Using the New Fourier Neural Operator and Real-Time GOES-R Satellite Data for Improved Precipitation Retrievals in the Southern Great Plains

In the Southern Great Plains (SGP) region, mesoscale convective systems (MCSs) are highly active and hard to predict during the summer months. Compared to the Northern Great Plains, MCSs in the Southern Great Plains exhibit higher precipitation intensity and greater total precipitation amount. Accurate retrieval of these storm-related precipitation enhances the precision of hydrological models used in water resource management, which has a significant impact on water resource availability. Current methods for retrieving precipitation totals rely on data from radar analysis and in-situ observations, such as rain gauges. The distribution of these observation tools is uneven across the United States, complicating accurate public warnings. Geostationary satellites, in contrast, possess unparalleled capabilities to capture atmospheric data for entire continents at high spatial and temporal resolutions.

Recently, machine learning (ML) models have been employed as a data-driven approach to quantify precipitation from severe storms. In this research, we evaluate the performance of a ML model using the new Fourier Neural Operator (FNO) for obtaining hourly precipitation retrievals in the Southern Great Plains region. The FNO-based model utilizes water vapor-absorbing band brightness temperatures, lightning flash counts, and lightning average flash areas from NOAA's latest generation of Geostationary Operational Environmental Satellites (known as GOES-R) as inputs to produce hourly precipitation retrievals at approximately 4-km horizontal resolution. The “ground truth” rainfall data are hourly NCEP Stage IV precipitation analysis totals, a product typically finalized after 12-18 hours. Utilizing the high-accuracy of our “ground truth” data, and the low-latency satellite products, the proposed ML architecture demonstrates the ability to produce reasonable precipitation retrieval totals with a fraction of the lag time of the NCEP Stage IV precipitation analysis. Our results show improvements over the operational GOES-R Quantitative Precipitation Estimation (QPE) product in both precipitation estimation and detection at hourly intervals.