Unsupervised machine learning of tornado-producing storms in the southeastern United States

The east-southeastern US is uniquely affected by storm and tornado-related damages, costs, injuries, and deaths. Based on doppler radar, satellite, and modeled data, previous research sought to understand these different types of storms that produce strong tornadoes. Many approaches to storm classification are time intensive, complex, and vary significantly across the literature. The purpose of this work is to (1) explore the radar-derived data structure and spread of strong tornado-producing mesoscale storms in the east-southeastern US; (2) use K-Means unsupervised machine learning methods to elucidate clusters (storm types) and clustering attributes; and (3) assess the utility of K-Means as a storm typing algorithm. Convective and stratiform strength, length, width, shape, and area, as well as the position of stratiform rain relative to convection, were evaluated using principal component analysis and K-Means clustering. The results show that convective strength and length attributes best explained variance in the dataset and minimized cluster overlap. While other attributes wielded explanatory power, they did not separate into distinct clusters. After testing K-Means on 2 through 8 clusters, the intensity and shape attributes generated three strong tornado-producing storm types: Weaker Linear (n=146), Stronger Cellular (n=222), and Weaker Cellular (n=292). These storm types were most like those in the literature and yielded the highest K-Means Silhouette score (0.41). It was found that 54% of storms were placed with less than 80% certainty, which emphasizes that storm types belong on a continuum and are not easily divided into groups. However, climatological analyses of the storms from 2000–2020 reveal that Weaker Linear, Stronger Cellular, and Weaker Cellular storm types produce similar spatial and temporal signals to those of seminal storm classification literature. This work can provide a framework for future storm taxonomies and highlights the value of machine learning in place of subjective, manual storm classification.