Researchers led by Florida Atlantic University have built a deep-learning method that detects and tracks beluga whales in sequences of aerial drone images, working with Fisheries and Oceans Canada and the University of Manitoba. The study appears in the journal Frontiers in Marine Science

2

1

.

Counting belugas from the air has traditionally meant people going through thousands of images by hand. Calves are especially easy to miss; they are smaller and look different from one image to the next

2

1

. The team's answer is semi-supervised learning, where an AI model can learn from a small set of expert-labeled examples and then work through a much larger pile of unlabeled images on its own. "Instead of requiring researchers to painstakingly label thousands of images before AI can become useful, semi-supervised learning allows the technology to learn from a small amount of expert guidance and then make use of the much larger body of unlabeled imagery," said Hanqi Zhuang, the study's senior author and an associate dean in the university's engineering department

2

.

What the tests showed

The team trained five AI models on 7,655 high-resolution drone images of belugas, each whale labeled as adult or calf. The models were then compared across seven labeling levels running from 1 percent to 50 percent of the available labels

2

1

.

Source: Phys.org

Source: Phys.org

Per the study, the semi-supervised model scored a mean average precision of 19.2 after seeing just 34 labeled images

2

1

. That was a reported 20 percent relative improvement over the best supervised model in the comparison. Mean average precision is a standard accuracy score for detection systems. The gap widened further at the 5 percent labeling level. There the reported lead reached 6.6 percentage points

2

1

.

The study reports a bigger advantage for calves. At the 1 percent labeling level, calf detection improved by a reported 24.3 percent relative to the supervised baseline, against 16.8 percent for adults

2

1

. Calf counts matter for conservation: they offer clues about reproduction, population growth and long-term population health, and belugas are described as a vital part of the Arctic ecosystem

2

1

.

The researchers estimate comparable detection could be achieved with roughly five to 10 times less annotation, which would cut the time and cost of preparing images for analysis

2

1

. "It gives us the potential to look across far more images and observations than we could realistically process by hand, allowing us to better understand where belugas are, how they move, how successfully they reproduce and how those patterns are changing," said Greg O'Corry-Crowe, a National Geographic Explorer and research professor at the university's Harbor Branch Oceanographic Institute

2

.

The reported precision figures come from the researchers' own benchmark tests; the published articles include no independent evaluation of the method

2

1

. The payoff, if the estimate holds, is monitoring that a conservation program with a modest budget could actually run.

© 2026 TheOutpost.AI All rights reserved