Global-local uncertainty-contrastive physics-guided inversion network for unsupervised cloud removal.

Abstract

Recently, unsupervised cloud removal methods have been widely explored and have achieved significant progress. However, existing unsupervised approaches often lack effective constraints on reconstructed images, limiting their ability to produce high-quality cloud removal images. Moreover, the physical mechanisms that govern cloud formation have not been effectively incorporated into existing cloud-removal models, leaving them susceptible to color distortions and structural inconsistencies. To overcome these challenges, we propose a Global-Local Uncertainty-Contrastive Physics-Guided Inversion Network for Unsupervised Cloud Removal (GUPI-Net). Specifically, the proposed Bidirectional Global-Local Uncertainty-Contrastive framework (BGLUC) introduces an uncertainty-guided weighting mechanism that adaptively adjusts the relative importance of global and local contrastive losses based on the predicted uncertainty of each region, thereby enhancing detail restoration in high-uncertainty areas while reinforcing global consistency in low-uncertainty regions. Additionally, we design a Physics-Guided Inversion Module (PGIM) that reformulates the atmospheric scattering model in the feature domain by simultaneously estimating channel-specific atmospheric light priors and spatially varying transmission maps, and fusing these parameters through a physics-inspired inversion scheme, thereby forming a physically consistent representation aligned with the atmospheric scattering mechanism at the feature level. Finally, we construct a benchmark multi-surface remote sensing image cloud removal dataset, named CloudQuad. Experiments on both our proposed dataset and public datasets demonstrate that GUPI-Net outperforms several advanced unsupervised haze and cloud removal methods.