Use of a U-Net Architecture to Improve Microwave Integrated Retrieval System (MiRS) Precipitation Rates

We report on the implementation of a U-Net convolutional neural network (CNN) architecture to improve operational satellite retrievals of instantaneous precipitation rate (PR) from the NOAA microwave integrated retrieval system (MiRS). The U-Net architecture was implemented using NOAA-20/advanced technology microwave sounder (ATMS) passive microwave retrievals from the MiRS system. Training data consisted of input features that included operational retrievals of PR, total precipitable water, latitude, and longitude. Training target data (i.e., reference) were hourly PRs from the operational multiradar/multisensor system (MRMS) over the conterminous United States. (CONUS). The U-Net was trained using one year of collocated MiRS and MRMS data over the CONUS during 2021. Independent validation of U-Net was performed using data from 2022. Validation results showed that U-Net predictions were clearly improved relative to the original MiRS retrievals in terms of bias and root–mean-square error (RMSE), as well as categorical scores. The improvement mainly stemmed from a much better depiction of light rainfall distribution. Categorical scores such as the probability of detection (POD) and Heidke skill score were also significantly improved, as were aggregate error statistics. For instance, Heidke skill score and false alarm rate (FAR) improved from 0.42 to 0.50 and 0.057 to 0.014, respectively. Bias improved from 0.033 to −0.006 mm/hr. The spatial distribution correlation coefficient of the accumulated precipitation improved from 0.77 to 0.89. Once trained, the extremely low computational requirements of the U-Net model predictions highlight a potentially attractive means of improving operational retrievals of satellite PRs, where the latency of product dissemination is an important consideration.