AI-Integrated Portable VIS-NIR Spectroscopy for On-Site Authenticity Evaluation and Quality Grading of Mutton-Veal Minced Meat.

Abstract

Minced meat for home cooking is typically made up of 40% mutton and 60% veal, providing a balanced flavor and texture. However, some producers reduce the mutton content to increase profits, leading to consumer dissatisfaction. This study aimed to develop a handheld Visible-Near Infrared (VIS-NIR) spectroscopy system coupled with artificial intelligence (AI) algorithms as a rapid, accurate, and non-destructive method for predicting and grading the mutton-veal composition of minced meat samples. The weight ratios of the veal in the compositions ranged from 0 to 100% at 2% intervals. Several chemometric and AI methods were used to analyze the spectral data of the prepared minced meat samples captured by a handheld Vis-NIR spectrometer under LED and bulb light sources. The spectral data were divided for modeling using the stratified and Kennard-Stone algorithms. Results indicated that the best model was Multi-Layer Perceptron (MLP) as an artificial neural network algorithm, which achieved R2 CV and R2 P (0.879 and 0.916, respectively) for the best scenario of the samples' authenticity prediction. In quality grading, the MLP achieved the highest performance in classifying samples into three quality classes-excellent (Grade A) for 0%-28% veal, medium (Grade B) for 30%-58% veal, and poor (Grade C) for 60%-100% veal-with an accuracy of 92.4%, and precision, recall, and F1 score all reaching 96.2%. These findings confirm that the proposed approach provides a fast, reliable tool for quantifying veal content in minced meat, supporting the development of effective meat quality control systems.