Partial Least Squares Regression Combined with Various Variable Selection Techniques to Determine Pork Adulteration in Ground Beef and Lamb Using Visible Near Infrared Hyperspectral Imaging

Meat adulteration raises special concerns regarding quality, safety, and socio-cultural values. This study integrated visible-near-infrared hyperspectral imaging (Vis-NIR HSI) at 397.4 - 1,003.8 nm and partial least squares regression (PLSR) to determine the concentration of pork added in ground beef and lamb. Moreover, this study compared various wavelength-selection methods, i.e., Competitive Adaptive Reweighted Sampling (CARS), Monte Carlo Uninformative Variables Elimination (MCUVE), Successive Projections Algorithm (SPA), and Variable Importance in Projection (VIP). Pork at concentrations of 2%, 5%, 10%, 20%, 30%, 40%, and 50% by weight was added to ground beef and lamb, and the 34 samples were scanned using the Vis-NIR HSI. PLSR models were developed using full-wavelength and selected-wavelength data obtained from CARS, MCUVE, SPA, and VIP. Among feature-selection approaches, the SPA produced the most accurate models with coefficient determination (R²) values of 0.974 for calibration, 0.952 for cross-validation, and 0.955 for prediction for the Beef + Pork model, as well as yielded R² values of 0.922 for calibration, 0.847 for cross-validation, and 0.803 for prediction for the Lamb + Pork model. Chemical maps visualized the distribution of pork adulterants, with clearer spatial patterns in beef than in lamb. The results confirm the use of Vis-NIR HSI for rapid, non-destructive detection of meat adulteration. HIGHLIGHTS Vis-NIR HSI (400 - 1,000 nm) quantified pork adulteration in beef and lamb. Preprocessing affected PLSR accuracy. SPA-selected wavelengths improved PLSR prediction of pork concentration in beef and lamb. At the NIR region, beef samples showed higher separability from pork than from lamb. Pixel-level prediction enabled spatial mapping of pork distribution in beef and lamb. GRAPHICAL ABSTRACT