Real-Time and Climatological Storm Classification Using Machine Learning

We present a machine-learning based system to classify storms across the Continental United States (CONUS). Following the classifications developed by Thompson et al (2012), we classify storms across CONUS into six categories. These classifications can be used in real-time by forecasters or retrospectively to create climatologies of storms in certain areas.

Labels for the training data were provided by Smith et al (2012) and Obermeier. Storm characteristics were drawn from radar data provided by the Multi-Year Reanalysis of Remotely Sensed Storms (MYRORSS, Ortega et al. 2012). Near Storm Enviroment (NSE) data was also created by interpolating soundings from RUC/RAP archives to the storm location using SharpPy (Blumberg et al, 2017).

This talk is part 1 of two highlighting two different machine learning approaches to classification. In this talk, we focus on tree-based classification systems including Random Forests (Breiman, 2001) and Gradient Boosted Classifiers (Friedman 2002). These systems achieve a preliminary Peirce Skill Score of 0.54 when trained and tested on data from 2011. We are currently augmenting the training set to include 2008-2010.

A preliminary version of this system was tested in real-time in NOAA’s Hazardous Weather Testbed (HWT) in the summer of 2017 and specifically within Probabilistic Hazard Information (PHI) (Karstens et al. 2015).