This project expands upon the work outlined in McGovern et al (2007) and Rosendahl (2007) in two important ways: changing the definition of storm cells and extending the data mining techniques. The overall goal of each of these projects is to identify key parameters for tornadogenesis by applying data mining techniques to a full set of meteorological fields. This significantly differs from the traditional approach of analyzing radar data containing only precipitation intensity and the radial wind component. Long term goals include lowering the false alarm ratio (FAR) and raising both the probability of detection (POD) and lead time for tornado warnings.

This paper introduces a new storm tracking algorithm that more reliably detects the critical features of a storm such as rear flank downdrafts and regions with high pressure perturbations by incorporating both updraft and reflectivity into the definition of a storm cell. The new storm cell definition improves the results presented in McGovern et al (2007) and Rosendahl (2007) and is designed to move toward a relational model that incorporates high level features such as rear flank downdrafts, updrafts, rain/hail regions and gust fronts and the relationships that exist between those high level features.

We also extend the data mining techniques in two ways. We first extend the rule-finding approach presented in McGovern et al (2007) and Rosendahl (2007) to include boosting. This significantly improves the results and gives us a more varied set of rules which can be used to improve our understanding of tornadogenesis. We also introduce preliminary results with a novel dynamic relational model: the dynamic relational probability tree. This model is designed to learn human-readable models of the critical storm processes and requires the new storm cell definition to identify the high-level features and relationships.