The project was led by Dr Cormac Purcell, Dr Andrew Walsh and Dr Andrew Colefax (Sci-eye) in close collaboration with Dr Paul Butcher at the NSW Department of Primary Industries in Australia.

Intelligent Systems for Planetary Good, led by Dr Cormac Purcell, and other researchers have published a study that investigated the ability of commodity deep learning algorithms to detect marine objects in video footage from drones, with a focus on distinguishing between shark species. The objective was to improve upon current drone-based shark surveillance methods which are effective, safe and affordable, but are labor intensive and rely heavily on pilot skill to detect sharks. (The Conversation)

AI to detect sharks

Dr Purcell and team set out to create a very reliable mobile app to be used by lifesavers on Australian beaches as a shark detector.

Since 2018 the New South Wales government has used trained surf lifesaving pilots to fly drones over the ocean at a height of 60 m, all the while watching the live video feed on portable screens for signs of sharks swimming below the surface of the sea.

However, this method has its drawbacks, including problems with water clarity, sea glitter, animal depth, pilot experience, and fatigue – all of which combine to achieve a shark detection success rate of only 60%.

Using AI has previously shown promise with the detection accuracy improving to over 90%. Unfortunately these systems could not be scaled up to realise real-world advantages on NSW beaches.

The researchers tried to combat these difficulties in their new shark detector mobile app by gathering a vast dataset of drone footage that was then labelled by shark experts for use in a machine learning model.

The model learned to recognise 10 kinds of marine life including dangerous sharks such as the great White and whaler sharks. This AI model was then integrated into a new mobile app for use on live drone footage and detection of risky marine life near beaches.

The mobile app was tested on five beaches in NSW during the summer of 2020, and the results were encouraging. The app identified dangerous sharks on a frame by frame basis 80% of the time in realistic conditions.

The authors of the study conclude that AI is now ready to be deployed in drone-based shark detection operations across Australian beaches.

However the system would need to be monitored and data updated frequently to maintain its reliability. But the payoffs are significant: Reliable identification means fewer unnecessary beach closures, fewer marine animals or sharks killed, and safer beaches in general.

Related Story: AI Helps More Accurately Count The Number Of Sea Lice On Salmon