SARBA-Net: a boundary awareness network for impervious surface extraction of SAR images

The Impervious Surface (IS) is closely related to human activities and ecological environments, whose monitoring is vital for establishing safe, resilient, and sustainable human settlement environments. Synthetic Aperture Radar (SAR), with its timely, cloud-penetrating, and geometrically sensitive observational ability, provides a promising solution for large-scale, high-frequency IS extraction. However, accurately identifying the IS with clear boundaries in SAR amplitude images is a huge challenge due to speckle noise and the scattering heterogeneity of the IS. In response to the above challenge, we innovatively propose a Boundary Awareness Network (SARBA-Net) for IS extraction from SAR images. SARBA-Net employs a multi-scale Boundary Preserving Structure (BPS), which consists of multiple Boundary Awareness and Enhancement Modules (BAEMs) to progressively restore the boundary topology. Within BAEMs, a novel Boundary Contrast Loss (BCL) is designed to generate boundary class awareness, and a hard boundary anchor mining strategy further facilitates BCL to discriminate boundaries. Finally, a Boundary Consistency Auxiliary Loss (BCAL) is introduced in SARBA-Net to refine the boundary detail by the boundary duality between the ground truth and prediction. Extensive experiments on SAR images across diverse IS scenes, including dense urban, rural, road-dominated, and complex mountainous regions, show that SARBA-Net achieves remarkable IS segmentation performance in detail perception, shape preservation, and boundary refinement, with a 2.81% improvement in IoU and a reduction in boundary consistency error to 5.44%. These results demonstrate the effectiveness of SARBA-Net and highlight its potential for robust IS monitoring, particularly under cloudy and rainy conditions where optical imagery is limited.