Researchers at Florida Atlantic University have trained an AI model to find beluga whales in drone video from the Canadian Arctic using only 34 hand-labelled images. The study appeared in Frontiers in Marine Science on 4 September. The university announced it on 19 November.
Working with Fisheries and Oceans Canada, the team used 7,655 frames from 4K video shot by DJI Phantom 4, Mini, Mavic 3E and Matrice drones in 2022 and 2023. In semi-supervised learning, experts label a few images by hand and the model then learns from many more that carry no labels. The authors say it gives comparable results with five to ten times less labelling.
Calves are the hardest whales to spot from the air, and counting them shows whether a population is reproducing. With 34 labelled images the semi-supervised model, SEMI-DETR, found 27.8% of the calves in the test set, against 1.3% for a conventionally trained YOLO11 model. With 1,709 labelled images it found 66.6% of them.
The paper puts the misses down to resolution. The team's permits allowed flights at 20 metres, where the smallest calves cover 1 by 2 pixels. It recommends flying lower where the rules allow, or using cameras with more zoom. Senior author Hanqi Zhuang, a professor in the university's College of Engineering and Computer Science, said the method lets the technology learn from a small amount of expert guidance.
Background
Hanqi Zhuang
Professor at Florida Atlantic University
Professor in Florida Atlantic University's College of Engineering and Computer Science.
Greg O'Corry-Crowe
Research professor at Florida Atlantic University
Research professor at FAU Harbor Branch Oceanographic Institute.



