Lightweight SAR building segmentation via enhanced multi-scale context modeling and bidirectional guided feature fusion

Accurate building extraction from synthetic aperture radar (SAR) imagery is important for urban mapping, disaster assessment, and remote-sensing interpretation, but remains difficult because speckle noise, heterogeneous scattering, radar shadow, weak texture, and ambiguous building boundaries can degrade pixel-level predictions. At the same time, heavy segmentation networks are difficult to deploy in computation-constrained settings. This study presents a task-oriented lightweight SAR building segmentation framework derived from DeepLabV3+. MobileNetV3 is used as an efficient backbone, a depthwise separable atrous spatial pyramid pooling (DS-ASPP) module retains multi-scale context at reduced cost, and a bidirectional guided fusion module (BGFM) combines boundary-sensitive low-level information with deep semantic features. On SARBuD1.0, the model achieves 94.64% OA, 79.89% MIoU, and 76.45% Kappa with 11.3 M parameters and 58.12 FPS. Across five independent runs, the proposed model achieves 79.86 ± 0.17% MIoU compared with 78.52 ± 0.20% for HSB-Net, with a significant paired difference ( p < 0.001). On HR-SARBuD, the model achieves 79.75% MIoU; the five-run mean is 79.77 ± 0.15%, compared with 79.68 ± 0.16% for FPANet, and the difference is not statistically significant ( p = 0.214). Under the batch-1 complexity protocol, the proposed network requires 8.9 GFLOPs and 17.21 ms per image, versus 42.7 GFLOPs and 50.89 ms for DeepLabV3+. These results support a favorable accuracy-efficiency trade-off, while cross-dataset and edge-device limitations are explicitly discussed.