CoMS-UNet: A Cohesive Multi-Scale Network for Semantic Reorganization and Scale-Aware Context Modeling in Remote Sensing Image Segmentation
Semantic segmentation is essential for fine-grained land-cover interpretation from high-resolution remote sensing imagery. However, existing methods still show limitations in coordinating multi-stage encoder features, providing scale-aware contextual guidance at the bottleneck layer, and facilitating cross-channel information exchange during decoder reconstruction. To address these issues, this paper proposes CoMS-UNet, a cohesive multi-scale U-Net that integrates semantic reorganization, scale-aware context modeling, and channel-aware decoder reconstruction. In encoder–decoder skip connections, the Multi-Scale Semantic Reorganization Module (MSRM) reorganizes multi-stage encoder features through a fusion–separation–refusion process to achieve cross-scale semantic alignment. At the bottleneck stage, the Scale-Aware Module (SAM) constructs predefined receptive-field branches and adaptively weights their contextual responses to enhance scale-relevant semantic guidance. During decoder reconstruction, the Channel-Spatial Shuffle Mamba Block (CSSMBlock) promotes cross-channel information exchange and long-range spatial dependency modeling to improve structural detail recovery. Through this unified feature-processing framework, CoMS-UNet strengthens semantic consistency, scale-aware context aggregation, and channel-aware decoder reconstruction. Experiments on the ISPRS Vaihingen dataset and the labeled LoveDA validation set show that CoMS-UNet achieves a mean mIoU of 86.11±0.35% over three independent runs on Vaihingen and an mIoU of 52.48% on LoveDA.