Drones are meant to spot sharks early at New South Wales beaches without catching or killing marine wildlife. A new review of 900 flights shows how hard it is to identify bull, white and tiger sharks reliably on a small screen in real time.
Pilots reported at least one of those three target species on 269 flights. Later high-resolution video checks confirmed that call in only 101 cases. The remaining 168 reports, or 62 percent, were misidentifications.
Non-lethal surveillance under time pressure
Drones are a central part of the New South Wales Shark Management Program. Since 2017, more than 160,000 flights have been conducted at 56 beaches. Unlike shark nets or traditional drumlines, drones are intended to reduce risk for swimmers without trapping sharks and other marine animals.
The patrols usually run between 9 a.m. and 4 p.m. in suitable weather. A drone surveys up to one kilometre of coastline, commonly flying at about 60 metres behind the surf zone, while pilots must interpret the live feed quickly enough to warn beach managers.
The problem is not seeing animals, but naming them
The study separates detection from identification. A dark shape in clear water may be easy to spot, but assigning it to one species from above is much harder. Body shape, swimming pattern, water glare, depth and screen quality can all change the call.
That matters because management actions are tied to species. Bull, white and tiger sharks trigger stronger responses than many other large marine animals. If harmless or lower-risk animals are logged as target sharks, beaches may be cleared unnecessarily and shark presence can look higher than it is.
False alarms distort the risk picture
In the live drone reports, target sharks appeared roughly twice as often as they did after expert review. The authors argue that unverified real-time reports can therefore inflate perceived risk and complicate decisions about beach closures, public alerts and future investment.
The result does not make drones useless. It shows where their limits sit. Drones can provide rapid, non-lethal information, but species-level calls need training, quality control and, where possible, post-flight verification.
Why AI alone will not solve it immediately
Automated recognition could help, but the study is cautious. Algorithms still need high-quality labelled data, and coastal conditions are messy: glare, waves, turbidity, animal depth and partial views can defeat simple image rules.
For beach management, the practical lesson is modest but important: drone programs should report uncertainty clearly, distinguish sightings from confirmed identifications, and avoid treating every live target-shark call as a verified threat.
