Enhancing Production of Synthetic Radar Images from Geostationary Satellite Observations through Generative Diffusion Models

Abstract The limited coverage of radar sites has given rise to a demand for transforming the extensive coverage of weather satellite observations into high-resolution and accurate synthetic radar reflectivity imagery. In this study, we introduce a new method that utilizes generative diffusion models to address this challenge. Starting from pure noise, our diffusion model takes infrared images from the Himawari geostationary weather satellite and lightning observations from a ground-based network as inputs to control the generation process. The model’s iterative diffusion and denoising process helps capture the intrinsic uncertainty of satellite-to-radar transformation by generating probabilistic results, whereas nongenerative methods can only produce deterministic outputs. Our new technique improves the granularity and spatial accuracy of synthetic radar reflectivity imagery compared to previously published nongenerative U-Net models. In our experiments, the new technique enhances the emulation of severe weather by capturing finer visual structures in areas with strong radar echoes. Results show that images generated by our model outperform traditional U-Net models on key metrics such as the fractions skill score (FSS) across multiple thresholds, with the average FSS increasing from 0.40 to 0.50, and also produce a much improved statistical distribution of reflectivity, especially at the low and high ends of the distribution. Significance Statement Radar observations are essential for severe weather monitoring and nowcasting. However, the geographical limitations of radar coverage, particularly in Australia’s remote regions, present obstacles to providing radar-based warnings for high-impact weather events. This study introduces an innovative machine learning approach to transform widely accessible satellite infrared images and lightning observations into synthetic radar reflectivity images to overcome this limitation. Compared to previous methods, our approach improves the spatial resolution and accuracy of synthetic radar imagery, extending the applicability of severe weather nowcasting to regions beyond the reach of traditional radar systems. This advancement may improve severe weather monitoring and nowcasting in areas not covered by operational radar networks.