MSFANet: A Multi-Scale Feature Alignment Model For Optical/SAR Change Detection

Optical and Synthetic Aperture Radar (SAR) images exhibit inherent differences in imaging mechanisms and feature representation, making it difficult to directly compare their features, which impacts the performance of cross-modal change detection. To address this issue, we propose the Multi-Scale Feature Alignment Network (MSFANet), which utilizes multi-scale features to enhance feature representation and achieve effective feature alignment. We introduce the Low-Dimensional Feature Fusion Module (Dual-Modal Attention Fusion Module, DMAF Module) and the High-Dimensional Feature Selection Module (Channel Ranker). DMAF Module employs an attention mechanism to weight and fuse low-dimensional optical and SAR features, and uses a global self-attention mechanism in the transformer to capture the relationships between cross-modal features, achieving spatial and semantic alignment of optical and SAR features. The Channel Ranker utilizes a Squeeze-and-Excitation mechanism to select important channels from high-dimensional features, reducing feature redundancy and improving the effectiveness of feature representation. Experimental results demonstrate that MSFANet outperforms other methods on multiple evaluation metrics, validating its effectiveness in optical and SAR change detection tasks.