Derek Rosendahl
Identifying Precursors to Strong Low-Level Rotation within Numerically Simulated Supercell Thunderstorms: A Data Mining Approach

Members - Faculty, students, and collaborators
News - Recent news and publicty from members of the IDEA lab
Theses and Dissertations - Publications and code releases for student theses and disserations
Publications - Recent technical papers and presentations
Software - Recent software releases


Despite considerable progress made in recent decades in the observation, modeling, and theoretical understanding of tornadoes, warning and forecasting their occurrence remains a considerable challenge. Quantitative statistics clearly show that warning probability of detection (POD) and lead time have plateaued in recent years, with false alarm ratio (FAR) remaining relatively constant, principally because existing surveillance radars and hazardous weather detection methodologies suffer from fundamental limitations that allow key meteorological quantities and associated features to go undetected. New advances will be required if substantial improvements in warning and forecasting accuracy are to take place.

One promising avenue is the use of a data assimilation procedure, in real-time, which is capable of increasing the number of meteorological quantities available for use in detection algorithms. Applying algorithms to these real-time, gridded analyses should provide a considerable advantage over existing techniques, which, in the case of tornadoes, mostly utilize radar radial velocity and reflectivity data in their native spherical-polar coordinates. However, using assimilated data rather than direct observations will necessitate the development of a new suite of algorithms able to operate on regular grids and accommodate retrieved fields. The performance of these algorithms will depend upon their ability to identify storm features and feature interrelationships prior to the development of a severe weather event (e.g., tornado).

This study provides an initial framework for identifying important features and feature interrelationships, intended for future hazardous weather detection algorithms, which herald the development of strong low-level rotation within deep convective storms. Numerical simulations were used instead of observational data because simulations provided all meteorological fields within a 3-D gridded structure, analogous to future assimilated analyses, and could generate a large number of storms under controlled conditions that utilize varying initial background environments. Assimilated data sets based upon observations could not be used because too few of these data sets exist, and it would be difficult to verify the retrieved fields of those available. The low-level rotation qualifier was used because tornadic vortices can not be resolved with the model grid spacing employed here, though as computational advances foster higher resolution simulations, the methodology may be extended to include smaller scale vortices.

A total of 1168 numerically simulated storms were generated within initial environments characteristic of supercell storms and were categorized by whether they produced strong, weak or no low-level rotation. A computational data mining procedure was developed to search the gridded fields for meteorological precursors, occurring in repeatable patterns, that lead to the development of strong low-level rotation. An analysis was performed on storms producing “strong” and “no” low-level rotation and a separate analysis was performed on storms producing “strong” and “weak” low-level rotation.

Our results identified sets of precursors, in the form of meteorological quantities reaching extreme values in a particular temporal sequence, unique to storms producing strong low-level rotation. Statistical analyses were performed on the sequences (termed rules) to identify their significance and the highest rated rules consisted of the same meteorological quantities with small variations in temporal ordering. This implied that the order in which quantities reached extreme values was less significant than the requirement that the quantities simply reach extreme values. With this in mind, frequency distributions of quantity occurrence in the top rated rules were generated, identifying the most important quantities reaching extreme values prior to the development of strong low-level rotation. The top five quantities identified in storms exhibiting “strong” and “no” low-level rotation were: maximum in vertical vorticity stretching below 2 km, minimum in baroclinic vertical vorticity generation below 2 km, minimum in vertical vorticity stretching below 2 km, minimum vertical perturbation pressure gradient force above 2 km and maximum vertical perturbation pressure gradient force below 2 km. The second analysis group (storms with “strong” and “weak” low-level rotation) generated rules with comparable quantities but less statistical significance. The large number of rules identified by this study should prove useful in the development of algorithms for anticipating strong low-level rotation in real-time 3-D gridded assimilated analyses.

Simulations with finer grids could be used to extend this concept to the tornado scale.


Derek Rosendahl. (2008). Identifying Precursors to Strong Low-Level Rotation within Numerically Simulated Supercell Thunderstorms: A Data Mining Approach.Master's Thesis, School of Meteorology, University of Oklahoma.

Papers from thesis

  • McGovern, Amy; Rosendahl, Derek H; Brown, Rodger A; and Droegemeier, Kelvin K. (2010) Identifying Predictive Multi-Dimensional Time Series Motifs: An application to severe weather prediction. To appear in Data Mining and Knowledge Discovery. [submitted version. Final version will be linked in when it is online. pdf 1.8 MB]



Created by amcgovern [at]

Last modified June 12, 2017 12:57 PM