Key points are not available for this paper at this time.
Abstract Adulteration in edible oils threatens food safety and regulatory compliance, creating a need for rapid, non-destructive screening. This study presents an integrated hyperspectral imaging and machine learning framework that combines Savitzky–Golay smoothing with multiple band selection techniques (Correlation, Entropy, Manifold Learning, Mutual Information (MI), Signal-to-Noise Ratio (SNR), and Variance-Based methods) were employed to extract the most informative features/wavelengths. Using 240 samples acquired with a Specim FX10 camera, covering pure rose, rosehip, tea tree, and turmeric oils and oils adulterated with lemon, vitamin E, garlic, and orange, the selected wavelengths were used to train decision tree models (C4.5, C5.0, ID3) and gradient boosting models (HistGradient Boosting, XGBoost). HistGradient Boosting achieved the best results, with 100% training accuracy and 93% validation accuracy, and complementary metrics (Cohen kappa, MCC, Jaccard index, precision, recall, F1 score, and Hamming loss) supported robust generalization. The novelty lies in jointly exploiting information-theoretic relevance and redundancy control to yield a compact spectral subset, coupled with rigorous benchmarking of boosting methods across multiple oil/adulterant pairs, enabling fast, accurate, and practical quality control.
Aqeel et al. (Mon,) studied this question.