PORTABLE NIR AND SUPPORT VECTOR MACHINE APPLIED IN THE DETECTION OF TEFF FLOUR ADULTERATION

Food fraud is still a problem with numerous occurrences. Methods for on-site evaluation, like near-infrared (NIR), are being developed to speed up inspection. Nutritionally interesting foods like Eragrostis tef (Zucc.) Trotter are targets of fraud due to their commercial value. The objective of this research was to evaluate the portable NIR with chemometrics to quantify adulteration in teff flour. Flours from two brown teff crops were adulterated with rice, whole wheat, oat, and rye flours by simplex-lattice mixture design {5,4} (65%-100% of teff and 0-35% of adulterant). The portable InnoSpectra NIR-M-R11 spectrometer (1350-2150 nm) was used to analyze the samples. The 144 mean spectra were smoothed using cubic splines and pre-processed with MSC (multiplicative scatter correction), SNV (standard normal variates), first derivative, second derivative, and combinations. The calibration and prediction sets were divided using Kennard-Stone and SPXY (joint x-y distance) methods. The mean-centering, autoscaling, Pareto, and Poisson normalizations were compared. PLS (partial least square) and SVM (support vector machine) regression models were obtained to quantify the teff content in the mixtures. The SVM parameters were adjusted using the particle swarm optimization (PSO) applying a 10-fold cross-validation. Kennard-Stone sampling was more efficient. The best models were obtained for MSC, Pareto normalization, and SVM regression (r2PLS_prediction = 0.66 and RMSEPPLS = 0.0543, r2SVM_prediction = 0.76 and RMSEPSVM = 0.0491, r2 is the correlation coefficient and RMSEP is the root mean square error of prediction). Furthermore, similar performances were obtained for epsilon-SVR with polynomial kernel and nu-SVR with the radial base kernel (r2 = 0.76 and RMSEP = 0.0491 and r2 = 0.76 and RMSEP = 0.0493), but the radial base has fewer parameters to optimize. The models were able to predict adulteration with an error lower than 5%. Therefore, the proposed methodology can be used for fraud detection in teff flour.