Objective: The moisture, protein, and fat content in meat are among the most critical characteristics determining its quality. Traditional methods for measuring these attributes are typically time-consuming, destructive, and costly. Substituting these with spectroscopic methods could be a significant step toward developing rapid techniques in this field. Combining linear or nonlinear machine learning methods with near-infrared spectroscopy enables the quantitative determination of these components in meat samples. This study aims to compare the performance of linear and nonlinear machine learning methods in determining these three attributes in meat samples.
Materials and Methods: In this study, near-infrared (NIR) spectral data from 240 ground meat samples were collected using an NIR spectrometer in the wavelength range of 850 to 1050 nm. Of these, 170 samples were used for calibration and 70 for testing. Two machine learning methods, partial least squares (PLS) as a linear method and radial basis function artificial neural network (RBF-ANN) as a nonlinear method, were employed to model and predict the concentrations of moisture, protein, and fat. Modeling was performed using MATLAB 2018b with the MVC1 package.
Findings: Spectral data from 240 meat samples were analyzed using partial least squares (PLS) and radial basis function artificial neural network (RBF-ANN) methods. A total of 170 samples were used for calibration, and 70 samples were used for testing. The RMSEP values for predicting moisture, protein, and fat using RBF-ANN were 0.8500, 0.6538, and 0.6761, respectively, and for PLS, they were 2.3132, 2.8023, and 0.7615, respectively. The REP% values for RBF-ANN were 1.3494%, 3.5680%, and 3.8254%
for moisture, protein, and fat, respectively, while for PLS, they were 3.6725%, 15.2922%, and 4.3083%, respectively.
Conclusion: Partial least squares (PLS) and radial basis function artificial neural network (RBF-ANN) methods were applied to analyze spectral data from 240 meat samples. The variables of each machine learning method were optimized during the calibration phase using near-infrared spectra from 140 meat samples, and the corresponding linear and nonlinear models were used to predict the concentrations of three analytes (moisture, protein, and fat) in 70 independent meat samples. Although the nonlinear RBF-ANN model, as an artificial intelligence approach, demonstrated higher efficiency in predicting concentrations, the linear PLS model also provided acceptable performance in determining fat content in meat samples.