CoMS-UNet: A Cohesive Multi-Scale Network for Semantic Reorganization and Scale-Aware Context Modeling in Remote...

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Semantic segmentation is crucial for detailed land-cover analysis in high-resolution remote sensing imagery.
• Existing methods struggle with multi-stage feature integration, scale-aware context, and cross-channel communication in decoders.

Purpose of the Study:

• To propose CoMS-UNet, a novel U-Net architecture designed to overcome limitations in current semantic segmentation techniques for remote sensing.
• To enhance feature representation by integrating semantic reorganization, scale-aware context modeling, and channel-aware decoder reconstruction.

Main Methods:

• CoMS-UNet employs a Multi-Scale Semantic Reorganization Module (MSRM) for cross-scale semantic alignment via feature fusion-separation-refusion.
• A Scale-Aware Module (SAM) at the bottleneck provides scale-relevant semantic guidance using adaptive receptive fields.
• A Channel-Spatial Shuffle Mamba Block (CSSMBlock) in the decoder facilitates cross-channel exchange and long-range dependency modeling.

Main Results:

• CoMS-UNet demonstrated strong performance on the ISPRS Vaihingen dataset, achieving a mean mIoU of 86.11±0.35%.
• The model also obtained an mIoU of 52.48% on the LoveDA dataset, indicating its effectiveness in diverse remote sensing scenarios.

Conclusions:

• CoMS-UNet offers a unified framework that improves semantic consistency, context aggregation, and detailed reconstruction for remote sensing semantic segmentation.
• The proposed architecture effectively addresses key challenges, leading to significant advancements in land-cover interpretation accuracy.