Koustubh Sharma is what you could call a cat scientist with a daunting task, as a wildlife biologist studying one of the world’s most magnificent, fluffy-tailed and elusive big cats: snow leopards.
Based in Kyrgyzstan, Sharma spends a lot of time trying to solve the riddle of how to study the hard-to-study, threatened species. The alpine cats live in frigid, barren landscapes; roam hundreds of miles and are so adept at solitude that they’re dubbed “ghosts of the mountain.”
In the nearly 11 years that Sharma has studied snow leopards in the highlands of Central Asia, he has seen the thick-furred, rosette-marked feline only twice. His one close encounter was with a large male with a scarred face in southern Mongolia, while standing on a mountain ledge near a freshly killed ibex, a favorite meal for snow leopards.
“Suddenly, I hear a chuff and a snow leopard walks right in front of me. He looked at me with big round eyes, almost as if saying, ‘How did you get so close to me without me knowing it?’ They’re almost overconfident about their invisibility. Then he turned around and fled away like a ribbon,” says Sharma, a senior regional ecologist at the Snow Leopard Trust, a Seattle-based conservation nonprofit that works in China, Kyrgyzstan, India, Mongolia and Pakistan.
The cats’ covert nature hasn’t deterred Sharma and his team from studying them with camera traps, which helps the nonprofit understand and protect the species from poaching, mining, climate change and other threats. Snow leopards have dwindled to an estimated population of 4,000 to 6,000, making them so rare that remote cameras are one of the only feasible ways to study them.
The cameras, equipped with heat and motion sensors, have their share of issues. They are sometimes buried in avalanches, washed away in floods or knocked over by animals. Other times, they take photos of the “wrong” animals, like goats, camels or horses that plant themselves in front of a lens and ruminate for hours, or walk by in hordes. Or the cameras are triggered by a blade of grass swaying near a sun-heated rock.
Then there is the overwhelming volume of photos. Each camera survey lasts several months, covering a 400- to 500-square-mile area and generating 200,000 to 300,000 images from 30 to 60 cameras. Sorting the images – into photos with snow leopards and photos without – has traditionally been a tedious, manual task taking hundreds of human hours for the nonprofit.
A new Microsoft AI solution is accelerating the process, with a machine learning model that can identify snow leopards and automatically classify hundreds of thousands of photos in a matter of minutes.
“When you’re camera-trapping in an area, you’re basically giving snow leopards an opportunity to take their selfies,” says Sharma. “You need to know where they are and how many they are to conserve them. But sometimes you end up with thousands of pictures that are not relevant. That’s where AI comes to the rescue.”
Built by engineers from the Azure Machine Learning team, the scalable system will help conservationists focus more of their resources on researching population health, location and range. It can be integrated with Power BI so the Snow Leopard Trust can visualize and explore data from their cameras, which will help it develop and assess conservation programs.
Those efforts help snow leopards co-exist with humans, whether it’s reducing the impact of man-made infrastructures or working with herders who sometimes kill the cats after losing their livestock as prey.
The next step for the technology is automating the identification of individual snow leopards, based on their unique markings.
“The question is, ‘Is the snow leopard in Picture 1,240 the same snow leopard in Picture 1,000,240?’” says Microsoft software engineer Mark Hamilton, who built the model with deep neural networks, an AI technology that learns patterns like how a human brain learns.
Sharma’s teams have manually identified individual snow leopards in photos throughout the years, but have a backlog of roughly 10,000 photos awaiting cat IDs. Machine learning will alleviate the bottleneck, leading to more precise data and better population estimates.
The image classification model will be especially useful in supporting a new large study of the world’s snow leopard population, says Sharma. Announced last month, the study is part of the Global Snow Leopard & Ecosystem Protection Program, an alliance of conservation groups and the governments of all 12 countries where snow leopards roam.
“This is a massive undertaking,” says Sharma, the program’s international coordinator. The study will develop and use standardized research methods, collect enormous amounts of data and be expected to finish in about five years.
The project launched in the wake of last year’s controversial reclassification of snow leopards from an endangered species with a “very high risk” of extinction to a “vulnerable” species with a lower – but still high – risk of extinction.
The Snow Leopard Trust opposed the decision with other conservation groups, arguing that the best available science didn’t support it and that no one really knows how many snow leopards exist in the world. The best numbers are guesstimates, and a premature “down-listing” of snow leopards can compromise protection efforts for the cat.
“More research is needed,” Sharma says. “AI can help us speed up the task of estimating the world’s snow leopard populations.”
Other than their striking beauty, why study and protect snow leopards in the first place?
Two reasons, says Sharma. The cats are a “thermometer” of fragile mountain ecosystems that provide water for large human populations. And in addition to being “ghosts of the mountain,” they carry another distinguished title.
“Snow leopards cross borders without needing passports or visas,” he says. “When you talk about them, countries come together and appreciate the importance of conserving them. That’s why we call snow leopards ambassadors of the mountains.”
Top image: Photo by SLCF Mongolia/Snow Leopard Trust.