A Novel Cloud Detection Approach Integrating Mamba and Multi-scale Feature Fusion for Remote Sensing Images

Existing remote sensing cloud detection methods struggle with complex scenes and subtle feature extraction. To overcome these challenges, we introduce a U-shaped network enhanced by Mamba modules. The network employs a hybrid design of residual blocks and Mamba branches to balance local-global feature extraction. To strengthen decoder information transfer, spatial and channel attention mechanisms are integrated. Additionally, multi-scale skip connections are utilized to enhance cloud feature perception across scales. Experiments on 38-Cloud and WHUS2-CD+ datasets demonstrate that the proposed model achieves state-of-the-art performance. Compared with baseline methods, it improves precision by 3.58%, recall by 1.52%, IoU by 4.1%, and F1-score by 3.62% on the 38-Cloud dataset. On the WHUS2-CD+ dataset, it increases recall by 1.85%, IoU by 2.71%, and F1-score by 1.63%, significantly enhancing cloud detection reliability.