GUANet: Gaussian Uncertainty-Aware Network for Cloud Removal of Spaceborne Optical Images

In Earth’s surface observation, cloud coverage frequently impacts image quality, presenting substantial challenges for accurate analysis. Existing methods rely on static binary masks, which cannot capture changes in cloud thickness. It is hard to remove different shapes of clouds among different regions with the same loss function. As a result, accurately restoring land-surface structures and texture details under mixed-cloud conditions is challenging. Inspired by recent state-of-the-art approaches in probabilistic modeling and uncertainty learning, a Gaussian uncertainty-aware network (GUANet) architecture is proposed to explicitly address the stochastic uncertainties encountered during cloud removal. Additionally, the Bayesian pixel reconstruction loss (BPRL) and the spectral structural consistency loss (SSCL) are proposed. Specifically, BPRL adaptively modulates reconstruction weights according to pixel-level uncertainty, while SSCL preserves spectral fidelity and maintains edge sharpness. The proposed approach is evaluated on the RICE-I and RICE-II datasets. Experimental results demonstrate that the framework significantly outperforms existing models, achieving a PSNR of 34.59 and an SSIM of 0.9929 on RICE-I, and a PSNR of 33.50 and an SSIM of 0.9916 on RICE-II while maintaining color consistency and preserving fine structural details.