Seasonal Climate Prediction for the Australian Sugar Industry Using Data Mining Techniques

The ability to predict rainfall with adequate certainty and lead time is beneficial to both industry and public. Periods of high or low seasonal rainfall can have many follow on effects to agriculture, industry, public health and, water supply and management. In order to implement decisions, planning and management strategies to contend with these issues, the ability to predict seasonal rainfall quantities is of great importance (Klopper et al., 2006). Climate conditions are known to influence the cultivation of Sugarcane influencing planting, harvesting and milling (Muchow and Wood, 1996; Everingham et al., 2002; Jones and Everingham, 2005). Unforeseen climate events such as excessive rainfall, can adversely effect the agricultural practices related to Sugarcane cultivation. The Australian Sugarcane harvest period commences in May/June and aims to finish by November/December before the start of the rainy season (Everingham et al., 2002). The risk of excessive rainfall disrupting harvest operations is greatest towards the end of the sugarcane harvest period (Muchow and Wood, 1996; Everingham et al., 2002). Therefore, improved seasonal rainfall prediction during the October-December period is beneficial. Statistical prediction of seasonal rainfall can be performed using a variety of techniques including: regression (Singhrattna et al., 2005), classification methods (Drosdowsky and Chambers, 2001), canonical correlation analysis (Landman and Mason, 1999) and neural networks (Mason, 1998). All statistical models require predictor variables which act as proxies for describing the behaviour of response variables (Hastie et al., 2001). When considering a seasonal forecast model, it is useful to draw predictor variables from a climate data set that is both historically and spatially complete (Washington and Downing, 1999). One of the most temporally and spatially resolute climate parameters is sea surface temperature (SST) data. Consequently, SST data are often used as an empirical measure of the ocean-atmosphere interaction in statistical climate models. However, a vast proportion of potential SST predictors may be redundant. Therefore employing data mining methods for the purpose of feature extraction and data reduction is advantageous. Principal component analysis (PCA) is a commonly used feature extraction method that reduces data dimensionality whilst retaining the majority of the variability (Jolliffe, 1986). As sea surface temperature data sets are large, it is useful to perform PCA data reduction such that the bulk of the variability is contained in a small subset of variables (Wilks, 1995). PCA also referred to as empirical orthogonal function (EOF) analysis is commonly used throughout climate research (Wilks, 1995). PCA is popular because it is available in most