Convective Meteorology (Mesoscale Dynamics)

Machine Learning Predictions of Flash Flooding

Race Clark

School of Meteorology

15 April 2016, 2:00 PM

National Weather Center, Room 5600
120 David L. Boren Blvd.
University of Oklahoma
Norman, OK

Machine learning techniques allow computers to sift through large amounts of data in order to detect patterns useful for making predictions. In this study, machine learning techniques (primarily “random forests”) are applied to a lengthy archive of outputs from numerical weather prediction models (primarily the Global Forecast System) to characterize the environments in which dangerous flash flooding events occur. This prediction framework will enable forecasters to quickly identify areas worthy of additional attention for local flash flood forecasting. At analysis time, this method correctly identifies more flash flooding events with marginally fewer false positives than the equivalent operational methods, where deterministic quantitative precipitation forecasts (QPFs) are compared to flash flood guidance (FFG) values. Unlike QPF/FFG comparisons, this method also enables probabilistic forecasts of flash flood environments. Results of these forecasts for two cases in the Southern Plains are presented, as well as statistics that summarize the performance of the technique on a national scale over multiple years. In general, raw probabilities (or “scores”) from the random forest technique can be successfully calibrated to more accurately reflect the true probability of historic flash flooding events as recorded in the Storm Data publication. Finally, examples of real-time global flash flood predictions using machine learning techniques will be discussed.

Convective Meteorology (Mesoscale Dynamics) Seminar Series website