Abstract
The extraction methods used in wet chemical analysis of honey prior to antioxidant constituent analysis may decrease their activities due to decomposition. Therefore, near-infrared (NIR) spectroscopy technology has been used to measure these compounds. Besides, partial least squares regression (PLSR) was also applied to establish the calibration models with 105 samples and 45 unknown honeys to affirm the robustness of the established model. The outcomes of calibration models excellently affirmed the quantification of the concentrations of phenolics, flavonoids, carotenoids, ferric-reducing antioxidant power (FRAP), and 1,1-diphenyl-2-picrylhydrazy (DPPH) with high correlation coefficients (0.95–0.96), and acceptable standard error of cross validation (RMSECV) ranged from 0.92 to 13.60. The ratio performance deviation (RPD) for phenolics, flavonoids, carotenoids, FRAP, and DPPH ranged from 3.45 to 3.89, hence indicating that the NIR equations established were appropriate for unknown samples. Moreover, two discrimination modeling techniques including least squares support vector machine (LS-SVM) and linear discriminant analysis (LDA) were applied for discrimination of honey varieties. Cross validation was applied to optimize the performance of the models. The results revealed that LDA model was superior to LS-SVM model with discrimination rate of 100 % in both the training and prediction sets, whereas LS-SVM model had 99.05 and 93.33 % for the training and prediction sets, respectively. In general, the findings indicated that NIR coupled with chemometrics and linear/non-linear classification algorithms could be applied successfully to determine the antioxidant properties and discriminate honeys of different floral sources.