Self-organizing maps (SOMs) are artificial neural networks that can receive as input a large number of maps of environmental parameters [in our case, 240 x 240 km2 grids of Rapid Update Cycle (RUC) or Rapid Refresh (RAP) mesoanalysis data centered on the location of all 11,579 tornado events and the location of maximum STP within all 39,025 tornado warnings issued from 2003—2012] and can cluster them in such a way as to reduce the total variance within each cluster. The statistics within these clusters can then be calculated: is probability of detection higher within a particular cluster of maps than within the dataset as a whole? Are false alarm ratios higher? Are fatality or injury rates anomalously high? Are tornadoes produced by quasi-linear convective systems disproportionately common compared with those produced by right-moving supercells?
The examination of two-dimensional maps representing the mean environment of each clustered “node” can be used to identify not only bulk environmental parameters associated with, for instance, significant (EF2+) tornadic activity, but also to depict the two-dimensional structure of these environments. These analyses highlight the importance of the positioning of atmospheric boundaries such as drylines or fronts relative to the tornado, and also serve as a reminder that tornadic activity does not always necessarily simply correlate with a “bullseye” of a given parameter.
This proof-of-concept work highlights the value of moving beyond pinpoint bulk parameters to incorporate two-dimensional environmental heterogeneity and representations of meso-synoptic features into studies of the tornadic near-storm environment.