Sea Ice Semantic Segmentation in Optical Image Based on Adaptive Training Sample Selection and Cross-Attention ResUNet

The formation of numerous channels among Arctic sea ice provides potential routes for Arctic navigation and the identification and semantic segmentation of sea ice becomes a crucial task. This letter proposes a sea ice semantic segmentation method with adaptive training sample selection and cross-attention mechanism to enhance the robustness under the complex climatic conditions of the Arctic. First, the image is divided into patches. An adaptive iterative clustering on them automatically selects the training samples. Second, ResUNet with a cross-attention mechanism is used for image segmentation. This approach enhances contextual understanding with relatively low computational overhead, enabling better focus on relevant features across different layers of the network. The experimental results demonstrate that the proposed method can achieve high accuracy segmentation with a small training set. Furthermore, the proposed method exhibits segmentation consistency across two datasets and various types of sea ice.