Authors: Hari Dhital, Sujan Maharjan, Bong-Chul Seo – Missouri University of Science and Technology; Witold F. Krajewski – The University of Iowa; Wendy Pearson, Scott Dummer – Missouri Basin River Forecast Center, NOAA
Title: Evaluation of Deep Learning-Based Short-Term Rainfall Forecasting for Streamflow Prediction
Presentation Type: Poster Presentation
Abstract: Rainfall-driven hazards such as pluvial flooding require accurate rainfall forecasts issued with sufficient lead time to support flood monitoring, warning, and emergency response. However, quantitative precipitation forecasts (QPFs) from numerical weather prediction (NWP) models contain uncertainties associated with model initialization, which can lead to errors in the timing, spatial organization, and intensity of rainfall systems. These rainfall uncertainties can propagate through hydrologic models and influence streamflow predictions, often exceeding errors from other modeling components such as rainfall-runoff models and quantitative precipitation estimates (QPEs). In the United States, short-range streamflow forecasts from the National Water Model rely on NWP-based High-Resolution Rapid Refresh (HRRR) model, which can exhibit errors such as spatial displacement and disorganized precipitation structures.
This study evaluates deep learning-based approaches for short-term rainfall forecasting, with emphasis on their ability to represent both storm motion and rainfall evolution, including growth and decay. Although radar-based extrapolation methods (e.g., optical flow) have been widely used for nowcasting, recent advances suggest that deep learning-based approaches can better represent complex storm motion and evolution. We use Radar-derived QPE, Multi-Radar Multi-Sensor (MRMS) as the primary input to generate high-resolution rainfall forecasts up to 3 hours over the Kansas City Metropolitan area, where a dense rain gauge network supports flood monitoring and warning. We evaluate deep learning-based forecasts against established nowcasting methods such as optical flow and LINDA, as well as HRRR. In addition to rainfall forecast evaluation, the study performs a watershed-based analysis over three selected small watersheds of approximately 150 km². This analysis examines how rainfall forecast errors propagate through the NextGen modeling procedures.