AI Could Transform How Drones Identify Sharks At Australian Beaches

The humming noise above you does not sound natural. And, if you’re at a beach in New South Wales, Australia, during the summer, chances are it is not natural. In fact, it is most likely a drone flying overhead, scanning the water for sharks. And when a shark is spotted in the live video feed, the pilot manning that drone has only moments to decide what they are looking at. Is it a white shark? A bull shark? Something else entirely? What they say — that decision — can have immediate consequences. A warning might be issued, followed by a deafening siren, with people being asked to leave the water by lifeguards via megaphone. A rescue vessel could even be deployed if a person is close to what that shadowy thing is lurking under the water.

But what happens if the animal was not actually a shark that posed a significant risk to people? What if it was another shark species, a guitarfish or even a fish? That is the question at the heart of new research examining the accuracy of real-time shark identification during drone surveillance along the New South Wales coastline. And the results reveal just how difficult it can be to identify an animal from above, even with experienced observers and high-quality equipment.

Researchers examined 900 drone flights conducted as part of the NSW Shark Management Program. During these surveys, professional drone pilots identified sharks in real time as either a "target species" (i.e., bull sharks, white sharks and tiger sharks, the three species that are the focus of shark mitigation measures in NSW) or an "other shark." The researchers then went back to the recorded footage; unlike the pilots working in real time, they could pause the video, slow it down, replay it and examine the animal more carefully. This post-flight assessment was treated as the best available estimate of what species had actually been present.

And… well, it wasn’t great results.