Semi-Supervised AI Boosts Beluga Whale Calf Detection in Arctic Drone Surveys

For Arctic researchers, counting beluga whales from the air can mean painstakingly examining thousands of aerial images—searching for animals that may be difficult to see and even harder to distinguish from one another. The challenge is especially significant when researchers are trying to identify calves, whose smaller size, lower numbers, and variable appearance can make them easy to miss.

Two beluga whale cow-calf pairs in a shallow river estuary in the High Arctic. (Image credit: Greg O’Corry-Crowe and Cortney Watt, Arctic Whale Research Program – FAU/DFO)
Two beluga whale cow-calf pairs in a shallow river estuary in the High Arctic. (Image credit: Greg O’Corry-Crowe and Cortney Watt, Arctic Whale Research Program – FAU/DFO)

Knowing how many calves are present can provide critical clues about reproduction, population growth, and the long-term health of beluga populations.

Now, researchers from Florida Atlantic University‘s College of Engineering and Computer Science and Harbor Branch Oceanographic Institute, in collaboration with Fisheries and Oceans Canada and the University of Manitoba, are turning to artificial intelligence to make that process faster and more precise. Their new approach uses deep learning to detect and track beluga whales across sequences of aerial images, helping researchers automate a task that traditionally requires extensive manual annotation.

The approach could provide conservation scientists with a more efficient way to monitor beluga populations while reducing the amount of labor-intensive human effort needed to train AI systems for wildlife detection.

For the study, researchers used a dataset of 7,655 high-resolution aerial drone images of beluga whales, with each whale identified as either an adult or calf. They trained and compared five AI models to detect the whales in the images, testing how accurately the models performed when trained with different amounts of human-labeled data. The researchers evaluated the models using seven levels of labeled data, ranging from just 1% to 50%, to determine how effectively AI could detect belugas when only a limited amount of manually annotated imagery was available.

Results of the study, published in the journal Frontiers in Marine Science, reveal that AI can detect beluga whales in aerial imagery with substantially less human-labeled data than conventional supervised AI systems require. The researchers found that their semi-supervised approach, SEMI-DETR, achieved a mean average precision of 19.2 using just 34 labeled images—only 1% of the labeled data available—representing a 20% relative improvement over the best-performing supervised model. At the 5% labeling level, the advantage increased to 6.6 percentage points.

“Artificial intelligence is opening a new path for conservationists to make sense of the enormous volumes of imagery that modern monitoring technologies can collect,” said Hanqi Zhuang, Ph.D., senior author and associate dean and professor in the Department of Electrical Engineering and Computer Science within FAU’s College of Engineering and Computer Science. “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. This has the potential to transform how we monitor wildlife—making sophisticated computer vision tools more accessible, scalable, and practical for scientists working in some of the most challenging environments on Earth.”

The benefits were particularly pronounced for detecting beluga calves, which are smaller, less frequently observed, and more variable in appearance than adults. At the 1% labeling level, SEMI-DETR improved calf detection by 24.3% relative to the supervised baseline, compared with a 16.8% improvement for adults. Overall, the researchers found that comparable detection performance could be achieved with approximately five to 10 times less annotation, potentially reducing the time and cost required to prepare imagery for automated analysis.

“This research demonstrates how artificial intelligence and engineering can be applied to address complex environmental challenges,” said Stella Batalama, Ph.D., Dean of FAU’s College of Engineering and Computer Science. “By developing innovative technologies that make wildlife monitoring more efficient and accessible, our researchers are creating practical solutions that can strengthen conservation efforts and help protect vulnerable species and ecosystems.”

The findings provide a benchmark for using semi-supervised AI in marine mammal monitoring and offer guidance for conservation programs with limited resources for manually annotating imagery.

“For those of us who study wildlife, the value of this technology goes far beyond simply counting animals,” said Greg O’Corry-Crowe, Ph.D., co-author, research professor at FAU Harbor Branch, and a National Geographic Explorer who studies wildlife evolution and behavior. “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. Belugas are a vital part of the Arctic ecosystem, and as their environment changes, better tools to monitor their populations will be increasingly important for conservation.”

Additional co-authors are Mohammed G. Al-Jassani and Chiron Bang, both Ph.D. students in FAU’s Department of Electrical Engineering and Computer Science; Cortney A. Watt, Ph.D., a research scientist; and Mark Gillespie, a researcher and field technician, both with Fisheries and Oceans Canada, University of Manitoba.

This work was supported by the National Geographic Society, FAU Harbor Branch, Fisheries and Oceans Canada, the World Wildlife Fund, the Nunavut Wildlife Management Board, and Natural Resources Canada’s Polar Continental Shelf Program.

latest edition

From Arctic sea ice to the North Sea to the waters offshore Louisiana, applied marine science is looking beneath the surface, testing which new observing technologies earn the trust needed for wider adoption, and what it takes to turn ambitious science plans into safe and reliable work practices in marine environments.

got marine science news?

Send us your latest corporate news, blogs, or press releases

Search