HCV-CVAE: A hierarchical convolutional variational transformer for thin cloud removal in remote sensing imagery

To address the challenge faced by existing thin-cloud removal methods in balancing global structure reconstruction and local texture restoration under complex cloud conditions, this paper proposes a remote sensing image de-clouding approach based on a Hierarchical Convolutional Variational Vision Transformer (HCV-CVAE). Built upon the conventional CVAE framework, the proposed model introduces an HCV-ViT encoder that integrates the strengths of convolutional networks and Transformers to enhance local texture representation while capturing global semantic dependencies. Furthermore, strategies such as KL-divergence annealing, cross-dimensional weighted mutual information loss, and test-time augmentation are incorporated to improve the stability of the latent space and the robustness of the generation process. The proposed approach exhibits superior performance over existing algorithms on the RICE2 and T-Cloud datasets, with the highest PSNR and SSIM reaching 40.93 dB and 0.9872, respectively. The HCV-CVAE effectively restores fine details and spectral characteristics beneath clouds while maintaining global structural consistency, exhibiting significant advantages in both visual quality and quantitative metrics.All implementation code and pretrained models are publicly available at:https://github.com/Kyperio/HCV-CVAE.