Sea Ice Semantic Segmentation with Sentinel-2 Data Based on Adaptive Sample Training on U-Net Network

The rapid melting of Arctic sea ice presents significant opportunities and challenges for humanity. The formation of numerous channels between the ice offers potential for Arctic navigation. The identification and semantic segmentation of sea ice is a crucial task in sea ice monitoring. To reduce the influence of complex climatic conditions in the Arctic region on the robustness of the sea ice segmentation model, this paper proposes a sea ice semantic segmentation model based on adaptive training sample selection on U-Net for Sentinel-2 data. The method adaptively selects training samples through unsupervised iterative clustering and inputs them into U-Net network for image segmentation. In addition, the iterative efficiency of clustering is improved by building subspaces. The experimental results on Sentinel-2 data show that the proposed method can effectively achieve sea ice segmentation with a high positive detection rate and low false alarm rate.