Assessment of Italian honey origin using near infrared spectroscopy and multivariate modeling

Honey is a natural sweetener synthesized by bees thanks to the blooming of flowers and plants, whose nutritional properties depend solely on the same floral origins. It can provide various benefits for human health, which places honeys market value in a better position than other sweeteners. As a result, honey has always been adulterated, making its authenticity a concern for researchers and producers worldwide. The most ancient and popular method for determining the floral and geographical origin of honey is melissopalynology, performed by analyzing the pollen contained in honey. However, this method has several limitations in terms of both the lengthy duration of the analyses and high costs, as well as the need for specialists for its execution. Finally, it is unable to detect fraudulent pollen contamination. In recent decades, Near InfraRed (NIR) Spectroscopy (NIRs) has been widely applied in agriculture and the food industry, revealing to be a valid predictive and explorative analytical technique to determine the quality of products.NIR offers numerous performance advantages being a fast, non-destructive, reliable, cost-effective method and does not require preliminary preparation of samples. The purpose of this thesis is to evaluate the feasibility of using Near-Infrared (NIR) spectroscopy as a rapid analytical tool to determine the geographical origin of a pool of polyflower honey samples produced in the Italian context. The samples (n = 227) of honey were classified according to their regions and macro-areas of production: SL = South Lowland, below 600 m above sea level (asl); NM = North Mountain, above 600 m asl; NL = North Lowland, below 600 m asl. The samples were analyzed using three portable instruments: a visible spectrophotometer (VIS), a VIS-NIR spectrophotometer and a NIR spectrophotometer. The spectral data collected by the three analytical instruments VIS, VIS-NIR and NIR were pre-processed and used to define four classification models, such as Random Forest, K-Nearest Neighbor (KNN), Support Vector Machine (SVM) linear and SVM radial, applying machine learning (ML) algorithms. The data were randomly divided into a training test (70% of the data, n = 162), necessary for the training of the four classification models, and a test set (30% of the data, n = 65), where their predictive performance was evaluated using confusion matrices. The results show that the multi-instrumental approach, which combines NIR spectroscopy and colorimetric analysis, has a poor accuracy (0.49) in distinguishing the origin of polyflower honey from specific regions, but is more effective in classifying honey in wider geographical areas, with an accuracy of 0.58. The low performance of the predictive models is probably due to the high intrinsic variability of the polyflower honey product, which is influenced by multiple factors such as the chemical-physical composition, mineral content, and colorimetric characteristics. These conditions can also vary within the same region due to the botanical and flowering diversity of the plant species affected by bees, varying climatic conditions and beekeeping practices. Moreover, since geographical areas are divided by altitude, they can include more or less vast territories, thus contributing to greater intrinsic variability. The use of portable instruments, with a limited wavelength range, posed greater challenges, making it more difficult to meaningfully distinguish the origin of honey. The results suggest that, although NIR spectroscopy and multivariate modeling approaches are widely used in the food industry, further technological improvements are needed in order to develop an integrated analytical system capable of accurately authenticating and discriminating the geographical origin of Italian polyflower honey.