• A bear stands in a creek with a fish in its mouth

    A collaboration between wildlife ecologists at the University of Victoria and Silicon Valley software developers has led to the creation of an AI-powered software that can identify the faces of individual brown bears. (Photo: Wilma McKenzie/Can Geo Photo Club)

A collaboration between wildlife ecologists at the University of Victoria and Silicon Valley software developers has led to the creation of an AI-powered software that can identify the faces of individual brown bears. 

The idea came to be when Melanie Clapham, a conservation biologist based in Knight Inlet, B.C. studying grizzly bears, noticed she had developed the ability to recognize individual bears. 

In Silicon Valley, software developer Ed Miller noticed the same thing. 

“My whole career has been just typical tech-oriented. But in my free time, I’m an avid outdoors person,” Miller says. “That’s always been an important part of my life.” 

While watching the live webcam broadcast from Katmai National Park, Alaska — the same park that hosts the Fat Bear Week contest — he was able to tell which bear was which. At that point, the idea was just a matter of teaching an artificial intelligence software to do the same thing. 

Clapham and Miller’s idea took flight when they met on WildLabs.net, an online platform for interdisciplinary collaboration. Along with software engineer Mary Nguyen and conservation scientist Chris Darimont, they created the technology they’re calling BearID. 

The Deep Learning software — a type of machine learning that uses examples — was trained and tested on 4,674 photos of 132 brown bears in British Columbia and Alaska. 

The researchers’ analysis found the software has almost 84 per cent accuracy rate at classifying each bear — that is, it knew which bear it was 84 per cent of the time. Its accuracy rate for locating the face on the bear was higher — 98 per cent. 

The AI decides what data points are relevant to its identification of each bear, such as the distance between different points on the bear’s face, as opposed to the researchers telling it what to do. 

“It actually picks that up for itself. That's probably also similar to what our human brain is using,” Clapham says. “We’re just probably not aware of it.” 

Clapham says the technology can be used in replacement of marking or GPS collars — which she said can be expensive — to track animals for conservation purposes. 

“We don't have to capture them and put collars on them, or interact with them at all. We can do this in a non-invasive way,” she says. 

Clapham said in addition to behavioural research, the technology can be used for population estimates for conservation, by combining it with camera trap photography. 

She says it can also be used to resolve human-on-wildlife conflict, such as bears that get into people’s trash. 

“Is this just one bear that’s perhaps getting into somebody’s garbage and causing trouble? Or is this a number of different bears?” she says. “Management decisions may differ (if) it’s just one bear or if it’s 10 different bears. It gives us more understanding of what’s going on in the population or with specific [bears].” 

Clapham says the fact that brown bears don’t have distinguishing markings, and gain significant amounts of weight around winter, makes it more difficult for the AI to learn how to identify bears, in comparison to other species.  

“They don’t have stripes like tigers, or a human fingerprint for example, that can be used to individually identify them,” Clapham says. 

Rinjan Shrestha, a World Wildlife Fund conservation biologist, says the technology is a “significant step forward” for the study of the behaviour and ecology of brown bears. 

Shrestha specializes in Asian big cats, such as snow leopards, and has done most of his work in Nepal and Bhutan. He says reading about the study was exciting for him because it gives him hope that a similar technology could be developed for the species he works with. Snow leopards, unlike brown bears, have distinguishing markings, which eliminates one hurdle the B.C.-Silicon Valley team faced.  

He says one area of improvement could be to work on providing the software with far more examples to learn from, despite the 84 per cent accuracy rate being “impressive.”

The team is currently working on collecting more photos and examples for the AI to learn from to try to increase its classification accuracy rate, says Clapham. 

Shrestha says the need for more photos could be an opportunity to engage citizen scientists, by inviting local people to take photographs and participate in the research on wildlife in their area.