Accurate precipitation prediction with support vector classifiers: A study including novel predictive variables and observational data

This paper presents a study on the Support Vector Machine (SVM) performance in a problem of daily precipitation prediction. Several novelties are included in the proposed analysis: first, a large set of novel predictive variables is considered, including upper air sounding data, variables derived from a numerical weather prediction model as well as the synoptic pattern of the atmosphere (by means of the Hess–Brezowsky classification). The importance of several of these predictive variables in the SVM performance is analyzed in the paper. In addition, two types of observational rain data are used in the experiments: first data from rain gauges (pluviometers) are considered, in order to establish the precipitation prediction, and then observational data from airports (METAR and SPECI reports) are used to carry out a similar study. The excellent performance of the SVM approach is shown by comparison with several alternative neural computation-based approaches (multi-layer perceptron, Extreme Learning Machine) and with classical algorithms such as decision trees and K-nearest neighbor classifier. Finally, the results of the persistence model are used as reference to certify the good performance of the proposed technique.