Developing an Ensemble ML Framework for Flash Flood Prediction

As part of an ongoing project in the NOAA Hydrometeorological Testbed, the Center for Analysis and Prediction of Storms (CAPS) at the University of Oklahoma (OU) is working to develop an ensemble ML forecast system for prediction of flash flooding using Convection Allowing Model (CAM) ensembles as input. This project builds upon prior efforts at CAPS developing U-Net ensemble forecasts for the prediction of rainfall and snowfall. The goal of this research is to produce forecasts of the likelihood of flash flooding in terms of probability of flash flood occurrence or probability of flash flood warning issuance with 36 hours of lead-time.

Forecasts are produced using an ensemble of U-Net machine learning models, trained on previous flash flood warnings and flash flood event data. ML model inputs are obtained from CAM forecasts produced by the High-Resolution Ensemble Forecast (HREF) and CAPS FV3-LAM ensemble members. The inputs include NWP predictions of QPF, radar reflectivity, wind, temperature, and moisture fields, as well as relevant information from other data sources describing land use, burn scars, and previous rainfall. Development efforts thus far will be presented, preliminary ML forecast results will be evaluated, and plans for future research, including the investigation of forecaster-focused ML tools such as WFO-specific models that take into account specific geographic and human factors which vary widely from WFO to WFO, will be discussed.