How Swiss AI recognises wild animals – and helps biologists worldwide

Strong noses: the AI model can identify individual bears based on details such as their snout. Beth Rosenberg

From identifying bears in Alaska to tracking deer in the Alps, Swiss-built AI is opening new frontiers in wildlife conservation. But machines are no miracle workers: they rely on researchers who have spent a lifetime in nature.

Beth Rosenberg has spent the past 20 years observing bears in remote areas of Alaska, without electricity or running water. The biologist and ecologist, who works at the Alaska Pacific University, can now recognise individual animals by the shape of their heads and muzzles, as well as by small scars or distinctive behaviours.

“Some bears always fish in a certain way, or like to play with each other. If you spend enough time observing them, individual differences quickly become obvious,” she says.

Now, Rosenberg is sharing her experience with an artificial intelligence (AI) model. In collaboration with researchers at the Swiss Federal Institute of Technology Lausanne (EPFL), Rosenberg helped to train an AI system on how to recognise individual bears. The training relies on Rosenberg’s long experience and thousands of photos she collected over six years along the McNeil River in Alaska, where hundreds of brown bears converge every year to catch salmon migrating upstream.

Studying these predators is vital for understanding the health of their ecosystem and how species respond to climate change. But doing so non-invasively in remote environments is challenging. This is where AI can make a difference. Potentially, it could open up “enormous opportunities to better understand wildlife and ecosystems, supporting conservation and management,” Rosenberg says.

Recognisable snouts

The AI model developed with EPFL can reliably identify bears along the McNeil River based on physical features such as head shape or profile. It can track their movements over time and across space by linking images of the same individual animal taken at different times and locations and can even flag new bears it has never seen before – a result Rosenberg describes as unprecedented.

By analysing a large number of images, AI also helps researchers to identify recurring patterns – such as where bears feed, rest or move – and to study their behaviour.

“This helps us better understand bears, their population dynamics, and answer many important ecological questions,” the researcher explains.

A bear fishing in the McNeil River, Alaska. Beth Rosenberg

Reaching this point, however, was far from easy.

Unlike zebras or leopards, bears do not have easily distinguishable patterns. This makes them particularly difficult to identify for computer vision systems – AI technologies which analyse and interpret images and videos in a way which mimics human vision. Bears also look dramatically different at various times of the year: before hibernation, they can gain over 100 kilograms, and their fur changes completely between summer and winter.

“That’s enough to confuse even the most trained eyes,” Rosenberg says.

To train the system, Rosenberg manually curated around 73,000 images of over 100 bears, taken in the rain, at different times of day, and from multiple angles. The team focused mainly on the animals’ heads, identifying features that remain quite stable over time, like muzzle shape, brow structure, ear placement and profile.

“AI systems are very good at recognising recurring patterns,” Rosenberg explains. “But it’s much harder when dealing with species that lack obvious distinguishing marks.”

Preparing the dataset took years. In 2018, Rosenberg contacted Alexander Mathis, a professor at EPFL’s Brain Mind Institute, to propose a collaboration.

“Alexander and I spent much of the Covid period in front of our computers, coding non-stop,” she says.

According to Rosenberg, the eight years needed to build the dataset were just enough for the model to learn to identify individuals – a timeframe which “says a lot about the complexity of the human brain”, she says.

The next step will be to test the system on a larger scale, in other regions and with animals the system has never seen before. 

From Alaska’s forests to the Swiss Alps

Rosenberg believes the AI model could eventually be applied to other species with pronounced snouts, such as wolves and deer, including those found in the Swiss Alps.

However, transferring models from one environment to another is a major challenge.

“It’s unlikely that a model trained in Alaska will work well in Switzerland,” says Devis Tuia, who heads the Environmental Computational Science and Earth Observation Laboratory at EPFL.

Tuia’s team has developed AI models capable of recognising Alpine wildlife and analysing their behaviour using images and videos collected by camera traps. But achieving this required extensive manual work, including labelling thousands of frames and fine-tuning the model’s parameters.

“AI is not magic,” Tuia says. “A model is only as good as the data it is trained on – and that depends on the time and effort people put into it.”

Watch how AI manages to recognise two deer in canton Graubünden in this EPFL video:

Improving coexistence between humans and wildlife

Despite the limitations, Tuia believes the technology has enormous potential. AI systems can automatically analyse millions of images and videos in seconds, helping biologists and rangers to better understand how ecosystems function and how they’re changing, which would then lead to more informed decisions about conservation.

These insights could also improve coexistence between humans and wildlife. “If we understand how animals behave, we can, for example, plan hiking trails in safer and more appropriate ways,” Tuia explains.

In places like Switzerland, where debates about wolf management – particularly attacks on livestock – are increasingly heated, such tools could prove valuable.

In the future, these models could be applied on a much larger scale and in many more areas through citizen science initiatives, where members of the public upload their observations to shared platforms such as iNaturalist.

“We are building a global dataset of the planet, one photo at a time,” says Tuia.

Rosenberg also sees great potential in this approach. This summer, she and her colleagues will launch a platform to collect images of brown bears taken by people around the world.

“We can use people’s photographs to create maps and understand where animals move on a much larger scale,” she says. “The opportunities for learning – without invading their space – are enormous.”

Join the debate:

Edited by Gabe Bullard/VdV,dos