Semantic-Geometric Consistency-Enforcing With Mamba-Augmented Network for Remote Sensing Image Segmentation

Remote sensing semantic segmentation (RSSS) remains challenged by semantic inconsistency, where spatially contiguous objects of the same category are fragmented into conflicting labels, and geometric distortion, manifesting as blurred boundaries and deformed shapes due to unmodeled directional anisotropy. To address these issues, we propose SGCMamba, a novel semantic-geometric consistency-enforcing network that leverages the Mamba architecture to synergize semantic coherence and geometric fidelity through three core innovations. First, our dual-path semantic-enhanced module (DPSM) integrates self-attention with adaptive dual-pooling to enforce intraclass label consistency by capturing long-range circumstantial dependencies. Second, the dual-dimension geometric-preserving module (DDGM) decouples height–width feature learning and employs multiscale kernels to preserve structural integrity and enhance boundary precision. Finally, the dual-axis Mamba-augmented module (DAMM) refines cross-consistency via spatially gated attention and dynamic channel reweighting, effectively suppressing error propagation while enhancing semantic consistency and contextual awareness. Comprehensive experiments and analysis have been conducted on three classical benchmark datasets: ISPRS Vaihingen, ISPRS Potsdam, and LoveDA, demonstrating SGCMamba’s superior performance in both accuracy and efficiency.