A Traceable Six-Band Sentinel-2 Cloud Removal Dataset With Diffusion-Mask-Driven Synthetic Augmentation

Abstract

Cloud removal models require paired multispectral observations, yet such data remain scarce, and existing synthetic methods often fail to preserve both realistic cloud geometry and band-wise spectral consistency. We present a traceable six-band Sentinel-2 benchmark dataset alongside a diffusion-mask-driven synthetic augmentation pipeline. The dataset contains approximately 73 000 paired patches from six globally distributed Sentinel-2 tiles, with a stratified in-domain test subset that supports evaluation across cloud-coverage levels. Spectral consistency filtering on clear pixels helps ensure reliable restoration targets. To mitigate data scarcity, we decouple geometric modeling (using a denoising diffusion probabilistic model (DDPM) to capture cloud structures) from radiometric rendering (via a band-sensitive scattering model). Evaluations show two complementary effects: additive synthetic augmentation improves in-domain real-data efficiency, while fixed-budget RGB mixed real/synthetic training improves cross-domain RICE performance. Curated real pairs remain important for spectral anchoring. The dataset and pipeline offer a reproducible benchmark for multispectral cloud removal under data-limited regimes.