How AI Helped Snow Leopards

The surreal photo of an Afghan girl canoodling with an adult snow leopard would have been crowned “Photo of the Year”, with the duo striking just the right chord with those trapped in the concrete jungle-if the image hadn’t been fabricated.

To many’s disappointment, not only was the photo proven to be AI-generated, but the entire story circulating online was an invention of a creator named Babrak Khan. He conjured up the tale of the snow leopard being rescued by an Afghan girl when it was a cub, and then continuing to visit her out of gratitude after she released it back to the Pamir Mountains. Khan later explained that he wanted to raise awareness of the importance of harmonious human-animal relations.

All of this is a bit strange. Are human beings so detached from our entire ecosystem that they can only coexist with animals in digital fantasies? And, is this the most creative way humans can use their newfound AI superpower to improve the lives of their cohabitants on Earth?

Today, we’ll explore how Chinese AI technology is being used not to paint a snow leopard, but to track the elusive big cats in order to better protect them.

Snow leopards are listed as “Vulnerable” on the IUCN Red List, with only 3,920 to 6,390 individuals left in the wild, which shares the same threatened status with the iconic giant panda. In China, they are classified as a Class-A protected species, with hunting punishable by more than ten years in prison. The intense protection stems not only from their rarity but also from their cultural and ecological importance.

As early as 2000 years ago, snow leopards had already made their appearance in the ancient Chinese texts 山海经（Classic of Mountains and Seas）and Tibetan mythology. Even today, major central Asian cities like Bishkek are using the snow leopard as their emblems.

Ecologically, snow leopards are considered a barometer of high-altitude ecosystem health. They inhabit the source regions of China’s three major rivers, making their survival essential to maintaining the fragile balance of the Tibetan Plateau. This, in turn, impacts water security for China’s 1.4 billion people.

However, conservation efforts for snow leopards have been met with challenges. Snow leopards live in rugged, high-altitude terrain between 3,300 and 5,000 meters, spread across an area of two million square kilometres—the size of Greenland or Mexico. Each snow leopard has an enormous home range, with some individuals covering up to 4,500 square kilometres. Their low population density makes it incredibly difficult for researchers to locate them.

Infrared cameras used to monitor wildlife are also limited in their efficiency when used alone to track snow leopards. For instance, Ma Duifang, head of the wildlife protection department at the Qilian Mountain Nature Reserve, started receiving reports from local herders about snow leopards in the area as early as 1990. However, it wasn’t until 2012 that his infrared camera finally captured one of them.

Moreover, infrared cameras struggle to capture the target animal, leaving the scientists with voluminous empty footage or footage of non-target animals. A fully charged infrared camera can run for four months, capturing about 2,000 photos and 800 videos during that time. This means that over the past 22 years, Mr. Ma has manually gone through 132,000 photos and 52,800 videos before finally locating that one shot, a time-consuming and costly process that hampers all conservation efforts.

Mr. Ma Duifang (left)

Fortunately, in 2021, Chinese tech giant Tencent proposed a solution. An AI model of Tencent’s AI Lab is now capable of automatically detecting snow leopards from 31 wild animals in camera footage, and filtering out empty frames. This breakthrough reduced researchers’ workloads by over 50%, allowing them to deploy more cameras without being overwhelmed by data.

This wasn’t the first time AI had been used in snow leopard research and conservation. In fact, as early as 1998, Professor Philip Riordan from Southampton University used neural networks (SOM) and Bayesian methods to identify individual snow leopards by comparing snow leopard footprints. With the greater ambition of identifying more traits of the snow leopards, Tencent’s algorithm is therefore required to process much more data.

The Tencent AI model is tasked with first distinguishing the snow leopards apart from the natural environment where they hide, camouflaged by their spots and stripes, and then with recognizing animals appearing in the shot. Given the low frequency of capturing snow leopards on film, they had to train the model using the bare minimum amount of video footage.

During the training process, Tencent used over 100,000 photos of snow leopards and other wildlife. Human reviewers were brought in to label the images to speed up machine learning. As part of this, Tencent partnered with a Chinese charity that supports people with disabilities, allowing them to participate the project remotely from home and earn money by simply clicking a mouse. Thanks to the joint efforts of the charity, the algorithm was completed in just four months and was first deployed in the Qilian Mountain Nature Reserve in China.

According to staff feedback, what used to take a team of people 2 to 3 weeks to process now only requires two people to upload the data and let the AI handle the identification automatically. In its first year of use, the AI model processed 60TB of data, leading to the discovery of over 400 snow leopards and helping researchers publish more than 10 papers on snow leopard conservation.

In addition, the accuracy of Tencent’s AI is also astonishing, achieving a 98% success rate. For comparison, Facebook rolled out DeepFace in 2015, a facial recognition system that was trained on four million images uploaded by Facebook users, its accuracy is only up to 97.6% on the Labeled Faces in the Wild (LFW) dataset. Similarly, the FBI launched its own facial recognition system, and by the end of 2017, the FBI completed internal tests of the NGI-IPS algorithm, which showed an accuracy rate of just 85%.

China is now home to the largest number of wild snow leopards in the world. Around 60% of the world’s snow leopard habitat and its 63.78% wild population are found in China. Taking into account the fact that wild animals are not bound by state borders, Tencent launched an online data-sharing platform to support wildlife conservation across borders. By coordinating data from various nature reserves, the platform helps track snow leopard populations and migration patterns.

Since 2012, Tencent has partnered with local communities to train herders and farmers in snow leopard conservation and monitoring. They helped establish 21 grassroots wildlife protection groups, involving 350 local residents. By 2022, they had recorded 97 species of wildlife and 221 plant species. Tencent also visited local farmers to assess livestock losses caused by snow leopards, bears, and other predators, providing compensation through Human-Animal Conflict Foundation, a fund jointly supported by the government, Tencent, and commercial insurance companies. In Yaqiu, a village located in Qinghai province, 804 incidents were documented, and 184 herding families benefited from the aforementioned fund. This effort also reminded local governments of the necessity of providing diverse forms of support to herders, such as commercial insurance.

All this information on wildlife observation and conservation is uploaded to an online database developed by Tencent, helping other nature reserves in China make more informed decisions. This scientific collaboration even extends beyond China. Specifically, conservationists from Nepal, where snow leopards are found along its border with China, have been among the groups that benefited most from this data-sharing platform. With the wisdom of hindsight, I couldn’t help but view Tencent’s success as a product of seamless cooperation between the Chinese government, conservationists, and private enterprises like Tencent.

While AI technologies elsewhere are applied to ChatGPT or MidJourney, which generate fictional content in the blink of an eye, Tencent went another route engineers chose to immerse themselves in the real world, however harsh it might be.

I find the words of Anton Chekhov particularly fitting for this moment: And, you know, whoever has once in his life caught perch or has seen the migrating of the thrushes in autumn, watched how they float in flocks over the village on bright, cool days, he will never be a real townsman, and he will have a yearning for freedom to the day of his death.

For Tencent’s engineers, a glimpse of a snow leopard caught by the camera might just be enough to inspire a lifetime of commitment.