RePIC-Seg: Reciprocal Prototype–Image Consistency-Driven Semantic Segmentation for Remote Sensing Images

Remote sensing image semantic segmentation is crucial for applications such as land monitoring and urban governance, yet remains challenging due to high interclass similarity, large intra-class variability, and complex high-resolution structures that often cause boundary leakage. To address these challenges, we propose a Reciprocal Prototype-Image Consistency-driven segmentation framework for remote sensing images, RePIC-Seg. Specifically, we design a RePIC Interactor layer that establishes reciprocal consistency interactions between class prototypes and image features, enabling stable prototype updating and effective feature refinement, thereby alleviating prototype drift and cross-class leakage. Furthermore, a Reciprocal Margin Loss (RML) is proposed to enlarge the margin between the ground-truth class and the most competitive negative class, further improving inter-class separability and boundary discrimination. Extensive experiments on the LoveDA, Potsdam, and Vaihingen datasets demonstrate that RePIC-Seg consistently achieves superior performance. In particular, it outperforms competing methods by 1.9% mIoU on LoveDA. The released code is available at https://github.com/Flyzhuhs/RePIC-Seg.