LGE-CR: Integrating Nonlocal Gated Attention and Two-Flow Residual Dense Learning for Cloud Removal

Optical remote sensing images often suffer from information loss due to cloud cover. This problem can be alleviated by employing synthetic aperture radar (SAR) and optical data fusion to remove clouds based on deep learning. However, the different imaging mechanisms of SAR and optical images increase the complexity of fusion and hinder the effectiveness of cloud removal. To address this, we propose a two-flow residual dense block (TFRDB) and a nonlocal gated attention block (NL-GAB) within the network of locally globally enhanced cloud removal (LGE-CR) to enhance model characterization capabilities. The TFRDB focuses on extracting local features through densely connected convolutional layers and integrating hierarchical features from the auxiliary branch into corresponding feature layers of the main branch, enabling the adaptive learning of efficient features from the local hierarchical features of SAR and optical images. In addition, the NL-GAB selectively captures long-range dependencies to enhance the network’s global learning capabilities without introducing unnecessary cloudy information. The efficacy and superiority of the proposed method are demonstrated on the SEN12MS-CR dataset, achieving improved visual effects and accuracy metrics (PSNR =32.0375, SSIM =0.9214, SAM =2.0145, and MAE =0.0228).