Predicting eating quality of pork shoulder butt using Vis/NIR hyperspectral imaging and machine learning models

This study proposed a framework for predicting eating quality clusters (EQCs) of pork shoulder butt using visible/near-infrared hyperspectral imaging (HSI). HSI data from 189 samples were acquired, and their EQCs were defined by applying k-means clustering (k = 2 and 3) to an integrated feature matrix composed of results from sensory and electronic tongue analyses. Three datasets, including muscle, fat, and whole-meat regions, were generated using a partial least squares-discriminant analysis-based segmentation procedure. The muscle region outperformed other datasets in both class classification tasks. Optimal feature selection and data augmentation methods were then explored. In the binary classification, the combination of uninformative variable elimination, combined transformation, and a logistic regression model achieved the best performance. In the three-class classification, the combination of a support vector machine with all variables and scale and noise addition showed the best performance. Consequently, HSI would serve as a rapid and non-destructive tool for predicting pork EQCs.