Identification of severe weather outbreaks using kernel principal component analysis

Abstract

A new adaptive approach to severe weather outbreak compositing and discrimination is described for datasets of known non-tornadic and tornado outbreaks. Kernel principal component analysis (KPCA) is used to reduce the dimensionality of the dataset and provide input for cluster analysis (CA) of the outbreaks to discern meteorological characteristics unique to each outbreak type. Results are compared to traditional principal component analysis (PCA). The KPCA methodology and CA assigned outbreaks to different composite (maps that have a close correspondence) sets than did PCA and CA. The clusters associated with each method were used as training for a support vector machine classification scheme. An independent subset of the outbreak dataset was retained for cross-validation classification of outbreak type. Significant differences in the two composite methods are observed, and a support vector machine classification scheme demonstrates compelling effectiveness in distinguishing outbreak types based on the resulting composites.