Hail in Northeast Italy: A Neural Network Ensemble Forecast Using Sounding-Derived Indices

Abstract

n a previous work, the hailpad data collected over the plain of the Friuli Venezia Giulia region in northeast Italy during the April?September 1992?2009 period were studied through a bivariate analysis with 52 sounding-derived indices from the Udine?Campoformido station (WMO code 16044). The results showed statistically significant relations but, nevertheless, were not completely satisfactory from a practical point of view. In the current work, a prognostic multivariate analysis is performed, using linear and nonlinear approaches, finding the best results with an ensemble of neural networks. For the hail occurrence?classification problem, a novel method for combining binary classifiers (a variant of the Mojirsheibani major voting algorithm) is introduced. For the hail extension?regression problem the ensemble is built by choosing the members with a bagging algorithm, but combining them with a linear multiregression, in order to increase the forecast variability.