Towards Improving a Machine Learning Model for Excessive Rainfall in the Southwest United States

The Colorado State University-Machine Learning Probabilities (CSU-MLP) system is an operational model for excessive rainfall in the contiguous United States (CONUS), composed of eight regional random forest models that produce real-time probabilistic forecasts. CSU-MLP excessive rainfall forecasts are trained on 10+ years of output from NOAA’s Global Ensemble Forecast System (GEFS) Reforecast dataset, observations of excessive rainfall (24-hr ARI exceedance and flash flood local storm reports), and driven by the operational GEFS in real-time. Aggregate skill of CSU-MLP forecasts is lowest in the southwest (SW) region of the model (encompassing southeast California, southern Nevada and Utah, Arizona, western New Mexico, and southwest Colorado) due to low skill forecasts in association with the North American monsoon. It is hypothesized that the random forest model in the SW region is over-reliant on quantitative precipitation forecasts (QPF) from the GEFS and does not "use" other predictor variables (i.e., convective available potential energy, precipitable water, 2-m specific humidity) as heavily as it could. There is also a possibility that this region could see improvements by including additional hydrological variables. This work shows current progress towards improving CSU-MLP forecast skill in the SW region. Various predictor assemblies are tested, including training a model from reforecasts during the monsoon season only (as opposed to reforecasts from the entire year), withholding QPF as a training variable, and inclusion of hydrological variables (i.e., soil moisture, terrain).