GOES2MRMS: Implementing hybrid Diffusion-Regression Models to predict Radar Reflectivity from Geostationary Satellites

Weather forecasters and hydrologists rely heavily on radar reflectivity data to estimate surface precipitation globally. However, radar coverage remains sparse, particularly over oceans, challenging terrains, and in developing nations. This limitation hinders accurate precipitation estimation and weather forecasting in many regions. Despite the consensus that geostationary satellites provide comprehensive radiance observations at high spatio-temporal resolutions, their potential for predicting radar reflectivity has not been fully explored. Furthermore, while machine learning techniques have shown promise in various meteorological applications, their use in bridging the gap between satellite radiance data and ground-based radar reflectivity fields remains understudied.

This work develops a machine learning approach to predict ground-based composite radar reflectivity (REFC) from the Multi-Radar Multi-Sensor (MRMS) observational dataset using Geostationary Operational Environmental Satellite (GOES) level 1 infrared and near-infrared radiance data. Our method combines a deterministic regression model (U-Net) with a generative diffusion model (incorporating spatial embedding; Karras et al., 2022) to facilitate this diagnostic prediction. The U-Net predicts REFC using GOES radiances, while the diffusion model further refines the prediction by using the regression-predicted REFC in its latent space while denoising the REFC field. Our model was trained using data from 2018-2022, validated on 2023 data, and tested on 2024 data. Spatially, we trained on the eastern US (100°W-80°W, 30°N-50°N) and performed inference on the entire CONUS, coinciding with MRMS data availability. Additionally, we conducted inference tests over open oceans (tropical Atlantic and Pacific), comparing our results against available radar data and GOES quantitative precipitation estimates (QPE). We demonstrate that our approach can reproduce REFC across the entire Continental United States (CONUS) with high accuracy for a range of precipitation events. When applied to out-of-sample GOES full disk data over the tropical Atlantic, our model demonstrates promising results, highlighting its potential for global applications. This observation-based approach has significant implications for both operational use and research applications, serving as a first step towards building an observational model without relying on physical priors from reanalysis or model datasets.

Our study underscores that significant advancement in REFC prediction is possible by leveraging diffusion and regression models for geostationary satellite datasets. However, further investigation is needed to improve the diffusion model capabilities to denoise such datasets both in space and time, as diffusion models tend to be over-dispersive. Future work will focus on adapting generative diffusion models for weather and climate problems, with a special emphasis on multi-modal observational datasets.

References

Karras, Tero, Miika Aittala, Timo Aila, and Samuli Laine. "Elucidating the design space of diffusion-based generative models." Advances in neural information processing systems 35 (2022): 26565-26577.