Using AI/ML Techniques to Improve Quantitative Precipitation Estimation from the Multi-Radar Multi-Sensor System over the Western US

Quantitative Precipitation Estimation (QPE) has advanced with the integration of sophisticated AI and Machine Learning techniques. Previous research has demonstrated that the Multi-Radar Multi-Sensor (MRMS) system enhances QPE accuracy through the use of Convolutional Neural Networks (CNNs). In this study, we present new experiments aimed at further improving QPE accuracy specifically in the region covering Washington State.

The CNN model in this study processes spatial data from radar, environmental, and geographical variables to predict localized precipitation. Prior studies within the MRMS research group have shown significant improvements from the CNN over conventional methods, as evidenced by case studies and a year-long statistical analysis. The model is particularly effective at capturing spatial variability in regions with pronounced elevation gradients, where traditional approaches often fall short.

Preliminary results from training the CNN model with data from an atmospheric river event in Washington State in October 2021 indicate that the model's precipitation estimation accuracy improves with increased training data. Specifically, training with 48 hours of data and validating with 24 hours yielded better results compared to using 12 hours for both. Building on these findings, our ongoing research focuses on refining and validating the CNN model, which utilizes 29 variables, by using data from precipitation events in the Pacific Northwest. This involves data processing, adjusting the amount of training and validation data, and making final revisions to the model.

Currently, we are generating and analyzing simulations for 5-10 independent precipitation cases to document performance statistics and spatial distributions. Additionally, the CNN QPE performance will be compared with benchmark QPEs, such as the MRMS physically-based radar and multi-sensor QPE, to identify strengths and weaknesses and suggest future improvements. The training and validation datasets include key periods marked by significant Atmospheric River events in Washington State, such as those in October 2021; December 2022 to January 2023; and December 2023. The model is further validated against specific events, including those on November 12-14, 2021; November 28, 2021; December 18, 2021; November 5, 2022; April 9, 2023; October 18, 2023; and November 2, 2023. Initial results, including a detailed analysis of an Atmospheric River event in the Pacific Northwest, suggest that the CNN model significantly improves spatial sensitivity and prediction accuracy.