Semi-Automated Seal Detection on the Western Antarctic Peninsula: An Unsupervised Machine Learning Approach for Detecting Ice Seals in Aerial Survey Data

Abstract

Over the past 25 years, the Western Antarctic Peninsula (WAP) has experienced dramatic shifts in sea ice extent. This change has coincided with rapid alterations in ice-dependent ecosystems, including those supporting crabeater seals - the most abundant Antarctic seal and one of the largest mammalian consumers of krill. Despite their ecological importance, population estimates for ice seals remain scarce due to the difficulty of surveying large-scale, remote, ice-covered habitats. In 2023, during an abnormally low sea ice year, we conducted aerial surveys over Crystal Sound and Marguerite Bay during the end of the breeding season, flying over 1,000 km of transects. Seals were extremely sparse in the resulting imagery - occupying less than 1% of the surveyed area. This posed a significant challenge for both manual annotation and automated detection. Here we present a semi-automated, rules-based image analysis pipeline to substantially reduce human annotation time. Our method leverages hierarchical clustering with just two tunable parameters, avoiding the computational burden and opacity of deep learning models. Using this method, we identified 758 seals within a [~]350 km2 survey subset, achieving a test recall of 82%. In the absence of concurrent tagging data to estimate haul-out corrections, we refrain from extrapolating to a population estimate. However, the low observed densities highlight the urgent need for continued monitoring. Our improved data processing pipeline is a key step in facilitating the large-scale analysis required to inform conservation strategies for this key species.