Severe storm cell classification using derived products optimized by genetic algorithms

Meteorological volumetric radar data is used to detect thunderstorms that are responsible for most severe summer weather. Discriminating between different storm types is a difficult problem, however. A radar data processing system conducts a volume scan by stepping a continuously rotating antenna through a series of elevation angles at regular intervals. Systems exist that allow meteorologists to focus their attention on regions of interest within the radar scan known as storm cells. When a cell is found, a number of derived features are computed. Each cell is also assigned a storm type label as determined by ground observers of the actual storm. Many of these products are computed using rules that incorporate threshold values. By varying these thresholds, the requisite features may be modified for each cell. We propose using a genetic algorithm to determine optimal values for the thresholds based on the labelling of the cells. We may evaluate the performance of the threshold values using linear discriminant analysis (LDA) on the derived features by comparing the computed and desired output of each cell. Once the optimized thresholds have been determined, we compare the classification performance of LDA, trained using the cell features generated using the new thresholds, against LDA trained with the original cell features.