A deep learning method for building semantic segmentation in SAR remote sensing images

Accurate building segmentation from synthetic aperture radar (SAR) images is important for all-weather urban mapping, change detection, and post-disaster assessment. However, speckle noise, side-looking imaging, scattering overlap, and geometric distortion often cause blurred boundaries, incomplete building regions, and confusion between buildings and background clutter. To address these problems, this study proposes an improved DeepLabV3 + -based framework incorporating a Gated Multi-Scale Edge Attention Module (GMEAM) and an Adaptive Scattering–Deformation Decoupling Module (ASDDM). GMEAM combines multi-scale contextual modeling with gated edge selection to enhance reliable building contours while suppressing noise-induced responses. ASDDM uses deformable feature sampling and foreground–background confidence modeling to adapt to irregular building geometry and reduce scattering-related ambiguity. Experiments on the public SARBuD1.0 and HR-SARBuD datasets demonstrate that the proposed method achieves favorable performance in region-level and boundary-level evaluation metrics. Ablation and visualization results further show that GMEAM improves boundary localization, whereas ASDDM strengthens structural recovery under geometric deformation and complex scattering interference. The proposed framework provides an effective approach for binary building extraction in complex SAR scenes.