Qualitative classification of sugar processing stream products by near infrared spectroscopy

The Sugar Milling Research Institute NPC (SMRI) is an integral and essential part of the sugar industry as it provides a quality control service among other consultation services to sugar mills in South Africa and other parts of Africa. SMRI uses various prediction equations with near infrared spectroscopy (NIRS), in transmission mode, to predict analyte concentrations present in the various sugar stream products. In this study, chemometrics was used to develop a classification model using discriminant analysis, which could be applied to the process analysis to choose the correct prediction equation for a specific sugar stream product. Samples were selected based on various geographical and environmental factors to ensure variability between the samples. Two different types of data sets were explored to determine the best classification model. The first method used the spectral data of absorbance and wavelength of each sample: Pre-processing was carried out to eliminate any scattering effects. Principal component analysis (PCA) was then applied to reduce the data so that only the necessary information remained. Various classification models, namely, K-nearest neighbour (KNN), Classification tree, Support vector machine (SVM), and Logistic regression, were tested and validated by comparing the predicted sample types against actual sample types. Results showed that the KNN (3) model with the Savitzky Golay filter and three principal components (PCs) provided the best separation between the various sugar stream products. The second method used the analyte concentrations for pol (apparent sucrose content), Brix (total dissolved solids), sucrose, fructose, glucose, and ash for the various sugar stream products. These results were standardised before PCA was applied. The same classification models were applied, tested, and validated using actual samples. These results showed that the Logistic regression model with two PCs performed best. The optimum model from each investigation was compared against each other by evaluating the performance measures of the two models. Based on the analyte concentration data, the Logistic regression (lasso) model with two PCs provided the best separation between sugar stream products. The F1 scores and classification accuracies determined this for the calibration and independent validation sample data set, which were 99.4 and 100 %, respectively.