Exploring the Treatment of Predictors for Forecasting Excessive Rainfall with Random Forests Based on a Deterministic Convection-Allowing Model

Machine learning applied to high-resolution numerical weather prediction forecasts is gaining popularity as a prediction tool for high-impact weather events, but a systematic examination of the role of spatial and temporal aggregation of predictor information is lacking. Here, we describe sensitivity experiments exploring the impact of a number of predictor assembly approaches upon a random forest (RF) system for excessive rainfall prediction using a deterministic convection-allowing model. Forecasts are significantly improved when the system uses hourly model predictors rather than three-hourly predictors, and when the RF uses data from all nearby model gridpoints, rather than just sparse input points. Forecast improvements, with a more than doubling of the Brier skill score, are greatest in the interior southwestern US, and during the months of July and August, mirroring the spatial and temporal domain of the North American monsoon (NAM). This reflects the relatively small spatial and temporal scale of NAM heavy rainfall events (as well as their environmental indicators). Examination of predictor contributions with the tree interpreter algorithm reveals that using spatial aggregation allows the RF to obtain more information from storm attribute variables like accumulated precipitation for these small-scale events, and use of a shorter predictor time step allows the RF to learn signals associated with hourly changes in environmental fields that are relevant for heavy rainfall. These results hold promise for improving excessive rainfall prediction in the challenging NAM environment, where damaging flash floods are missed by the operational excessive rainfall outlook more often than in any other region.