Machine Learning calibration of CAM ensemble probabilistic forecasts of convective mode with conditional probabilities of severe weather hazards using the OU MAP ensemble from HWT SFEs.

The OU MAP (University of Oklahoma Multiscale data Assimilation and Predictability) lab has been contributing real-time convection allowing model (CAM) ensemble forecasts to the annual Spring Forecasting Experiments (SFEs) in NOAA’s Hazardous Weather Testbed. Calibration and post-processing are essential aspects of the end-to-end generation of skillful probabilistic forecasts of severe convective weather. Encouraging results have been shown in past studies for calibrating CAM forecasts with machine learning techniques such as Logistic Regression (LR) and Gradient Boosted Regression Trees (GBRT). In particular, past studies have demonstrated that GBRT, with both storm-based and environment-based predictors, can provide highly skilled probabilistic forecasts for severe hail. An object-based probabilistic (OBPROB) post-processing technique has been demonstrated in recent SFEs as a method for visualizing quantitative ensemble probabilities of specific storm modes and storms representing specific hazardous weather threats (i.e., persistent mesocyclones, large hail, heavy rain, and strong winds). The goal of this study is to explore LR and GBRT as techniques for calibrating storm-based (rather than spatial neighborhood based) predictands in the OBPROB framework.

In this study, object-based probabilities (i.e., the probability of observing a similar storm mode nearby), and the corresponding conditional severe hazard probabilities given the occurrence of a similar storm mode, are calibrated using LR and GBRT models. Emphasis will be placed on the sensitivity of the calibrated forecast skill to different configurations of model parameters and predictors, as well as the training sample size, in order to guide the choice and configuration of statistical post-postprocessing of CAM ensemble forecasts for hazardous weather prediction.