A Structure-Aware Semantic Prototype Constraint Network for Remote Sensing Image Segmentation

Remote sensing images are critical for extracting land surface information, supporting land-use planning and ecological protection. While remote sensing image segmentation automatically identifies target regions within complex scenes, existing methods often fail to balance semantic discrimination with boundary recovery due to intricate object morphologies and blurred edges. To address this challenge, this study introduces SSPCNet, a structure-aware semantic prototype-constrained network designed to enhance target recognition. SSPCNet introduces a structure-aware feature representation strategy to enhance the extraction of local textures, directional structures, and boundary-sensitive information, thereby improving the perception of complex object geometries. Furthermore, a semantic prototype-constrained decoder is developed to establish cross-layer semantic consistency by introducing shared semantic prototypes and prototype-constrained feature interaction, enabling effective integration of hierarchical features and improving the recognition of densely and sparsely distributed targets. Evaluated against eight representative segmentation methods across four public datasets using three independent runs with different random seeds, SSPCNet demonstrates strong and consistent performance. Notably, on the ISPRS Potsdam dataset, SSPCNet achieves an ACC of 85.23 ± 0.96%, an mIoU of 70.60 ± 1.40%, and a Harmonic Mean of 82.00 ± 1.01%, while requiring only 14.40 GFLOPs and 6.24 M parameters, demonstrating a favorable balance between segmentation performance and model complexity for remote sensing image interpretation.