Evaluating a Machine Learning Approach Leveraging MRMS to Improve Quantitative Precipitation Estimation for Atmospheric Rivers

Oscar Chimborazo

Howard University

Tuesday, Mar 25, 2025, 2:00 pm MT

JOIN THE WEBINAR:

Abstract

Atmospheric River (AR) events can lead to extreme rainfall, increasing the risk of flooding, landslides, and challenges in water resource management. Accurate Quantitative Precipitation Estimation (QPE) during AR events is critical for hydrological decision-making, including flash flood warnings, streamflow forecasting, and water resource management. However, current radar-based QPE products from the Multi-Radar Multi-Sensor (MRMS) system face significant challenges in mountainous regions of the Western U.S. due to radar beam blockage, complex terrain, and microphysical uncertainties.

To improve precipitation estimation in these complex environments, a Convolutional Neural Network (CNN)—a machine learning approach—is applied to develop a model that leverages spatial correlations in precipitation fields by integrating radar data, atmospheric predictors, and topographic features. The model is trained using MRMS radar data, incorporating key input features such as radar reflectivity, orographic forcing, precipitable water, and climatological precipitation patterns to produce enhanced QPE.

Preliminary case studies of AR events over Washington State suggest promising improvements in QPE accuracy, with CNN-enhanced estimates demonstrating stronger agreement with gauge observations compared to operational MRMS products. However, this research remains ongoing, and further refinements are needed to improve the model’s robustness across different AR events. Current efforts focus on testing ensemble approaches, optimizing hyperparameters, and expanding the training dataset to enhance the model’s ability to generalize across diverse atmospheric and topographic conditions.

Bio - Dr. Oscar Chimborazo is an Assistant Research Scientist at Howard University in Washington, D.C., working as an NCAS-M fellow specializing in AI-driven improvements to radar-based Quantitative Precipitation Estimation (QPE). His research focuses on applying deep learning techniques, particularly Convolutional Neural Networks (CNNs), to enhance MRMS precipitation products in complex terrains, including the Western U.S. He has expertise in numerical modeling, high-performance computing (HPC), AI applications in meteorology, and Python programming.