Machine Learning Predictions of Flash Floods

This study concerns the development, assessment, and use of machine learning (ML) algorithms to automatically generate predictions of flash floods around the world from numerical weather prediction (NWP) output. Using an archive of NWP outputs from the Global Forecast System (GFS) model and a historical archive of reports of flash floods across the U.S. and Europe, we developed a set of ML models that output forecasts of the probability of a flash flood given a certain set of atmospheric conditions. Using these ML models, real-time global flash flood predictions from NWP data have been generated in research mode since February 2016. These ML models provide information about which atmospheric variables are most important in the flash flood prediction process. The raw ML predictions can be calibrated against historical events to generate reliable flash flood probabilities. The automatic system was tested in a research-to-operations testbed enviroment with National Weather Service forecasters. The ML models are quite successful at incorporating large amounts of information in a computationally-efficient manner and and result in reasonably skillful predictions. The system is largely successful at identifying flash floods resulting from synoptically-forced events, but struggles with isolated flash floods that arise as a result of weather systems largely unresolvable by the coarse resolution of a global NWP system. The results from this collection of studies suggest that automatic probabilistic predictions of flash floods are a plausible way forward in operational forecasting, but that future research could focus upon applying these methods to finer-scale NWP guidance, to NWP ensembles, and to forecast lead times beyond 24 hours.