AI Model Counts Beluga Whales in Drone Images

Researchers say a semi-supervised deep learning method can detect beluga whales and calves in aerial photos using far fewer labeled images.

Updated on Nov. 19, 2026 in Wildlife

Flat vector editorial illustration of a white drone above dark blue water with pale beluga whale silhouettes and a calf beneath the surface.
Florida Atlantic University researchers say a semi-supervised deep learning model can count beluga whales and calves in drone images using 34 labeled examples. AI Illustration. Upload story photo >

Florida Atlantic University researchers developed a deep learning approach that detects and tracks beluga whales in aerial drone images. The semi-supervised method, called SEMI-DETR, reached a mean average precision of 19.2 using just 34 labeled images.

Why it matters

Counting beluga calves provides clues about reproduction, population growth and long-term population health, but manually reviewing thousands of aerial images is labor-intensive. A method that needs far less labeled data could make monitoring Arctic beluga populations more efficient as Arctic environments change.

The study used 7,655 high-resolution aerial drone images and compared five AI models across labeling levels from 1% to 50%. SEMI-DETR improved calf detection by 24.3% and adult detection by 16.8% over the supervised baseline at the 1% level, though the source does not report how these results hold outside the study data.

The players

Florida Atlantic University

A public research university in Boca Raton, Florida, whose researchers led the development of the beluga detection approach.

Harbor Branch Oceanographic Institute

A marine research institution affiliated with Florida Atlantic University that collaborated on the study.

The details

The approach uses deep learning to detect and track belugas across sequences of aerial images, learning from a small set of expert-labeled images and then from a larger body of unlabeled imagery. The researchers say comparable detection performance was achieved with roughly five to 10 times less annotation than supervised methods.

Timeline

  1. Study results were published in Frontiers in Marine Science on November 19, 2026.

Culture Shift

The study extends Florida Atlantic University's Harbor Branch Oceanographic Institute marine research into automated monitoring of Arctic wildlife. It reflects a broader move toward using drone imagery and machine learning to track animal populations where field observation is difficult.

If the method proves reliable in the field, conservation scientists could monitor beluga populations with less manual image review. The source does not give a deployment timeline, so practical use remains a future possibility rather than a current tool.

The takeaway

Automated tools that need fewer labeled examples could let conservation teams track vulnerable wildlife populations more quickly and with less manual effort.

Further reading

Explore more on Wildlife for related coverage of animal populations and conservation.

Source note: This article includes information reported by Phys.