Multivariate and machine learning approaches for honey botanical origin authentication using near infrared spectroscopy

In this work the feasibility of near infrared spectroscopy was evaluated combined with chemometric approaches, as a tool for the botanical origin prediction of 119 honey samples. Four varieties related to polyfloral, acacia, chestnut, and linden were first characterized by their physical-chemical parameters and then analyzed in triplicate using a near infrared spectrophotometer equipped with an optical path gold reflector. Three different classifiers were built on distinct multivariate and machine learning approaches for honey botanical classification. A partial least squares discriminant analysis was used as a first approach to build a predictive model for honey classification. Spectra pretreatments named autoscale, standard normal variate, detrending, first derivative, and smoothing were applied for the reduction of scattering related to the presence of particle size, like glucose crystals. The values of the descriptive statistics of the partial least squares discriminant analysis model allowed a sufficient floral group prediction for the acacia and polyfloral honeys but not in the cases of chestnut and linden. The second classifier, based on a support vector machine, allowed a better classification of acacia and polyfloral and also achieved the classification of chestnut. The linden samples instead remained unclassified. A further investigation, aimed to improve the botanical discrimination, exploited a feature selection algorithm named Boruta, which assigned a pool of 39 informative averaged near infrared spectral variables on which a canonical discriminant analysis was assessed. The canonical discriminant analysis accounted a better separation of samples according to the botanical origin than the partial least squares discriminant analysis. The approach used has permitted to achieve a complete authentication of the acacia honeys but not a precise segregation of polyfloral ones. The comparison between the variables important in projection and the Boruta pool showed that the informative wavelengths are partially shared especially in the middle and far band of the near infrared spectral range.