Objective classification of supercell environments using multivariate self-organizing maps for research and forecasting

Previous work has shown that self-organizing maps (SOMs) are effective in classifying univariate environmental profiles derived from supercell proximity soundings obtained from the Rapid Update Cycle (RUC). The performance of SOMs in creating clusters that discriminate between significantly tornadic, weakly tornadic, nontornadic, and nonsupercell storms was dependent on the single variable chosen to create the SOM.

We will present advancements on this technique that allow a single SOM to consider multiple variables during the training algorithm. This multivariate approach allows us to objectively classify storm environments using hodographs and thermodynamic (combined temperature and dewpoint) profiles. Using an expanded dataset, we will compare the results of this advancement to prior work. Assessments will include: the ability of multivariate SOMs to discriminate between storm type for a variety of profile heights, methods for deriving conditional probabilities of supercell tornadoes based on hodograph or thermodynamic classification, forecast skill using cross-validation techniques, and comparison with existing forecast methods.

Self-organizing maps also reveal specific regimes of hodograph shape (particularly at low-levels) that are conducive or detrimental to supercell tornadoes. Using idealized simulations we will explore the physical basis for the effects of particular hodograph shapes on the tornadic potential of supercells.