Cloud Removal in Remote Sensing Images Using Attention-Guided Dual-Stream U-Net with SAR

Remote sensing (RS) images are widely used for ecological and environmental monitoring. However, cloud cover often hinders direct acquisition of continuous, unobstructed RS images, which makes reconstruct clear RS images necessary. Traditional algorithms, however, struggle to fully exploit the information of multi-source data and face registration errors during fusion process. To solve these issues, we propose an attention-guided dual-stream U-Net encoding method by fusing cloud-covered optical images and SAR (Synthetic Aperture Radar) images in this work. The method employs dual-stream encoding for cloud-covered images and SAR images during the encoding phase, utilizing max pooling for down-sampling. Subsequently, we utilize a cross-attention block-guided fusion module for model integration, effectively harnessing the information from multi-source data and alleviating the minor misalignment between SAR and cloud-covered optical images. Finally, the designed decoder decodes the data, using Deconvolution for up-sampling, and ultimately outputs a cloud-free optical image. Compared to two mainstream SAR-guided optical image cloud removal methods, our proposed method significantly improves image restoration quality with about 0.5 reduction in RMSE.