Detection and quantification of C3-origin adulterants in carob honey using attenuated total reflectance Fourier-transform infrared spectroscopy and ensemble Machine learning

Glucose syrup and rice syrup, both C3-derived, are invisible to stable isotope ratio analysis, the regulatory gold standard for honey authentication. This creates a critical detection gap. We ask whether attenuated total reflectance Fourier-transform infrared (ATR-FTIR) spectroscopy of the fingerprint region (900–1800 cm−1), coupled with ensemble machine learning (ML), can simultaneously identify adulterant type and quantify concentration in authenticated Moroccan carob honey (Ceratonia siliqua L.). 325 spectra were acquired across six adulteration levels (5–40% w/w) using a 5-biological × 5-technical replicate design, preprocessed with Savitzky–Golay second-derivative transformation, and evaluated by stratified 5-fold cross-validation (CV) and out-of-bag (OOB) estimation. For adulterant-type classification, Random Forest Classifier (RFC) achieved perfect accuracy, F1, AUC-ROC, κ, and MCC (all = 1.000) across every fold, confirmed by OOB accuracy of 100.0%. Gradient Boosting reached 99.69 ± 0.62%. Logistic Regression and Support Vector Machine yielded 0% F1 for the minority pure-honey class despite high overall accuracy, a class-imbalance failure with direct regulatory consequences invisible to accuracy alone. For concentration regression (n = 300), RFC Regressor and Gradient Boosting Regressor achieved R2 = 0.971 and 0.970, MAE = 1.28 and 1.22% w/w, substantially outperforming Linear Regression (R2 = 0.913) and Support Vector Regression (SVR, R2 = 0.852). Gini importance identified distinct spectral discriminants: 1045 cm−1 for type classification and 1182 cm−1 (C–O–C glycosidic linkage, 33.1% of RFR importance) for concentration regression, providing mechanistic interpretability beyond predictive performance. These findings establish a rigorous framework for identifying and quantifying C3-origin adulterants in carob honey. Though multi-batch validation is required before regulatory deployment.