How Do You Spot an 800-Pound Sea Cow? With a Drone
Dugongs—Australia’s version of the lovably chubby manatees of Florida—are hard to keep track of.
Despite growing up to 10 feet in length and weighing 800 pounds, dugongs are just needles in a haystack when it comes to locating individuals in the open ocean. For scientists trying to monitor the endangered species, that’s a problem.
The marine mammals are under threat from boat strikes, fishing net entanglements, and habitat destruction, so obtaining accurate data on dugong populations is critical to conservation efforts.
Now researchers at Australia’s Murdoch University and Queensland University have combined drone technology with spatial imaging software to scan, photograph, and identify dugongs over large areas of open water.
Amanda Hodgson, a dugong researcher at Murdoch University, has been deploying drones since 2010 to take images of their habitat off the coast of Queensland, in the Torres Strait, and in Shark Bay off Western Australia.
While the drones did the photographing, humans like Hodgson were left to pore over the images, trying to identify the little gray specks that signified a dugong on the waters below.
The research team has stared at more than 30,000 photographs of blue water, searching for dugongs—a time-consuming endeavor. To speed things up, Hodgson recruited computer scientist Frederic Maire from Queensland University.
The team used Google-developed software called TensorFlow, a machine-learning platform that can be taught to pick out sea cows in photos automatically. But first, team members had to upload the images and point out where the dugongs were to the system, so it could learn to identify them.
The system is now searching images on its own. Maire told the Australian Broadcasting Corporation that the program is detecting about 80 percent of dugongs in the photos, and the rate is expected to improve over time.
“To train this deep neural network you need a lot of data. Initially we didn’t have much data, so we do this incrementally,” he said. “The more the neural network is provided with examples, the better it will get.”
The team believes the software could work for other species, including whales, koalas, and sharks.
“We can imagine a drone patrolling the beaches, trained to detect sharks, and if a shark is detected, then send a text message to lifeguards to warn them,” Maire said.