Foundation vision models for Antarctic sea ice floe segmentation and quantification from shipborne imagery

Sea ice floe size, ice concentration, snow cover, and thickness collectively drive the evolution of the marginal ice zone (MIZ) by influencing albedo, melt dynamics, wave-ice interactions, ice strength, and navigation conditions for icebreakers. Yet, reliable measurements of these parameters remain scarce, significantly limiting process understanding and model validation in the Antarctic. Satellite-based products are constrained by spatial resolution, and the scarcity of ground truth data prevents thorough validation and refinement of remote sensing retrieval algorithms. We demonstrate the use of the Segment Anything Model (SAM), a foundation vision model, to extract multiple physically meaningful sea ice properties from close-range, shipborne imagery. Using extensive datasets of high-resolution images collected during multiple Antarctic icebreaker voyages, SAM identifies and delineates individual ice floes, facilitating accurate estimation of floe sizes and sea ice concentration. Validation against manually segmented benchmarks shows robust agreement across diverse ice conditions. For snow cover and ice thickness estimations, SAM is specifically fine-tuned on manually annotated datasets to detect overturning ice events, enabling thickness measurement from exposed vertical profiles, and to classify snow-covered versus bare ice, quantifying snow cover fraction on individual floes. Overall, SAM enables systematic, scalable observations previously challenging to obtain, bridging the critical sea ice data gap by transforming images into quantitative datasets that support Antarctic sea ice process studies and improve observational detail beyond satellite capabilities.