Cloud-EGAN: Rethinking CycleGAN From a Feature Enhancement Perspective for Cloud Removal by Combining CNN and Transformer

Cloud cover presents a major challenge for geoscience research of remote sensing images with thick clouds causing complete obstruction information loss while thin blurring the ground objects. Deep learning (DL) methods based on convolutional neural networks (CNNs) have recently been introduced to cloud removal task. However, their performance is hindered by weak capabilities in contextual extraction and aggregation. Unfortunately, such play vital role characterizing complex In this work, conventional cycle-consistent generative adversarial network (CycleGAN) revitalized from feature enhancement perspective. More specifically, saliency (SE) module first designed replace original CNN CycleGAN re-calibrate channel attention weights capture detailed multi-level maps. Furthermore, high-level (HFE) developed generate contextualized cloud-free features suppressing components. particular, HFE composed both CNN- transformer-based modules. The former enhances local employing residual multi-scale strategies, latter captures long-range dependencies Swin transformer exploit global Capitalizing SE modules, an effective Cloud-Enhancement GAN, namely Cloud-EGAN, proposed accomplish tasks. Extensive experiments RICE WHUS2-CR datasets confirm impressive Cloud-EGAN.