The first step in building an intelligent Asian elephant ecological protection system is to establish an intelligent recognition model for Asian elephants. However, it is not simple to "recognize" Asian elephants from real-time shooting. At present, the accuracy of AI recognition models for wild Asian elephants in the world is only about 60%. Creating a high-precision intelligent recognition model for Asian elephants faces three main challenges.
Asian elephants are small in number and very scarce. Available images from pictures and or paintings collections are really limited and of poor-quality. Complete and clear Asian elephant images are scarce. Deep learning model training, however, requires massive image samples that are complete and clear for their efficiency. Therefor insufficient samples poses a great challenge to building the initial model.
Complex climate with frequent rain and fog, plus the nocturnal behavior of Asian elephants, result in dim and fuzzy videos and images from real-time observation. This require the new system to be able to precisely identify Asian elephants under poor light conditions from burred images, which translates into a higher demand on accuracy.
Using concealed equipment for image collection is necessary to avoid disturbing the wild Asian elephants and to protect the equipment from being damaged by the animals. Yet a limited view angle prevents the equipment from getting a complete picture of the giant elephants. Therefore, the AI model must accurately recognize Asian elephants by their parts like ears and tails, without mistaking them with domestic animals like cows and sheep.
For a more accurate identification model, sites were built across the habitat to capture more than 10,000 valuable images of wild Asian elephants, thus providing ample samples for model training. The subsequent training for deep learning focused on labeling and comparing data on local features of Asian elephants, including head, feet, back, and tail. During the three-month model training, researchers had been working on algorithm optimization. The samples were grouped to verify and compare different mathematical models which were repeatedly optimized for a higher accuracy. Eventually, the identification accuracy has been increased to a world-leading level of 96%.
Inspur not only greatly improved the recognition accuracy of Asian elephant models in a short time, but also provided powerful infrastructure support for large-scale Asian elephant data analysis with high-performance servers. Among them, Inspur AI server AGX-2 helps model training speed increase by 300% with its powerful computing power of 1 petaflops per second, providing efficient AI computing power support for Asian elephant image recognition processing and intelligent monitoring and early warning.
Backed by powerful AI solutions, the system can collect data on the behavioral patterns and migration routes of Asian elephants in the Xishuangbanna rainforest, identify elephants in milliseconds and send warnings to villagers in seconds via smart broadcasting, APP, and SMS. It is also able to locate Asian elephants and generate their routes in the APP in real-time.
With its efficient identification and warning functions, the system makes it easier for local authorities to deal with human-elephant conflicts and insure the observation no-mater the weather without disturbing the elephants. Meanwhile, it can perceive the ecological balance in the rainforest. With long-term scientific monitoring of temperature, humidity, and food chain in the forest, it can provides comprehensive, accurate data on the rainforest ecosystem research and preservation.
AI, a powerful force in transforming human life, is giving the planet’s ecosystems new hope. More endangered species will be saved, and human-nature harmony restored. "We strive to inspire a better world through intelligent computing," said Peter Peng, CEO of Inspur Information. In the future, Inspur will continue to utilize AI technologies to support ecological construction, contributing to a smarter and greener planet.