Classification of Australian Thunderstorms using Multivariate Analyses of Large-Scale Atmospheric Variables

Abstract

ightning accompanied by inconsequential rainfall (i.e. ?dry? lightning) is the primary natural ignition source for wildfires globally. This paper presents a machine-learning and statistical-classification analysis of ?dry? and ?wet? thunderstorm days in relation to associated atmospheric conditions. The study is based on daily lightning flash count and precipitation data from ground-based sensors and gauges, and a comprehensive set of atmospheric variables based on the ERA-Interim reanalysis for the period from 2004 to 2013 at six locations in Australia. These locations represent a wide range of climatic zones (temperate, subtropical to tropical). Quadratic surface representations and low-dimensional summary statistics were used to characterize the main features of the atmospheric fields. Four prediction skill scores were considered and ten-fold cross validation used to evaluate the performance of each classifier. The results were compared with those obtained by adopting the approach used in an earlier study for the Pacific Northwest, United States. It was found that: both approaches have prediction skill when tested against independent data, mean atmospheric field quantities proved to be the most influential variables in determining dry lightning activity and no single classifier or set of atmospheric variables proved to be consistently superior to their counterparts for the six sites examined here.