Enhancing Semantic Segmentation of Man-Made Structures Using Multiband SAR Imagery

Xinchen Li, Liang Guo, Xinrui Jiang, Kai Wei, Xiaohua Feng, Dan Jing, Mengdao XingIEEE GRSL 2025 · DOI

Abstract

The utilization of multi-band synthetic aperture radar (SAR) imagery holds considerable potential for distinguishing man-made structures. In particular, high-frequency band SAR provides more detailed information, while low-frequency band SAR offers higher penetration capabilities, making it beneficial for delineating the boundaries of man-made structures. However, since obtaining sufficient data for research is challenging, there are very few related studies. In this paper, we propose a method for enhancing the semantic segmentation accuracy of man-made structures using multi-band SAR imagery. Specifically, we use the DeepLabV3+ network for semantic segmentation. The network’s input consists of three polarization channels from S-band and Ka-band data, along with the extracted polarization difference. In experiments, we compare the segmentation accuracy achieved under various input scenarios. Results demonstrate that the combination of S and Ka-band SAR imagery preserves their distinct advantages for the segmentation of man-made structures, while incorporating polarization difference information further enhances segmentation performance.
BibTeX
@article{li2025enhancing,
  author = {Xinchen Li and Liang Guo and Xinrui Jiang and Kai Wei and Xiaohua Feng and Dan Jing and Mengdao Xing},
  title = {{Enhancing Semantic Segmentation of Man-Made Structures Using Multiband SAR Imagery}},
  journal = {IEEE Geoscience and Remote Sensing Letters},
  year = {2025},
  doi = {10.1109/lgrs.2025.3647274},
}

Loading related content...

Figure