Japan).
  • Romania estimates its bear population at about 11,500, and over five years bear attacks there caused 14 deaths (Council of the EU).

Why is it hard to identify individual bears?

Most automated photo identification works on animals with natural barcodes: the stripes of a zebra, the spots of a leopard, the notches on a whale fluke. Brown bears have none of these. The BearID authors describe them as a representative species among wildlife that "lack unique markings", and an EPFL summary of newer work adds a second problem: bears come out of hibernation thin and shaggy, then gain a lot of weight on salmon and shed their winter coat, which can confuse both experts and algorithms.

Without individual IDs, camera trap studies can count detections but not animals. The BearID Project notes that methods exist to correct for missing IDs in population surveys with cameras, "but concerns over reliability remain." Knowing which bear is which opens up standard tools such as mark and recapture, which the 2020 paper names as a cost effective route to population monitoring.

How does BearID recognize a bear's face?

BearID borrows the structure of human face recognition and runs it as a chain of four steps, all described in the Ecology and Evolution paper:

  1. Face detection. A detector finds the bear's face in the photo.
  2. Landmarks and alignment. The tool locates the eyes and nose, then rotates and crops the face into a standard "face chip".
  3. Embedding. A ResNet 34 network turns each face chip into a 128 dimensional embedding, a list of numbers where photos of the same bear sit close together.
  4. Classification. A support vector machine assigns the embedding to a known individual.

The team built the tool with the open source Dlib toolkit. A Raspberry Pi profile of the project explains that a Dlib example trained on dog faces gave the team a head start on labeling bear faces, because dogs have similar facial features. Training images came from two long running bear viewing sites: Knight Inlet in British Columbia and Brooks River in Katmai National Park, Alaska. Katmai alone held an estimated 2,183 ± 379 brown bears in a 2004 and 2005 National Park Service survey.

Accuracy was not uniform. In the test set, the classifier made 782 correct predictions out of 932 embeddings, and the paper's confusion matrix shows that some individuals were consistently identified while others were often confused. The authors suggest future work should test how age and body size affect accuracy.

What came after BearID?

In January 2026, researchers from EPFL and Alaska Pacific University published PoseSwin in Current Biology. According to the EPFL release, the model uses a transformer trained with metric learning and focuses on head features that change little over a season, such as muzzle shape, brow bone angle and ear placement, plus the pose of the head. That let the team use photos where the face is not shown straight on. Training data came from McNeil River State Game Sanctuary, where nearly 150 bears pass through each summer. The team then tested the model on visitor photos from Katmai and released its code and data on GitHub.

The two approaches differ mainly in what they ask of the photo:

Model Published Training data Method Reported result Code
BearID Ecology and Evolution, 2020 4,674 images, 132 bears, British Columbia and Alaska Face detector, landmarks, ResNet 34 embedding, SVM 0.98 detection precision, 83.9% ID accuracy hypraptive/bearid
PoseSwin Current Biology, 2026 Over 72,000 photos, 109 bears, McNeil River Pose aware transformer, metric learning Outperformed body only and pose blind models (EPFL) amathislab/BrownBear_ReID

How do camera traps and AI work together for bears?

Camera traps fire on motion, so they collect far more images than people can sort by hand. The usual pipeline first runs a general animal detector, then a species classifier, and only then individual identification. MegaDetector is the common first step: when the Hack the Planet team built a bear system in Romania, its engineers ran about a terabyte of the partner's photos and videos through MegaDetector to draw boxes around animals, then trained a custom YOLOv5 bear detector on that dataset.

The BearID 2020 paper argues that combining face recognition with camera trapping "could facilitate fine‐scale behavioral research such as individual spatiotemporal activity patterns." The catch is delay. A Microsoft write up on a BearID prototype notes that even with AI processing, "it still takes weeks to months to determine which bears were seen", mainly because someone has to retrieve memory cards from the field, while streaming video to the cloud needs too much bandwidth and power for remote sites.

Can bear AI run on low power hardware at the edge?

Running models on the camera itself solves the bandwidth problem, because the device only has to send a short message ("bear M025 seen at 06:12") instead of video. This is the direction BearID has taken. In the Raspberry Pi profile, the project's software lead describes a Raspberry Pi 4B version that reads a memory card in the field and runs the analysis locally, and says the end goal is for the camera to detect and identify the bear and send alerts or movement data "by some low power transmission protocol to a base station."

