Frequency-Spatial Domain Feature Fusion Network for SAR Building Extraction

Synthetic aperture radar (SAR) operates independently of weather and lighting conditions, enabling reliable performance in adverse environments such as clouds, fog, rain, and snow. This all- weather, day- and- night capability makes it highly valuable for building extraction applications. However, its unique imaging mechanism introduces challenges specific to SAR-based building extraction, including speckle noise. Additionally, persistent issues such as significant scale variations in target buildings and susceptibility to complex background interference remain problematic in SAR images. This paper proposes a novel SAR image building extraction network that utilizes wavelet transform to extract frequency domain features, which are then employed for denoising and enhancing spatial domain features. The network incorporates a frequency domain feature extraction module to capture multi-scale low-and high-frequency features, thereby improving the recognition capability for targets of varying sizes. By fusing these frequency features with spatial domain features according to their distinct spatial characteristics, the network enhances the structural integrity of extracted buildings while suppressing noise and background interference. Comparative experiments on the SpaceNet6 dataset demonstrate that the proposed model outperforms conventional building extraction methods, achieving a 75.22% intersection over union (IoU) for SAR image building extraction.