Integrating a data-driven classifier and shape-modulated segmentation for sea-ice floe extraction

Sea-ice images are foundational resources for studying the evolution of sea-ice floes (SIFs). However, extracting SIFs from diverse and complex sea-ice images using conventional segmentation methods is both tedious and challenging. The precision of identified SIFs is often compromised by various interferences, including low contrast, speckle noise, melt ponds, and slush ice. In this paper, we employ random forest machine learning to bolster the robustness of ice-water classification against interference. This data-driven approach reduces the reliance on subjective visual inspection for classification. Furthermore, we incorporated shape modulation within the framework of irregular watershed segmentation. By emphasizing contour convexity, we adeptly merge the oversegmented SIFs, thereby addressing the challenges inherent in traditional segmentation methods. Our methodology enhances accuracy and considerably reduces processing time. It also requires less manual intervention than traditional methods. This elevated level of accuracy and efficiency facilitates the extraction of SIFs from diverse types of sea-ice images, fostering research into SIF evolution under varying atmospheric and oceanic conditions.