A research team used a state-of-the-art artificial intelligence technique called deep learning to identify, count, and describe animals in the wild. The deep learning method uses motion-sensor cameras that record photographs which afterward are described by deep neural networks. Such a system reduces the time spent in manually identifying the animals by 99.3 percent while doing it at the same rate of accuracy (96.3 percent) achieved by human volunteers. Results of the study have been published in the Proceedings of the National Academy of Sciences journal, according to the University of Wyoming News.
Jeff Clune, the paper’s senior author, said wildlife data can now be gathered accurately, unobtrusively, and inexpensively by using their experimental procedure. Not only that but various fields in ecology, wildlife biology, zoology, conservation biology, and animal behavior can be transformed into big data sciences. Such transformation will improve mankind’s ability to study and conserve wildlife and other important ecosystems, Clune added.
The researchers used data available from Snapshot Serengeti, a citizen science project on the zooniverse platform. Snapshot Serengeti has a large collection of camera traps or motion sensor cameras in Tanzania that assemble millions of images of animals in their natural habitats such as lions, leopards, cheetahs, and elephants. The information found on these photographs is only useful once it has been changed into text and numbers.
The traditional method of extracting such information was to enlist the help of human volunteers to label each image manually. The data used by the present research consists of 3.2 million labeled images that have been processed by more than 50,000 over the course of several years. However, the new deep learning strategy not only was able to tell which is which among the 48 different animal species but also revealed where these animals are and what they are doing. It also showed if they are either sleeping or eating, and if there are offspring around. Such a technology would cut short the labeling of 3 million images by eight years. The present study proved the value of such technology to the wildlife community, according to the researchers.