Development of a Multi-Sensor Multi-Radar (MRMS) Machine Learning QPE Approach over the Southwest US

The Multi-Radar Multi-Sensor (MRMS) quantitative precipitation estimation (QPE) products are an important tool for forecasters and other end users, especially in the desert southwest where rainfall is often sporadic. Precipitation estimation in the southwest US from conventional methods is challenging due to terrain impacts on radar observations along with the complex evaporative processes common to arid regions. In prior studies, a machine learning approach to precipitation estimation has been shown to improve on the biases seen from the physically-based MRMS QPEs over California. In order to assess the broader applicability of the machine learning approach to areas outside of California, here we test the convolutional neural network (CNN) model on several cases over a southwest US domain with the goal of finding an optimized model and training data setup. The performance of the CNN QPE is evaluated relative to 24-hour rain gauge data and compared to the existing MRMS QPE performance. There is improved performance from the CNN QPE when retraining the model over the new domain compared to the results seen when transferring the original California model to the southwest US domain. The modified loss function from previous work was shown to improve the consistency of performance across heavy and light precipitation events over the California domain. The same mean absolute error (MAE) and normalized MAE combined loss function approach is explored here with application of a delineation threshold specific to the southwest US domain. A particular focus is given to case study analysis of the model performance over the summer months when monsoonal rains are the predominant precipitation mode.