SSDT: Scale-Separation Semantic Decoupled Transformer for Semantic Segmentation of Remote Sensing Images

I. Overview

Remote-sensing scenes combine large geographic structures with small, densely arranged objects. SSDT uses transformer components to separate scale-sensitive representations from semantic context and then coordinate them for pixel-level prediction.

II. Key Contributions

• Introduces scale-separation modules for heterogeneous object sizes.
• Decouples semantic interactions to reduce interference among scene categories.
• Uses transformer attention to integrate long-range context into segmentation.

III. Methodology

Multi-level visual features enter scale-separation and semantic-decoupling stages. Transformer attention models interactions within the refined representations, and a decoder fuses the resulting features to produce the final segmentation map.

IV. Research Focus

The method is intended to preserve fine spatial detail while using broad scene context, especially when the same semantic category appears at markedly different scales.

@article{zheng2024ssdt,
  title={SSDT: Scale-separation semantic decoupled transformer for semantic segmentation of remote sensing images},
  author={Zheng, Chengyu and Jiang, Yanru and Lv, Xiaowei and Nie, Jie and Liang, Xinyue and Wei, Zhiqiang},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing},
  volume={17},
  pages={9037--9052},
  year={2024},
  publisher={IEEE}
}