Physical-Model-Guided Dual-Branch Generative Adversarial Network for Thin Cloud Removal

Remote sensing (RS) imagery is essential for Earth observation applications, yet its utility is often compromised by cloud contamination, which reduces image quality and hinders downstream analysis. Although the semi-transparency of thin clouds permits the possibility of single-image restoration, most existing deep learning approaches rely on paired training data from the same geographic area and offer limited interpretability. To address these challenges, we propose a novel unsupervised framework, termed physical-model-guided dual-branch generative adversarial network (PM-DBGAN), for thin cloud removal using unpaired cloudy and cloud-free images. The core of this model is a dual-branch generator with a feature interaction module that decomposes a cloudy image into transmission map, atmospheric light, and cloud-free radiance. These components are recombined via a physical model to reconstruct the cloudy input and generate a synthetic cloudy image, enforcing cycle consistency at both image and physical levels. By explicitly modeling and inverting the atmospheric degradation process, PM-DBGAN enhances interpretability and versatility, effectively recovering clear surface information under thin cloud cover. Experimental results on two benchmark datasets demonstrate that the proposed method outperforms several state-of-the-art methods, achieving superior restoration quality and generalization capability across diverse RS scenarios.