Directional-Aware Dual-Branch Fusion Network for SAR Image Change Detection

Synthetic aperture radar (SAR) change detection (CD) is an essential yet challenging task in the field of ground surface observation. Existing transformer-based methods have promising context modeling capabilities for CD tasks, yet they struggle to preserve change-related details in high-resolution data. To deal with this limitation, a directional-aware dual-branch fusion network (DAFNet) is proposed based on a dual-branch backbone and a gated cross-fusion module (GCFM) for SAR image CD. First, an enhanced difference image (DI) is developed by integrating a log-ratio operator and A-law companding for high-quality training data generation. In the dual-branch backbone, a multidirectional convolution module (MDCM) is then constructed to extract high-frequency features comprehensively via multigradient Sobel operators with eight directions, which selectively captures details and contexts of images in a parallel manner with the Vision Transformer (ViT). Finally, the GCFM based on GeLU activation and cross attention is designed to fuse different dimensional features, enhancing related information against irrelevant interference. Experimental evaluations across three real SAR datasets verify the effectiveness of the proposed DAFNet in terms of visualization and quantitative indicators in CD.