A Cross-Scale 3D Attention and Boundary-Aware Network for Sea–Land Segmentation

Coastal segmentation struggles with complex terrain configurations and discontinuous edge demarcation. To resolve these challenges, we propose a novel deep learning model, the Cross-Scale Attention and Boundary-Aware Network (CABANet). The network synergistically enhances segmentation performance through two core innovations. First, we design a Cross-Scale 3D Attention Fusion Module, which creatively stacks multi-scale 2D feature maps from the backbone into a pseudo-3D feature volume. It then performs efficient contextual modeling across different geographic scales via a decomposed 3D attention mechanism, generating a highly optimized semantic summary feature. Second, we design a boundary enhancement module to strengthen the extraction and analysis of boundary features and improve the segmentation of the sea–land boundary, while utilizing contextual information to guide the boundary enhancement process. Evaluations on Landsat-8 OLI and SLSD datasets confirm state-of-the-art performance in both regional homogeneity and edge continuity.