Quantification of cumin powder adulteration with rice by-products using NIR spectroscopy combined with chemometric and machine learning models

Fraud in spices such as cumin (Cuminum cyminum L.) represents a major challenge for the food industry due to its high commercial value and the difficulty of detecting adulteration in powdered form. Rice by-products, including rice bran (RB) and small broken rice (SBR), have not previously been evaluated as cumin adulterants, despite their low cost and capacity to visually blend into the matrix. This study developed and compared near-infrared spectroscopy (NIRS) models coupled with partial least squares regression (PLSR), multilayer perceptron (MLP), and long short-term memory (LSTM) networks to quantify cumin adulteration with RB and SBR. Binary mixtures (5 %–40 %, w/w) were prepared, and principal component analysis (PCA) of NIR spectra (1100–2100 nm) revealed clear separation between pure cumin, adulterants, and mixtures. Regression models built with both full spectra and β-coefficient–selected wavelengths showed excellent performance (R² > 0.97; RMSE < 5.5 %). Among them, the MLP achieved the highest accuracy for both full (R² = 0.999–1, RMSE =0.74²%–0.963 %) and optimized (R² = 0.999, RMSE = 0.973 %–1.225 %) spectra, followed by PLSR and LSTM. These findings highlight NIRS combined with chemometrics and machine learning as a rapid, non-destructive approach for spice authentication, suitable for portable in situ monitoring.