Evaluating and Optimizing the MRMS Machine Learning QPE Performance over the Western CONUS

One of the main goals of the Multi-Radar Multi-Sensor (MRMS) system is to consistently provide high accuracy precipitation products for the various end users who depend on this information for flood forecasting, water resource management, and climate-related applications. This mission has been carried out mostly through the development of physically-based approaches for precipitation estimation, including radar-based and multi-sensor combined methodologies. In recent years there has been an effort to develop an MRMS machine learning quantitative precipitation estimation (QPE) product to augment the conventional methods in complex terrain areas. The high terrain results in radar blockages and orographic precipitation effects which make it challenging to maintain the accuracy of the physically-based QPEs. The machine learning QPE scheme under development is a convolutional neural network (CNN) model. This CNN model has shown consistent statistical improvement compared to the radar-based QPE products for a wide range of precipitation events over the western CONUS and Hawaii. There were incremental improvements in the CNN model’s representation of orographic and evaporative precipitation effects when additional input variables related to terrain and numerical weather prediction (NWP) model moisture information were added to the model. The work presented here focuses on further optimization of the CNN model and looks to improve on the consistency of performance across light and heavy precipitation events. A new normalized mean absolute error (MAE) loss function is tested and combined with the standard MAE loss function to help the model learn to better represent the full range of precipitation events. There were many high impact atmospheric river events affecting the western CONUS during the winter of 2022-2023 and the modifications to the CNN model will be tested for several of the events from this period. Efforts towards deploying the optimized version of the model in real-time will also be described.