Abstract
We present a generative modeling approach for nowcasting infrared (IR) brightness temperatures (Tb) from geostationary satellite observations (~ 10.8 μm) that couples a denoising diffusion probabilistic model with a 3D U-Net backbone. Using SEVIRI observations, the model ingests six hours of IR history and produces six-hour nowcasts at 15-min resolution. Deterministic evaluation is performed on an independent July-September 2022 test set and benchmarked against 3D U-Net, ConvLSTM, and Optical Flow extrapolation baselines. The diffusion model enhances prediction accuracy, yielding lower errors and higher correlation than all baselines across most forecast lead times, with a persistent advantage through the 2-hour lead time and reduced but still evident gains at longer leads. We complement the statistical assessments with a perceptual and a probabilistic diagnostic; diffusion achieves the highest SSIM at all leads and the lowest CRPS among all models (CRPS generates MAE for deterministic baselines). RAPSD analysis shows improved retention of high-frequency variance relative to deep learning baselines while avoiding the texture-only limitations of Optical Flow. Spatial maps demonstrate that our diffusion model significantly outperforms the baselines, indicating smaller errors over the study region. Two case studies show coherent structures in Tb forecasts, sharper gradients, and improved localization of evolving cold features. Overall, coupling diffusion with 3D U-Net delivers more structurally faithful satellite nowcasts than traditional extrapolation and deep learning baselines, particularly at short to intermediate leads.