Predicting the Onset of Australian Winter Rainfall by Nonlinear Classification

Abstract

Abstract A method for predicting the timing of winter rains is presented, making no assumptions about the functional form of any relationships that may exist. Ideas built on classification and regression trees and machine learning are used to develop robust predictive rules. These methods are applied in a case study to predict the timing of winter rain in five farming towns in the southwest of Western Australia. The variables used to construct the model are mean monthly sea surface temperatures (SSTs) over a 72-cell grid in the Indian Ocean, Perth monthly mean sea level pressure (MSLP), and monthly values of the Southern Oscillation index (SOI). A predictive model is constructed from data over the period 1949–99. This model correctly classifies the onset of the winter rains approximately 80% of the time with SST variables proving to be the most important in deriving the predictions. Further analysis indicates a change point in the mid-1970s, a well-known phenomenon in the region. The prediction rates are significantly worse after 1975. Furthermore, the important region of the Indian Ocean, in terms of SSTs for prediction, moves from the Tropics down toward the Southern Ocean after this date.