Self-Organizing Maps in the Near-Storm Environment: Toward Improving Tornado Forecasting

Self-organizing maps (SOMs) are a subset of artificial neural networks that use unsupervised learning to map input data onto nodes without losing the topology of the input space. In this work, the near-storm environments of over 11,000 United States tornado reports from January 2003 through June 2013 are characterized using mesoanalysis data based on the Rapid Update Cycle (RUC) or Rapid Refresh (RAP) models. We make use of SOMs to cluster into nodes the two-dimensional patterns of several environmental parameters of particular interest to severe weather forecasters, such as surface temperature, convective available potential energy, storm-relative helicity, lifting condensation level, and bulk wind difference. The clusters corresponding to each node are then characterized according to data such as probability of detection (POD), tornado warning lead time, time of day, time of year, geographical location, EF-scale value, as well as the morphology of the parent storm. By identifying and clustering the particular near-storm environments that lead to, for instance, a significantly higher percentage of nocturnal tornadoes, we can potentially improve our understanding of tornadic near-storm environments.