Precipitation Nowcasting using Diffusion Models on Satellite-Derived Radar Fields

Abstract

Accurate short-term precipitation prediction is essential for mitigating the hazardous impacts of severe weather events. Radar observations provide high-resolution measurements for convective activity, but their coverage is limited. Satellites can provide data in these uncovered areas, but do not contain inherent severity measures. This thesis investigates if generative diffusion models are able to bridge this gap by fore-casting radar reflectivity fields directly from multispectral satellite imagery. We construct a training dataset from 4 years of data with increased quantities of rain samples using importance sampling, keeping two seperate years meteorologically consistent for validation and testing. We propose a conditional denoising diffusion model with a convolutional UNet backbone and dedicated satellite encoder that fuses spatial, temporal and time-of-day information. We generate three consecutive radar fields given three preceding satellite images. Our model demonstrates realistic radar structures and competitive skill in standard verification metrics, especially when evaluating on ensemble forecasts.