24AI Score based diffusion nowcasting of GOES imagery

Data-driven weather forecasting has disrupted the weather forecasting space, enabling skillful and timely forecasts of many atmospheric quantities at a fraction of computational cost. Some of the latest advancements in data-driven forecasts include the use of generative AI models (e.g., diffusion models) to enable generating both sharper images and ensemble forecasts. Here we leverage the latest diffusion techniques, like the ones used in Google’s GenCast and NVIDIA’s CorrDiff, to generate nowcasts of satellite imagery, focusing on the Geostationary Operational Environmental Satellites (GOES). Our results show that score-based diffusion models outperform a standard U-net and persistence on several pixel-wise metrics (e.g., RMSE) as well as newer sharpness metrics (e.g., gradient metrics) on 1-hour nowcasts of infrared GOES imagery. We also explore the use of open-source pretrained autoencoders that compress the GOES imagery into a latent space to enable shorter training times and more timely nowcasts.