Investigating the identification of atypical sugarcane using NIR analysis of online mill data

Abstract In any given season thousands of tonnes of sugarcane with atypically low quality can pass undocumented through Australian sugarcane mills. Sugarcane with atypically low quality can negatively impact mill processes and throw off grower payment calculations. Mill laboratory operators often observe a small subset (1–5%) of cane consignments that have an unusually low juice Pol (Pij; a measure of sucrose content) relative to juice brix (Bij; a measure of dissolved sugars), that can indicate deteriorated or contaminated cane. Many mills only test a small subset of cane in the laboratory, with the majority of consignments analysed using fast near infrared (NIR) spectroscopic techniques. This means the true extent of ‘atypical’ consignments cannot be identified. To address this limitation, this paper compares five modelling approaches: Linear discriminant analysis (LDA), partial least squares discriminant analysis (PLS-DA), random forest (RF), Artificial Neural Networks (ANN) and Support Vector Machines (SVM). Model performance was reported as the correct classification rate (CCR) of typical and atypical samples based on independent test sets. The best performance was achieved by PLS-DA (CCRatypical = 88.65% and CCRtypical = 88.75%), while ANN had the lowest performance (CCRatypical = 85.27% and CCRtypical = 83.66%). The methodology used in this paper could be used to identify atypical consignments allowing mills to track occurrences to farms and if necessary develop process control operations for atypical cane. Furthermore, the use of a relatively simple modelling technique such as PLS-DA means model updates can be made efficiently and with confidence as PLS is already well established within the industry.