Ensemble data mining approaches to forecast regional sugarcane crop production

Abstract

Accurate yield forecasts are pivotal for the success of any agricultural industry that plans or sells ahead of the annual harvest. Biophysical models that integrate information about crop growing conditions can give early insight about the likely size of a crop. At a point scale, where highly detailed knowledge about environmental and management conditions are known, the performance of reputable crop modelling approaches like APSIM have been well established. However, regional growing conditions tend not to be homogenous. Heterogeneity is common in many agricultural systems, and particularly in sugarcane systems. To overcome this obstacle, hundreds of model settings (‘models’ for convenience) that represent different environmental and management conditions were created for Ayr, a major sugarcane growing region in north eastern Australia. Statistical data mining methods that used ensembles were used to select and assign weights to the best models. One technique, called a lasso approximation produced the best results. This procedure, produced a predictive correlation (rcv) of 0.71 when predicting end of season sugarcane yields some 4 months prior to the start of the harvest season, and 10 months prior to harvest completion. This continuous forecasting methodology based on statistical ensembles represents a considerable improvement upon previous research where only categorical forecast predictions had been employed.