Application of machine learning to an early warning system for very short-term heavy rainfall

Abstract

The purpose of an early warning system (EWS) is to issue warning signals prior to extreme events. Extreme weather events, however, are hard to predict due to their chaotic behavior. This paper suggests a method for an effective EWS for very short-term heavy rainfall with machine learning techniques. The EWS produces a warning signal when it is expected to reach the criterion for a heavy rain advisory within the next 3 h. We devised a selective discretization method that converts a subset of continuous input variables to nominal ones. Meteorological data obtained from automatic weather stations are preprocessed by the selective discretization and principal component analysis. As a classifier, logistic regression is used to predict whether or not a warning is required. A comparative evaluation was performed on the EWS models generated by various classifiers. The tests were run for 652 locations in South Korea from 2007 to 2012. The empirical results showed that the preprocessing methods improved the prediction quality and logistic regression works well on heavy rainfall nowcasting in terms of F-measure and equitable threat score.