Optimization of a Machine Learning Multi-Radar Multi-Sensor (MRMS) Precipitation Estimation Approach over the Western United States

Abstract Multi-Radar Multi-Sensor (MRMS) system precipitation products are relied upon for a wide range of important downstream applications relating to hydrometeorological impacts over different temporal and spatial scales. There are several challenges inherent to physically based precipitation estimation in complex terrain areas: observational siting issues, radar beam blockage effects, and highly variable rain-rate relationships in orographic forcing scenarios. A prior paper from the MRMS research group detailed the initial development and testing of a deep learning convolutional neural network (CNN) model to augment current precipitation estimation capabilities over the western United States. This study describes recent modifications made to improve the CNN model consistency of performance across event types. Additional numerical weather precipitation (NWP) model and terrain-related input variables have been added to the original 13 radar input variables, allowing the model to better capture warm rain processes and evaporative effects. Also, a custom loss function is applied to alleviate an overestimation bias seen in lighter precipitation events related to the use of the mean absolute error (MAE) loss metric in the initial model version. The updated model was tested over the year of 2021 along with a 15-day period with several high-impact atmospheric river events from the winter of 2022–23. Gauge-based evaluation of the 24-h quantitative precipitation estimation (QPE) fields shows that the new CNN model consistently outperforms both the MRMS radar-based QPE and the original version of the CNN model, building confidence in the eventual adoption of the CNN QPE within the MRMS operational framework. Significance Statement In this work, we look to refine a deep learning model for precipitation estimation in complex terrain areas with a focus on the western United States where observational limitations caused by the higher terrain inhibit physically based approaches. The initial version of the model showed improved accuracy relative to conventional methods. In the latest version of the model presented here, modifications were made to the input fields and loss function resulting in better consistency of performance compared to the initial version of the model. These results indicate that deep learning can likely provide critical improvements to the fidelity of estimated precipitation fields in areas with challenging terrain.