Enhanced-RSMamba: State space model for semantic segmentation of remote sensing images

Semantic segmentation of remote sensing images is vital for land use analysis and environmental monitoring but faces challenges in capturing long-range dependencies and multiscale features in high-resolution scenes. We propose an optimized MambaVision architecture that integrates 2D selective scanning with adaptive dilated convolutions to effectively capture local and global features. Our novel Multi-Frequency Multi-Scale (MFMS) decoder fuses frequency-domain features, derived from Discrete Cosine Transform, with multiscale spatial information and orientation-aware attention, improving texture and boundary detection. As the first application of MambaVision to remote sensing segmentation, our model achieves an mIoU of 84.24% on the ISPRS Potsdam dataset, a 6.57% improvement over baselines. Ablation studies confirm the effectiveness of 2D scanning and orientation attention, making our approach efficient for resource-constrained applications.