Rule-based support vector machine classifiers applied to tornado prediction

Abstract

A rule-based Support Vector Machine (SVM) classifier is applied to tornado prediction. Twenty rules based on the National Severe Storms Laboratory's mesoscale detection algorithm are used along with SVM to develop a hybrid forecast system for the discrimination of tornadic from nontomadic events. The use of the Weather Surveillance Radar 1998 Doppler data, with continuous data streaming in every six minutes, presents a source for a dynamic data driven application system. Scientific inquiries based on these data are useful for dynamic data driven application systems (DDDAS). Sensitivity analysis is performed by changing the threshold values of the rules. Numerical results show that the optimal hybrid model outperforms the direct application of SVM by 12.7 percent. © Springer-Verlag 2004.