AI can identify individual brown bears, opening new paths for wildlife research

Scientists from Alaska Pacific University (APU) and École Polytechnique Fédérale de Lausanne (EPFL) have developed a new artificial intelligence system that can reliably recognize individual brown bears in the wild, even as their appearance changes dramatically over the course of the summer. The advance opens powerful new possibilities for wildlife research, conservation, and management, particularly in remote and hard-to-study environments like Alaska.

“Differences between individuals is a foundational component in ecology,” said Beth Rosenberg, a researcher at the Fisheries, Aquatic Science, and Technology Laboratory (FAST Lab) at APU and a former staff member at McNeil River State Game Sanctuary, where the study began. “If you can’t reliably tell who is who within a species, you can’t truly understand behavior, movement patterns, survival, or population health.”

Being able to distinguish individual animals over time allows scientists to better understand movement patterns, behavior, population dynamics, and habitat needs. While existing computer vision systems can identify animals with distinctive markings, such as zebras, leopards, or giraffes, unmarked species like brown bears have remained a major challenge. Brown bears not only lack obvious visual identifiers, but also undergo extreme seasonal transformations: they emerge from hibernation lean and shaggy in spring, then rapidly gain weight and shed their winter coats as they feed on salmon throughout the summer.

Despite these challenges, the research team developed an AI program, PoseSwin, capable of recognizing individual bears from photographs alone, without capturing or tagging the animals.

Learning to see bears like biologists do

The work is grounded in years of field research at Alaska’s McNeil River State Game Sanctuary, home to the world’s largest seasonal gathering of brown bears. Nearly 150 bears traverse more than 500 square kilometers of protected land each summer, congregating on sedge meadows and salmon-rich waterfalls.

For four months each year, Rosenberg lives and works in the sanctuary. Between 2017 and 2022, she built an extraordinary dataset of more than 72,000 photographs of 109 individual bears, captured across seasons, lighting conditions, weather, and behaviors. “This dataset reflects how bears actually exist in the wild, so not posed or controlled, and instead, constantly changing,” says Rosenberg.

To analyze this data, the team developed an AI system called PoseSwin, designed to focus on anatomical features that remain relatively stable over time. Drawing on biological insight, the researchers trained the model to prioritize the shape of the muzzle, the angle of the brow bone, ear placement, and—crucially—the bear’s head pose.

“Body shape is unreliable because it changes so much with weight gain,” explains Alexander Mathis, professor at EPFL’s Brain Mind Institute and Neuro-X Institute. “Our intuition was that head features combined with pose would be far more stable and the data confirmed it.” By incorporating pose awareness, PoseSwin can successfully identify bears from a wide range of angles, including imperfect or partially obscured images that would normally be unusable.

Capturing identity, not just appearance

PoseSwin is built on transformer architecture, the same family of AI models that power large language models, adapted specifically for image analysis. Using a technique called metric learning, the system learns relationships between images by comparing groups of three photographs: two of the same bear taken at different times, and one of a different individual.

Over time, the algorithm learns to cluster images of the same bear together in a multidimensional space, while separating them from others. The result is a digital representation of each bear that captures something deeper than surface appearance.

“Each bear becomes its own constellation of points,” says Mathis. “That suggests the model is capturing something closer to identity, not just how the bear looks in a single moment.” Importantly, PoseSwin can also detect bears it has never seen before, which is an essential capability for open, unenclosed ecosystems where new individuals regularly appear.

From sanctuary to citizen science

To test the system beyond McNeil River, the team turned to photographs taken by visitors to nearby Katmai National Park and Preserve. Despite differences in photographers, equipment, and conditions, PoseSwin successfully identified several known bears, revealing seasonal movement patterns across the broader landscape.

“This shows the real potential of the technology,” says Rosenberg. “Visitor photos could help us understand how bears use this enormous region. I can tell us where they go, when they move, and what resources they rely on.”

Such insights are critical for long-term conservation planning and understanding how apex predators support healthy ecosystems.

A tool for wildlife research worldwide

While brown bears were intentionally chosen as one of the most difficult test cases, PoseSwin has already demonstrated strong performance on other species, including macaques. The researchers believe the system can be adapted to a wide range of animals, from small mammals to primates. “Bears are a complicated version of a mouse,” Mathis says. “If you can solve bears, many other species become much easier.”

The team has released PoseSwin and its accompanying datasets as open-source resources, allowing researchers and wildlife managers around the world to adapt the system to their own study species.

Looking ahead, the scientists plan to expand the system with additional data from other seasons and locations, with the long-term goal of automating population monitoring over time—without disturbing the animals being studied. Ongoing projects are already applying PoseSwin to bear populations in western Alaska and Montana, with plans to expand the dataset through 2026 and move toward automated, continuous monitoring. With this knowledge comes greater understanding of the needs of bears as habitat changes with climate or human development. This is of particular importance as current political forces move to develop remote regions of Alaska.

“Bears sit at the top of their ecosystems,” Rosenberg said. “Understanding their life cycle, movement patterns, and behaviors doesn’t just help bears — it helps us understand entire landscapes and systems.”