Protecting Snow Leopards with Technology

Framed by clouds and capped by a layer of permanent snow, the San Jiangyuan Nature Reserve on the Qinghai-Tibet Plateau and surrounding grasslands are home to 69 protected species, including the Tibetan antelope, wild yak, Tibetan fox, black-necked crane, and the snow leopard.

As part of the panthera genus, snow leopards derive their name from living above the snow line. These solitary hunters pace the mountains camouflaged by distinctive gray-black leaf markings, confident in their position at the top of the food chain. They are also an "umbrella species", meaning that their well-being in turn impacts other species and the ecosystem as a whole.

A snow leopard in profile

Snow leopards are classified as vulnerable on the International Union for Conservation of Nature (IUCN) red list. Fewer than 8,000 are thought to remain in the world, 60% of which are located in China. Established in 2007, the Shanshui Nature Conservation Center is committed to protecting species and their habitats, including multiple snow leopard habitats, on the Qinghai-Tibet Plateau, the mountain range in the south west, and around the surrounding cities.

Researchers from the center note that the nocturnal snow leopard has strong territorial instincts, roaming and hunting over an area of up to 200 square kilometers, which makes the elusive cat hard to find and protection and research extremely difficult.

AI recognition

To observe snow leopards without interfering with their behaviors, researchers usually install infrared cameras in their habitats. When an animal passes in front of the camera and is sensed by infrared rays, photos and videos are captured.

A snow leopard captured by infrared camera

Because the markings on snow leopards are as unique as a human fingerprint, researchers can identify individual animals, giving greater insight into their roaming, predation, and cub-rearing behaviors.

However, the main bottleneck in existing research has been the heavy reliance on manual labor, preventing camera footage from being quickly converted into actionable insights.

Field work conducted by researchers from Shanshui Nature Conservation Center

Conservationists or local herders needed to travel to the actual camera sites, retrieve memory cards, and import the images into a device for processing and analysis. This is complicated by the fact that snow leopards thrive in harsh terrain that is hard to access or connect with communications technologies, making the whole process time-consuming and labor-intensive.

When viewing the data, researchers also had to identify specific cats manually – an extremely tough challenge given the massive amounts of image data collected year-round and the fact that snow leopards tend to roam at night or at dawn, adding to the natural camouflage that sees them blend them in with the rocky terrain.

Processing 500,000 photos per year manually takes about 300 hours. Nevertheless, when combined with sufficient computing power, data on this scale provides a treasure trove of training material for deep learning. It is planned to use this information to compile databases for research and develop more effective conservation strategies.

The partners applied a full-scenario AI framework to process infrared camera footage, the first time that an open source model based on an AI framework has been used in this way. The AI framework for device, edge, and cloud scenarios is designed to build a new AI programming paradigm that allows developers to create better, efficient, and flexible AI software and hardware applications. Huawei and Shanshui developed a single inference display page and a batch inference tool based on the framework to process an initial batch of 280,000 infrared photos.

The following image shows the inference result and effect diagram:

Compared with manual recognition, it only takes about two and a half days for the machine to complete the preliminary screening of batches of 500,000 photos. Although final identification still requires expert analysis and review, AI halves the overall time required.

The trained model can accurately identify 10 common species or species categories, including the snow leopard, red fox, and blue sheep. The overall recognition accuracy in the validation set has reached about 92%, the most accurate of which is the snow leopard, for which recognition is 95%.

Species categories currently recognized by open source models

Infrared Camera Reasoning Photo: Snow Leopard

For developers interested in conservation, Huawei has open sourced the models and tools for species identification, lowering the threshold for developing related tools and applying datasets, data processing, and data cleaning. Huawei has also opened up its basic software and hardware platforms.

It's great to see AI starting to benefit biodiversity conservation, helping us process infrared camera data faster. I hope that the tech industry will get more involved in biodiversity – I believe that in the future, science and technology will bring more possibilities to nature conservation.Cheng Chen
Director, Nature Observation Project, Shanshui Nature Conservation Center

To protect biodiversity, ICT is proving to be an invaluable tool, improving efficiency and accuracy by orders of magnitude – benefits that should help ensure the longevity of vulnerable species like the snow leopard.