Portable NIR and HSI Technologies for Honey Authentication: A Non-Destructive Approach to Detect Food Fraud

Noemí Quintanal¹, Martín Hervello Costas¹, Armando Menéndez Estrada¹, Teresa Ávarez San Lázaro², Pelayo González González¹ ¹ASINCAR, Spain | ²IGP Miel de Asturias, Spain HSI SPECTRAL ANALYSIS Controlled lighting was critical to achieving consistent spectral information in HSI analysis. Two acquisition modes were tested: Diffuse reflectance mode: Proven most effective, with the camera placed directly above the sample and halogen lights positioned at a 45° angle. Transmittance mode: Produced shadow artifacts, especially when using a dome with halogen light ring. The optimized reflectance setup, combined with a white reference background beneath the sample, reduced shadows, minimized spectral artifacts, and significantly improved the discrimination between pure and adulterated honey. CLASSIFICATION MODEL Python was used to process and classify spectral data from NIR (930–1670 nm) and HSI (430–970 nm), excluding edge regions affected by noise. Spectra and labels were organized in structured data frames. Portable NIR and HSI Technologies for Honey Authentication: A Non-Destructive Approach to Detect Food Fraud INTRODUCTION CONCLUSION This study confirms the potential of portable NIR and HSI technologies as effective, rapid, and non-destructive tools for honey authentication. The combined use of spectroscopy and machine learning enabled accurate detection of sugar syrup adulteration and verification of botanical origin, with transmittance and transflectance proving optimal for NIR, and controlled lighting essential for HSI. While the results are promising, real-world implementation faces challenges such as environmental variability, sample presentation, and regulatory acceptance. The methodology, tested with common adulterants, shows potential for broader fraud detection. These findings support NIR/HSI as practical tools for food quality control in both industrial and regulatory settings. EXPERIMENTAL SETUP A total of 148 honey samples were analyzed, comprising 78 pure honeys and 70 laboratory-adulterated samples. Adulteration was performed using three types of syrups, two rice-based and one corn-based, at six concentration levels: 5%, 10%, 20%, 25%, 30%, and 40%.This controlled preparation introduced significant variability, strengthening model robustness and generalization. Samples were measured using portable NIR and HSI devices under varying configurations. To ensure consistency, samples were prepared with uniform thickness and maintained at 37°C, never exceeding 39–40°C, to prevent sugar crystallization. NIR SPECTROSCOPY Three measurement modes were evaluated using multiple portable NIR devices (900–1700 nm): Reflectance: Discarded due to strong water absorption, which significantly distorted the signal. Transmittance: Performed best with thicker honey samples, capturing clear spectral variation. Transflectance: Provided stable, high-quality spectra across a range of conditions. To assess the influence of sample thickness (0.4–10 mm) on spectral quality, multiple configurations were tested. Water absorption had a notable impact, especially in reflectance mode, while transmittance and transflectance modes yielded more stable and informative spectral profiles. Table 1 summarizes the tested configurations, including measurement modes and sample thickness ranges. HYPERESPECTRAL IMAGING Hyperspectral images were captured using a portable Specim IQ camera (400–1000 nm range). Several lighting configurations were evaluated to ensure uniform and shadow-free illumination. Two optimized setups were tested: Diffuse reflection: A white reference surface was used to bounce controlled halogen light at a 45° angle, enhancing even light distribution. Metallic dome: Used to eliminate direct shadowing and standardize illumination across the honey sample. Pixel-wise analysis of the spectral cube enabled the discrimination between pure and adulterated honey samples based on their spectral response. MODEL DEVELOPMENT The development of NIR-based models follows a structured workflow to ensure robustness and accuracy: 1. Experimental design: Selection of ~100 representative and diverse samples based on the target matrix and parameter. 2.Data acquisition: Collection of NIR spectra and reference values to build a reliable training and validation database. 3.Pre-processing: Enhancement of spectral data through techniques such as baseline correction, smoothing, normalization, and derivatives to reduce noise and improve feature consistency. 4.Model development: Predictive models are built using traditional methods like PLS regression, and advanced approaches including SVM and ANN depending on the classification or regression task. 5.Visualization: User-friendly outputs and performance metrics allow easy interpretation of model results, even by non-experts. This workflow enables the creation of robust NIR and HSI tools for food quality control, ensuring effective, efficient, and scalable solutions for industry applications. MATERIALS AND METHODS RESULTS AND DISCUSION Near-infrared (NIR) spectroscopy (700–2500 nm) and hyperspectral imaging (HSI) (400–2500 nm) are nondestructive techniques widely applied in food quality and authenticity assessment. These technologies allow rapid extraction of chemical and physical information and have been used to measure moisture, fat, protein content, and even detect bruises or classify products by origin. Recent advances in portable NIR devices enable real-time and on-site measurements, especially useful in decentralized quality control environments. Honey authentication has become increasingly relevant due to fraud involving syrup adulteration and mislabelling of floral origin. While tradicional adulterants could be more easily detected using conventional laboratory techniques, modern syrups are increasingly similar to honey in composition and spectral profile, making their detection more complex. This study investigates the potential of portable NIR and HSI tools for honey authentication, with a focus on optimized measurement conditions and data processing techniques. NIR SPECTRAL ANALYSIS Transmittance and transflectance modes provided the most reliable spectral data, while reflectance mode showed limited performance due to strong water absorption interference. Optimized configurations were identified for best performance: Transmittance: 10 mm and 2 mm sample thicknesses. Transflectance: 1.5 mm and 0.5 mm thicknesses. These setups ensured optimal light penetration and minimized scattering, enhancing reproducibility across samples. The resulting spectra revealed distinct patterns between pure and adulterated honey, especially in the 1250–1350 nm, 1450 nm, and 1650 nm regions. These correspond to C–H and O–H overtone bands, associated with sugar and water content (Figure X) Preliminary results showed >93% accuracy in detecting adulterated samples. Data were split into training and test sets, followed by 10-fold cross-validation for robust evaluation. To assess model performance, a confusion matrix was generated, confirming strong discrimination between pure and adulterated samples using the PLSDA-SVM model. A user interface was developed in Python (TkInter) to visualize predictions in real time, making the tool accessible for non-expert users and adaptable to different datasets. The steps followed to develop the classification model are summarized in this scheme ACKNOWLEDGMENT This work was performed as part of the WATSON project (A holistic frameWork with Anticounterfeit and inTelligence-based technologieS that will assist food chain stakehOlders in rapidly identifying and preveNting the spread of fraudulent practices), funded by the European Union under the Horizon Europe programme (HORIZON-CL6-2022-FARM2FORK-01, GA: 101084265). Examples of NIR devices used: left – Transmittance mode (NIR-S-T2, LabSpec 4); right – Transflectance mode (NIR-S-G1, Senseen NIR).