In April 2024, the BearID Project ran an AI for Bears Challenge with FruitPunch AI, the former operator of this domain and an organization this site is not affiliated with. The challenge was sponsored by Arm, with Hack the Planet and NXP as partners. According to BearID, 19 engineers spent more than 1,500 hours over 10 weeks in four groups: bear classification, face detection and segmentation, bear identification, and an ML pipeline on edge hardware. Arm's write up reports that the edge group demonstrated classification, detection and segmentation on the NXP i.MX93 and "showed 5-19 times speedup over CPU-based inference", with power savings per inference, though it does not publish absolute power or accuracy figures.

Power budgets matter as much as model size. For the Romanian deterrent system, Q42's hardware partner cut the repeller's standby draw to 0.06 watts, linked the repellers over LoRaWAN because a 4G modem running around the clock "would draw far too much power", and measured a LoRa range of 5 to 7 km in a field test.

How does AI help prevent human bear conflict?

Conflict is where bear AI meets public safety. Two official data sets show the scale:

Country Period People harmed Deaths Source
Japan Fiscal 2023 (April 2023 to March 2024) 219 6 Ministry of the Environment
Japan Fiscal 2025 (April 2025 to March 2026) 238 13 Ministry of the Environment
Japan Fiscal 2026, through October 76 6 Ministry of the Environment
Romania Past five years More than 150 seriously injured 14 Council of the EU
Slovakia Past five years 64 seriously injured 4 Council of the EU

Japan's figures are provisional numbers collected from prefectures and dated November 10, 2026. Most of the fiscal 2025 cases involved Asian black bears (232 people); brown bears on Hokkaido accounted for 6 people and 2 deaths. The Romanian and Slovak figures come from a June 2026 note the two governments sent to EU agriculture ministers, which also reports more than 30,000 emergency calls about bears in Romania over five years.

AI systems for conflict follow a simple loop: a 4G camera trap takes a photo, a model checks it for a bear, and the system either alerts people or triggers a deterrent. In Romania, the Hack the Planet system for the village of Rucăr sends camera photos to a cloud detector and, on a bear detection, activates a repeller that plays loud sound and strobes a light. The team assembled 14 repellers for the June 2023 rollout. A separate University of Cambridge project with Fundația Conservation Carpathia trained models on six years of camera trap data and runs an alert system for bears and wild boars in three locations, describing it as "the first known study to use remote processing of 4G-enabled camera trap images to operate a human-wildlife conflict alert system."

Individual identification adds something detection alone cannot: the 2020 BearID paper lists identifying problem individuals in conflicts as a practical use, so managers could act on one habituated bear instead of a whole area.

How many bears are there?

Brown bears are listed as Least Concern, with a global estimate that "exceeds 200,000 individuals", according to the San Diego Zoo Wildlife Alliance fact sheet citing the IUCN assessment by McLellan and colleagues. The same sheet lists more than 100,000 in Russia, 33,000 in the U.S., 25,000 in Canada, 15,400 in Europe and 2,200 or more on Hokkaido. Romania now estimates about 11,500 bears and Slovakia at least 2,500, which the two governments say is more than half of Europe's bears.

Other species are in worse shape. An IUCN assessment found six of the world's eight bear species threatened with extinction, with only the American black bear secure across its range. That is one reason BearID has started work on Andean bears and in 2024 helped camera trap researchers at Universidad San Francisco de Quito in Ecuador.

Which bear AI projects and tools can people use?

Frequently asked questions

Can AI really recognize individual bears?

Yes, with limits. BearID reached 83.9% accuracy across 132 bears, and accuracy varied by individual. The 2026 PoseSwin model outperformed models that ignored pose and can flag bears it has never seen. Both work best with clear, close photos.

What was the AI for Bears Challenge?

It was a 10 week challenge in 2024 hosted by the BearID Project with FruitPunch AI, sponsored by Arm with Hack the Planet and NXP. Its goals were an efficient classifier for a bear deterrent and models for a smart camera. This site is a separate, unaffiliated publication that now uses the fruitpunch.ai domain.

Why run bear models on the camera instead of the cloud?

Remote cameras often have weak signal and limited power. Streaming video needs considerable bandwidth and device power, while an on-device model only sends a short result. NPUs like the one on the NXP i.MX93 gave a 5 to 19 times speedup over CPU inference.

Do AI bear deterrents work?

Field evidence is still early. The Cambridge project reports that preliminary results from its deployment show a positive conservation impact, but no large controlled trial of AI bear deterrents has been published.

How many people do bears injure in Japan?

Provisional Ministry of the Environment data show 238 people harmed in fiscal 2025, 13 of them killed, the highest totals in the series since fiscal 2019. Through October 2026, the count for fiscal 2026 stood at 76 people and 6 deaths.

Spotted an error in this guide? Email editorial@fruitpunch.ai

Keep reading