DCAN-CR: A Dual Cross-Modal Attention Network for Cloud Removal Based on Optical and SAR Images

Cloud contamination in optical remote sensing significantly limits the application of high-resolution imagery in monitoring land surface changes. Synthetic aperture radar (SAR), which is insensitive to cloud cover and illumination conditions, provides complementary information under cloudy scenarios and serves as a valuable reference for optical image reconstruction. However, existing optical–SAR fusion-based cloud removal methods still struggle to preserve spatial details while maintaining spectral fidelity under complex conditions. To address these issues, we propose a dual cross-modal attention network for optical-SAR cloud removal (DCAN-CR), which aims to reconstruct cloud-contaminated optical images by leveraging auxiliary SAR data. The network captures deep cross-modal interactions through a dual cross-attention mechanism and enhances feature representation with a dense residual attention block. In addition, an SAR-derived edge prior is incorporated via learnable convolutional layers to facilitate the recovery of structural detail. An adaptive weighted composite loss function is designed to jointly enforce pixel-level accuracy, spectral consistency, and structural integrity. Experiments on the SEN12MS-CR dataset demonstrate that DCAN-CR outperforms four representative cloud removal methods across multiple metrics, achieving superior spectral fidelity and more refined structural reconstruction. Specifically, compared with the best-performing baseline, DCAN-CR improves peak signal-to-noise ratio by 3.7%, reduces RMSE by 12.4%, increases SSIM by 1.3%, and decreases spectral angle mapper by 19.1%. Ablation studies further confirm the effectiveness of each component and the proposed loss function. Overall, DCAN-CR effectively integrates optical and SAR information, providing a robust and reliable solution for reconstructing cloud-contaminated optical imagery.