DeepMind is working with conservationists to map the animal population in Africa’s Serengeti National Park. To this end, the company uses machine learning to relatively quickly recognize and count the animals in the millions of photos taken with motion cameras.
To better study the behavioral dynamics of the entire fauna population from the Tanzanian reserve, DeepMind uses machine learning methods. It uses millions of photos taken over the past nine years with 225 motion cameras placed in the park by the Serengeti Lion Research program. Millions of photos have been collected and stored in the past nine years, but identifying and counting the different animal species is a labor-intensive and time-consuming task, according to DeepMind. According to the company, this prevents scientists from doing simple research and makes it difficult for conservationists to respond to challenges that could disrupt the ecosystem.
The SnapShot Serengeti dataset classification process started in February 2018 and has already been published in 2015 in the scientific journal Nature. The researchers wrote Python scripts to extract the date, time and other information from the image files. Subsequently, more than 28,000 volunteers set to work to classify forty animal species. That process is currently only 24 percent. To speed this up, DeepMind trained machine learning based on the dataset. According to Google’s sister company, this has resulted in a model that works as quickly and efficiently as the volunteers, or better. In particular, it also helps in speeding up the processing procedure. It normally takes about a year for the photos taken by volunteers to be tagged; that is now reduced to nine months.
The motion cameras have been placed in an area of 1125 km² and work on the basis of infrared sensors. As soon as an object that is warmer than the ambient temperature comes in front of the sensor, pictures are taken. Most of the time, animals can be seen, although the tall, sunlit grass of the Serengeti can trigger the cameras as soon as it moves in the wind. Two models are used: the Scoutguard 565 and the DLC Covert Reveal. They can take color photos both during the day and at night and use a white flash light. According to the Serengeti Lion Research team, this does not deter the animals and they regularly return to the location of the cameras. The cameras need maintenance after two months, such as changing the batteries and SD cards. Sometimes they also need to be replaced or repaired in their entirety, for example if it rains too much or if hyenas have bitten them, elephants have trampled them or ants have taken possession of them.
The placement of the cameras in the area. In the overview map, the dotted line represents a study area for lions.