Scientists from Russia have presented an algorithm that will be able to monitor the walrus population and count their number. Previously, researchers had to do it manually.
Employees of the Nenets Reserve and scientists from the Moscow Institute of Physics and Technology (MIPT) began to train an algorithm based on deep learning to identify walruses from photos and videos.
The scientists trained the algorithm on footage they filmed during a three-week trip to an island in the Barents Sea. To make things easier for themselves, they chose this territory, since there is a rookery of Atlantic walruses under protection. So the researchers got an archive of more than 10 thousand images, with their help, you can start training the program.
“Every year neural networks are being used more and more widely in studies of various species of marine mammals, but no one has yet worked with walruses in this direction, at least in our country. In this regard, the Pechora walruses are excellent for practicing this approach – they are a small and largely isolated group. This means that the likelihood of re-meeting the same walruses at the rookery in different years is quite high, ”said Varvara Semenova, coordinator of projects for the conservation of Arctic biodiversity at WWF Russia.
This algorithm will replace the scientists’ previous technology for fixing animals – camera traps. With their help, researchers receive tens of thousands of images per season, but they need to process and analyze their images manually. The neural network will be able to do this automatically – it will sort by pinnipeds, count their number and give each photo a small signature. This will save researchers hundreds of hours of work.
In the future, researchers at MIPT want to make the algorithm even more advanced so that it can recognize the individual traits of walruses. AI will be able to detect wrinkles, blemishes and scratches that distinguish one walrus from another. So the AI will learn to count the exact number of walruses in a particular rookery.