Comparing Predictor Influence for a Machine Learning Model in the Southwest United States for Three Different Excessive Rainfall Events

The Colorado State University-Machine Learning Probabilities (CSU-MLP) system is a collection of random forest models that produce real-time probabilistic forecasts for excessive rainfall in the contiguous United States (CONUS). CSU-MLP excessive rainfall forecasts are trained on numerical output from the GEFS Reforecast dataset, with eight separate models trained on different regions of the CONUS and designed as a "first-guess" outlook to assist forecasters at NOAA's Weather Prediction Center (WPC) when creating the Excessive Rainfall Outlook (ERO). 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) largely due to low skill daily forecasts in association with the summer monsoon.

This work compares how predictor variables (i.e., QPE, CAPE, PWAT) for a convective monsoon case (19 August 2022) contribute to resulting forecast probabilities relative to two higher predictability events: a wintertime atmospheric river (9 January 2023) that affected the western portion of the SW region, and a landfalling tropical cyclone (Hilary; 20 August 2023) that affected southern California, Nevada, and western Arizona. Specifically, the tree interpreter package in python is used to calculate whether predictors provide positive contributions (I.e., increase the forecast probability) or negative contributions (decrease the forecast probability) to forecasts for excessive rainfall. Additionally, predictor values are compared relative to the training distribution to provide further insights into the random forest model. Results are used to direct future improvements to the CSU-MLP SW regional model.