P2.2 On the use of kernel density estimation to identify severe weather events

Monday, 11 October 2010
Grand Mesa Ballroom ABC (Hyatt Regency Tech Center)
Chad M. Shafer, University of Oklahoma, Norman, OK ; and C. A. Doswell III

In previous studies, the identification and ranking of severe weather outbreaks included the use of the middle-50% parameter, a variable describing the degree to which severe weather reports were scattered geographically across a specified domain. The middle-50% parameter is defined as a latitude-longitude area, in which the middle 50% of the latitudes and the middle 50% of the longitudes of the severe reports in a given 24-h period were retained. The area was calculated as the product of the differences between the 25th and 75th percentiles of the latitudes and longitudes. The use of this parameter effectively identified cases with large geographic scatter. However, days with spatially distinct clusters of severe weather reports were also identified using this technique. This is undesirable, as subjective perceptions of these days suggest that each cluster should be represented as a separate event.

This study introduces a new approach to identify severe weather events using kernel density estimation. Grids were developed covering the conterminous United States, and the reports for a specific 24-h period on each day from 1960-2008 either were converted to the grid before the method was implemented or were computed as distances from each grid point directly. Several types of map projections, using various grid spacings, were tested. Identification of severe weather clusters required subjective tuning of the bandwidth and the approximate probability density function threshold. Results indicate that kernel density estimation effectively distinguished spatially distinct clusters of severe reports as separate events and was relatively resistant to changes of grid spacing and map projection, as expected. After removal of severe weather clusters with relatively few reports or sparse coverage within the region associated with the cluster, the remaining cases were ranked and classified using a linear-weighted, multivariate index, as in recent studies. The new rankings and classifications of outbreaks were consistent with previous studies.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner