A rapid method to predict type and adulteration of coconut milk by near-infrared spectroscopy combined with machine learning and chemometric tools

Abstract

Coconut milk is a soft target for adulterators owing to its simplicity of chemical composition. Professionals and consumers want to control the originalitas of coconut milk, while sellers can profit by mixing fresh coconut milk from low-cost products into high-value fresh coconut milk. Non-destructively and rapidly identifying coconut milk classification goods may be useful in quality assurance settings. However, no studies to date have investigated this topic. In this study, near-infrared spectra (NIRs) were collected from fresh coconut milk (FCM), instant coconut milk (ICM), and adulterated fresh coconut milk (A-FCM) in order to investigate the prospect of non-invasively discriminating coconut milk type and at the same time predicting the level of A-FCM. Partial least squares (PLS), linear discriminant analysis (LDA), support vector machine (SVM), and multilayer perceptron (MLP) were employed to establish classification and regression models using NIRs. Combining 18 preprocessing types and hyperparameter optimization of individual machine learning algorithms is carried out together and evaluated using 5-folds cross-validation. All algorithms in this study (LDA, SVM, MLP) obtained the same satisfactory results with all the precision, recall, F1-score, and perfect accuracy (100%) to distinguish FCM, ICM, and A-FCM in both calibration and prediction. Regression models using the SVM obtained acceptable results, with a determination coefficient of calibration and prediction all over 0.93, root mean square error of calibration and prediction all below 8.30%, and ratio of prediction to deviation over 3.80. Last but not least, this study would help apply NIRs to detect the originality of coconut milk in real-world conditions.