Speaker: Dr Cameron Trotter, Machine Learning Research Scientist, British Antarctic Survey

Loss of marine biodiversity is a key issue facing the modern world. The removal of species from an environment can have profound effects on the overall ecosystem structure, though to what degree any species contributes to ecosystem stability is often unknown until they are removed. Due to its remoteness, relatively little is known about the structure of benthic ecosystems situated in the Southern Ocean around Antarctica. This region is among the most vulnerable to climate change and is currently one of the fastest-warming areas on the planet. Additionally, increasing human activity, including a growing number of vessels, poses further risks to these fragile ecosystems.

Traditionally, our understanding of Southern Ocean biodiversity has relied on nets or other sampling devices to bring benthic organisms to the surface. However, these methods are inherently destructive and provide limited insight into community structure. The development of underwater imaging technologies has enabled non-destructive, in-situ data collection, but analysing these images remains time-consuming and requires specialist expertise, as many of the organisms are found nowhere else on Earth. This has created a bottleneck, where data is collected faster than it be curated, significantly limiting our understanding of these ecosystems and how they are changing.

To address this challenge, we present the development of a deep-learning computer vision model trained to detect key taxa in Southern Ocean benthic imagery. Using only a small subset of labelled images from a high-resolution, downward-facing towed camera, the model learns to autonomously process unlabeled imagery, requiring only human verification of its output. This approach accelerates analysis and expands the spatio-temporal range of study compared to fully manual methods, offering a clearer picture of the current state of the Southern Ocean’s benthic ecosystems.