Rapid Assessment of Chemical Parameters in Italian Honey by Near-Infrared (NIR) Spectroscopy

With the continuously increasing global demand for food, ensuring the quality and authenticity of high-value products, like honey, has become critical to minimize economic losses and guarantee consumer safety. Traditional chemical analyses, while accurate, are often time-consuming, destructive, and require skilled personnel. Near-Infrared (NIR) spectroscopy offers a rapid, non-destructive, and environmentally friendly alternative, allowing the prediction of chemical parameters and the classification of botanical origin for honey samples. This study evaluated the feasibility of NIR-based prediction models for 80 Italian honey samples, provided by CONAPI (BO), collected in 2024 from Campania, Calabria, and Puglia regions, encompassing nine botanical origins, including polyfloral, chestnut, acacia, orange, clementine, honeydew, French honeysuckle, linden, and lemon blossom. Reference chemical analyses were performed according to harmonized methods of the International Honey Commission and Italian legislation, at the Istituto Zooprofilattico Sperimentale delle Venezie (IZSVe), specifically on the Centro di Referenza Nazionale per l’Apicoltura, assessing water content, glucose, fructose, sucrose, electrical conductivity, pH, diastatic activity, and hydroxymethylfurfural (HMF) content. Spectral data were acquired using a benchtop FOSS DS-2500 (850-2500 nm) and two portable instruments, NeoSpectra™ (1350-2500 nm) and AlbaNit (1102-1600 nm), at the LabCNX Laboratory of the University of Padua (Department of Animal Medicine, Production and Health - MAPS). Pre-processing, model calibration, and cross-validation were performed using a customized chemometric workflow, including Partial Least Squares Regression (PLSR), k-Nearest Neighbours (KNN), Random Forest (RF), and Support Vector Machines (SVM) algorithms for both regression and classification tasks. Model’s hyperparameters were optimized using a fine-tuning grid, and performance was evaluated through Venetian-blind 5-fold cross-validation and 100-iteration Bootstrap resampling. Metrics included the coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), balanced accuracy, macro-F1 score, and Matthews correlation coefficient (MCC). Two calibration strategies were compared: full-spectrum and band-selected models using siPLS/CARS-PLS algorithm with interval selection based on VIP and selectivity ratio profiles. The FOSS DS-2500 consistently yielded the highest predictive performance, achieving R² values up to 0.91 for glucose and electrical conductivity, with lower predictive quality for HMF and diastatic activity (R² ≤ 0.49). Classification of botanical origin reached a balanced accuracy of 0.90 and an F1 score of 0.88. Portable instruments showed moderate to low predictive performance, with NeoSpectra™ performing well only for glucose (R² = 0.85) and AlbaNit achieving moderate regression metrics (R² 0.70-0.85) and balanced accuracy of 0.62 for classification. Band selection slightly improved model robustness for lower-resolution portable instruments but had minimal impact on the benchtop instrument, highlighting the trade-off between spectral richness and noise reduction. Overall, NIR spectroscopy combined with chemometric modeling proved effective for rapid prediction of key honey quality parameters and moderate classification of botanical origin, particularly when using high-resolution benchtop instruments. Limitations remain for low-concentration analytes, such as HMF and diastatic activity, due to weak and overlapping spectral features. These findings support the potential integration of portable NIR devices for in-field screening and quality monitoring, complementing traditional laboratory analyses while reducing analytical time and costs.