Short-Range Forecasting of Flash Flood Warnings with Deep Learning

Abstract

Flash flood forecasting remains a major operational challenge because impacts developrapidly and are influenced by highly localized rainfall and hydrologic processes. This thesis investigates whether machine learning can complement existing forecast guidance by providing short-term, storm-scale probabilistic predictions of National Weather Service (NWS) flash flood warning (FFW) issuance. A U-Net convolutional neural network was developed to predict the likelihood of FFW issuance at 0-, 1-, and 2-hour lead times over the contiguous United States. Data included Multi-Radar Multi-Sensor (MRMS) precipitation, Flooded Locations and Simulated Hydrographs (FLASH) Coupled Routing and Excess Storage (CREST) maximum unit streamflow, High Resolution Rapid Refresh (HRRR) meteorological fields, land cover, and optional HRRR quantitative precipitation forecasts (QPF) from the warm seasons in 2021-2025. Verification results demonstrated strong regional variability in model performance, with higher skill in areas of frequent flash flood occurrence and reduced skill in lower-frequency regions. Spatial tolerance analysis showed that the models often correctly identified general regions of flooding but struggled to reproduce the precise spatial extent of FFW polygons, often exceeding the polygon bounds. Finally, improved metrics with alternative NWS flood products suggest that the models may be capturing a broader flood signal instead of strictly flash floods. Overall, these results indicate that machine learning can provide useful storm-scale probabilistic guidance for flash flooding. However, this prediction problem requires careful consideration of the data being used, as errors are strongly influenced by the limitations of FFW polygons as a training target. This suggests that future work should explore more physically representative labels, such as creating a flood/flash flood practically perfect dataset.