MCGFE-CR: Cloud Removal With Multiscale Context-Guided Feature Enhancement Network

Optical remote sensing imagery is often contaminated by clouds and cloud shadows, leading to the loss of ground information and limiting the application of optical images in fields such as change detection and object classification. Therefore, the removal of clouds and cloud shadows is one of the important tasks in the processing of optical remote sensing imagery. Currently, cloud removal methods with better performance are mainly based on Convolutional Neural Networks (CNNs). However, they fail to capture global context information, resulting in the loss of global context features in image reconstruction. The underlying architecture of Transformer networks is the attention mechanism, which can better capture global context features. Inspired by this, we propose a Multi-Scale Context-Guided Feature Enhancement Cloud Removal Network (MCGFE-CR), which can directly reconstruct cloud-free images from SAR and multi-cloud optical imagery. In MCGFE-CR, we embed a Multi-Scale Context-Attention Guidance (MSCAG) block, which can guide global and local context information at multiple scales into cloudy optical images. To enhance the global structural features after fusion and reduce the impact of SAR speckle noise, we incorporate a Residual Block with Channel Attention (RBCA). The network was trained and cloud removal was performed on a global dataset, and it was compared with the Hierarchical Spectral and Structural Preservation Fusion Network (HS2P), Deep Residual Neural Network and SAR-Optical Data Fusion Network (DSen2-CR), Global-Local Fusion Enhanced SAR Cloud Removal Network (GLF-CR), and Generative Adversarial Network for SAR Image to Optical Image Cloud Removal (GAN-CR). The method showed significant improvements in Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Peak Signal-to-Noise Ratio (PSNR), Spectral Angle Mapper (SAM), and Structural Similarity Index (SSIM). The MAE and RMSE were reduced by up to 0.0056 and 0.0129, respectively, while PSNR, SAM, and SSIM were increased by up to 3.8004, 3.2118, and 0.0385, respectively. The experimental results demonstrate that this method has higher spectral fidelity and richer structural texture information in reconstructing various types of ground information and optical images with different cloud coverage areas.