New study suggests warming seas are negatively affecting beluga whales' aggregation patterns

Until recently, technology limitations have made it challenging to effectively study the aggregation behavior of beluga whales. As climate change continues and sea surface temperatures rise quickly, the ability to do so becomes a priority, requiring methods that can capture data completely and accurately.

To this end, a team of researchers from Spain's Universities of Cadiz, Alicante, and Leon has used deep convolutional neural networks (CNNs) to analyze open-access satellite imagery from Arctic areas of Canada, Russia, and the U.S. state of Alaska in order to positively identify the presence of beluga whales in these areas, compare sea surface temperatures (SSTs) to beluga aggregation patterns, and—based on data from the Intergovernmental Panel on Climate Change (IPCC)—create three representative concentration pathway (RCP) scenarios through the end of the 21st century.

The work appears in Frontiers in Marine Science.

As of 2017—the year of its most recent assessment of this species—the IUCN had classified beluga whales (Delphinapterus leucas) within its Red List conservation category of Least Concern. However, these whales make their home in isolated Arctic waters and depend on sea ice for protection from predators such as orcas, and for nutrition. Algae growing within the ice draw fish that are a mainstay of the belugas' diet.

Research published in 2020 has found that before the end of this century, ice could conceivably disappear from the Arctic during summers.

Research teams have previously used direct-observation aerial surveys to study cetaceans; however, this method is expensive, time-consuming, and can result in incomplete data due to its range limitations. Especially in rapidly warming seas, the accuracy of species-distribution models (SDMs) created through aerial surveys is time-limited.

At the end of the 20th century, researchers began using remote sensing imagery to collect data on individual whales and aggregations, but for studies involving marine mammals, employing this method on its own has also proven inaccurate.

Satellite images meet a deep CNN

Wondering if applying AI image analysis to satellite images might provide a solution, the research team behind this new study accessed high-resolution (less than 1 meter/pixel) satellite images of Arctic coasts obtained between 2007 and 2020. They first trained a deep CNN model to identify beluga whales based on a set of free images from Google Earth and Mapbox. Of 1,400 images in the set, half were of belugas and the other half were of icebergs.

The researchers built up the volume of the training data by randomly flipping some of the images, rotating others to random factors between zero and 360 degrees, cropping some of them, and adjusting image scales and brightness levels. After testing the model, they trained it further using 700 aerial images "in which each beluga whale is annotated within a bounding box (the total number of bounding boxes is 846)," they write.

Two of the researchers visually assessed and annotated each image with data on its class (icebergs or whales) and the number of whales, using whale-watching websites to verify the presence of belugas in the pictured areas. They compared images from the same spots on different dates to discern belugas from the sea floor. This work became important to determining the model's accuracy later in the process.

Finally, the team applied their model to images in which the whales were most probably present, and used it to detect and count the whales within specific grid cells and to analyze aggregation patterns. However, icebergs presented a problem.