Tornado Detection with Kernel-Based Classifiers from WSR-88D Radar Data

Abstract

Detection of tornadoes that provides warning times sufficient for evasive action prior to a tornado strike has been a well-established objective of weather forecasters. With modern technology, progress has been made on increasing the average lead time of such warnings, which translates into a number of lives saved. Recently, machine learning (e.g., kernel methods) has been added to the collection of techniques brought to bear on severe weather prediction. In this chapter, we seek to extend this innovation by introducing and applying two types of kernel-based methods, support vector machines and minimax probability machines to detect tornadoes, using attributes from radar derived velocity data. These two approaches utilize kernel methods to address nonlinearity of the data in the input space. The approaches are based on maximizing the margin between two different classes: tornado and no tornado. The use of the Weather Surveillance Radar 1988 Doppler, with continuous data streaming every 6 min, presents a source for a dynamic data driven application system. The results are compared to those produced by neural networks (NN). Findings indicate that these kernel approaches are significantly more accurate than NN for the tornado detection problem.