Segmenting sea ice floes in close-range optical imagery with active contour and foundation models

Abstract

The size and shape of sea ice floes play a crucial role in influencing ocean-atmosphere energy exchanges, sea ice concentrations, albedo, and wave propagation through ice-covered waters. Despite the availability of diverse image segmentation techniques for analyzing sea ice imagery, accurately detecting and measuring floes remains a considerable challenge. This study presents a precise methodology for in-situ sea ice imagery acquisition, including automated orthorectification to correct perspective distortions. The image dataset, collected during an Antarctic winter expedition, was used to evaluate various automated image segmentation approaches: the traditional GVF Snake algorithm and the advanced deep learning model, Segment Anything Model (SAM). To address the limitations of each method, a hybrid algorithm combining traditional and AI-based techniques is proposed. The effectiveness of these approaches was validated through a detailed analysis of ice floe detection accuracy, floe size, and ice concentration statistics, with the outcomes normalized against a manually segmented benchmark.