"Sentinel-2 Cloud Removal Dataset"

"Developing robust cloud removal models is hampered by the scarcity of high-quality paired multispectral images and the limitations of existing synthetic alternatives to balance geometric realism and spectral accuracy. To address this, we present a traceable, six-band Sentinel-2 benchmark dataset alongside a diffusion-mask-driven synthetic augmentation pipeline. The dataset encompasses approximately 73,000 paired patches from six diverse global sites, stratified by cloud density. It applies rigorous spectral consistency filtering on clear pixels to ensure reliable restoration targets. To mitigate data scarcity, we propose an augmentation method that isolates geometric modeling, using a Denoising Diffusion Probabilistic Model (DDPM) to capture authentic cloud structures, from radiometric rendering via a physics-based scattering model. Evaluations reveal that while synthetic data cannot fully replace curated real pairs, it serves as a highly potent supplement. Under data-constrained conditions, this augmentation increases in-domain Peak Signal-to-Noise Ratio (PSNR) by up to 2.49 dB and cross-domain generalization by 2.14 dB."