Spectroradiometer-Assisted Quantitative Evaluation of Metanil Yellow Adulteration in Arhar Dal and Flour Using Partial Least Squares Regression

Abstract

Adulteration of food products with non-permitted dyes such as metanil yellow poses a significant threat to consumer health and food authenticity, particularly in arhar dal and flour. This study was aimed at developing a rapid, non-destructive, spectroradiometer-based method for quantifying metanil yellow adulteration in these matrices. Samples containing 0–2% (w/w) metanil yellow were prepared, with 20 independently prepared physical samples for each matrix. Reflectance spectra were acquired over 350–2500 nm. Five replicate spectra from each of 30 measurement positions were averaged, followed by averaging the 30 position-level spectra to obtain one representative spectrum per sample. Representative spectra were divided into calibration and prediction subsets using the Kennard–Stone algorithm (80:20). Spectral preprocessing combinations and latent variables were optimised using five-fold venetian blinds cross-validation on the calibration set, with the prediction set held out. Based on the lowest RMSECV, SNV with four latent variables was selected for arhar flour, whereas smoothing + MSC (median) + OSC with one latent variable was selected for arhar dal. The flour model achieved R²_cv = 0.842, RMSECV = 0.126%, and R²_p = 0.927, while the dal model achieved R²_cv = 0.899, RMSECV = 0.146%, and R²_p = 0.951. Independent hold-out validation yielded RMSEP values of 0.123% for flour and 0.150% for dal, with biases of 0.0002% and 0.005%, respectively. Method-specific LOD and LOQ estimates based on the integrated 500–600 nm reflectance response were 0.08375% and 0.25377% for flour and 0.25906% and 0.78499% for dal, respectively. Low-concentration analysis (0–0.20%) showed increased prediction uncertainty. Overall, the results demonstrate the preliminary feasibility of spectroradiometry combined with PLSR for metanil yellow estimation under controlled conditions. Further validation using larger, multisource datasets and diverse adulteration conditions is required.