Diffusion-Based Cloud Removal Under Systematically Varied Cloud Densities in Sentinel-2 Imagery

Cloud obstruction remains a major challenge for the operational use of optical satellite imagery. Recent studies have demonstrated promising results with diffusion-based generative models for cloud removal and geospatial scene reconstruction. However, the robustness of these models under systematically varied cloud densities remains insufficiently understood. In this study, we present a systematic evaluation of a custom diffusion model for cloud removal using a continuous cloud-coverage dataset derived from the SEN12MS-CR dataset. The dataset is stratified into low (0–30%), medium (30–70%), and high (>70%) cloud-density subsets to enable a structured assessment of model performance. Quantitative evaluation is complemented by qualitative inspection. Comparative experiments against state-of-the-art baselines, including SSGAN and DiffCR, highlight the strengths and limitations of diffusion-based reconstruction under increasing cloud obstruction. Results indicate that diffusion models maintain high reconstruction fidelity under low to medium cloud coverage but exhibit progressive blurring and uncertainty under dense cloud conditions. This study focuses on controlled single-image reconstruction and is not intended to replace operational multi-temporal or multi-sensor remote sensing pipelines.