Species classification of Antarctic pack-ice seals using very high-resolution imagery

Abstract

We introduce a semiautomated machine learning method that employs high-resolution imagery for the species-level classification of Antarctic pack-ice seals. By incorporating the spatial distribution of hauled-out seals on ice into our analytical framework, we significantly enhance the accuracy of species identification. Employing a Random Forest model, we achieved 97.4% accuracy for crabeater seals and 98.0% for Weddell seals. To further refine our classification, we included three linearity measures: mean distance to a group's regression line, straightness index, and sinuosity index. Additional variables, such as the number of neighboring seals within a 250 m radius and distance of individual seals to the sea ice edge, also contributed to improved accuracy. Our study marks a significant advancement in the development of a cost-effective, unified Antarctic seal monitoring system, enhancing our understanding of seal spatial behavior and enabling more effective population tracking amid environmental changes.