Robust method for detecting metanil yellow in turmeric: Integrating Vis-NIR spectroscopy and machine learning

Abstract

Food color significantly influences the quality and marketability of food, but harmful adulterants like metanil yellow (MY) are sometimes added to turmeric to improve its appearance, posing severe health risks. In this study, two datasets were created by mixing turmeric with different concentrations of MY (0, 0.1, 0.5, and 1-50 percent). The reflectance spectra were recorded in the 400-2050 nm range and pre-processed using Standard Normal Variate (SNV), Multiplicative Scatter Correction (MSC), and Savitzky-Golay (SG) filtering. Principal Component Analysis (PCA) was then applied to reduce dimensionality. For dataset-1, the K-Nearest Neighbors Regressor (KNR) combined with SNV-PCA achieved an R sup(2) of 0.998, an RMSE of 0.664 mg (w/w), and leave-one-out cross-validation (LOO-CV) results of R sup(2) = 0.997 and RMSE = 0.849 mg (w/w). For dataset-2, KNR with SNV-PCA showed R sup(2) = 0.998, RMSE = 0.833 mg (w/w), and LOO-CV R sup(2) = 0.999 with RMSE = 0.645 mg (w/w). When the datasets were combined, KNR with MSC-PCA recorded R sup(2) = 0.992, RMSE = 1.479 mg (w/w), LOO-CV R sup(2) = 0.998, and RMSE = 0.741 mg (w/w). The classification model achieved 97.0 percent accuracy, an F1 score of 97.0 percent, and a 10-fold cross-validation F1 score of 80.8 percent.