New A.I. Offers Facial Recognition for Grizzly Bears
The open-source software could help conservation scientists keep track of individual animals over years
/https://tf-cmsv2-smithsonianmag-media.s3.amazonaws.com/filer/92/43/92433590-617f-4e6c-a798-7458fae2843c/2020_nov11_grizzlybearai.jpg)
Grizzly bears have domed shoulders, tall foreheads, and pale-tipped fur that gives them their grizzled appearance. If you’re comparing two bears, one might be lighter or darker in color, or fatter for hibernation. But for the most part, there’s no universal, unique marker a person can use to tell two bears apart.
This issue is a challenge for scientists like University of Victoria wildlife conservationist Melanie Clapham, whose research on grizzly bear behavior requires her to monitor individual bears over years, Adam van der Zwan reports for CBC. But now, Clapham and her research team have developed a solution: facial recognition for bears.
Bears grow and shrink a lot depending on the season, and their appearance changes frequently during their 20- to 25-year-long lifespans. Clapham began to wonder if A.I. might be able to solve her problem the same way the technology recognize people’s faces. Luckily, software developers named Ed Miller and Mary Nguyen were wondering about this exact problem at the same time, Lesley Evans Ogden reports for the New York Times.
In 2017, Miller, Nguyen, Clapham and University of Victoria conservation scientist Chris Darimont connected on Wildlabs.net, which organizes connections between engineers and conservationists. Over several years, they developed a machine learning algorithm to identify individual bears. The paper describing their open-source software published in the journal Ecology and Evolution on November 6.
“Learning about individual animals and their life stories can have really positive effects on public engagement and really help with conservation efforts,” says Clapham, first author on the paper, to the Vancouver Sun’s Randy Shore.
Some national park programs have already found success in identifying specific bears for the public to rally behind. Every autumn, Katmai National Park in Alaska hosts Fat Bear Week, where fans of the park’s bear cams can vote for the chubbiest bear out of a cast of cubs with names like Chunk, Holly and this year’s winner, 747. In Canada’s Banff National Park, bears like Split Lip and The Boss have followings of their own, per the Vancouver Sun.
Unfortunately, some fans try to seek out their favorite bears in person, which puts both people and bears at risk. Clapham hopes that programs like the bear-recognizing artificial intelligence, dubbed BearID, will help reduce direct interaction between people and wildlife. For instance, scientists could use BearID with camera trap images to track a bear’s movement, instead of capturing and tagging an individual.
“Fifteen years ago when we started doing land use planning, there was just one provincial bear health expert for the whole province,” says Kikaxklalagee / Dallas Smith, a member of the Tlowitsis Nation and president of Nanwakolas Council, to the New York Times. With limited resources, it was difficult to understand the health of bears in their territory. But Smith says the introduction of technology like BearID could support their stewardship of local bears.
“We’re trying to make it a sustainable, limited footprint operation,” says Kikaxklalagee / Dallas Smith says.
To train the algorithm, the developers submitted over 3,000 identified bear images to the algorithm for it to study, learning to identify not just a bear in an image, but also remembering which bear it was. Then, they asked the program to spot differences between bears in 935 more photographs. It had an accuracy rate of 84 percent, and tended to mix up the same bears that Clapham also does sometimes, she tells the Times.
Without patterns of spots or stripes to differentiate between bears, the A.I. had to use other characteristics like the proximity between its eyes, nose, ears and forehead top to match a bear’s face with a name. But unintended biases in the training dataset—the first 3,000 images—can sometimes introduce mistakes in the results.
“It’s basically a black box. You don’t know what it’s doing,” says Fraunhofer Institute for Digital Media Technology research engineer Alexander Loos to the New York Times. For example, if all of the photographs submitted of one bear are taken in a bright environment, then the program might learn to take sunlight into account when categorizing pictures later.
The research team hopes that systems like BearID could be used on other bear species, like sloth bears, sun bears and Asiatic bears, or even caribou or wolves.
“The challenge is that you would need a few photos of 50 to 100 known individuals of a species,” Clapham tells the Vancouver Sun.