Predicting sensory characteristics of pork using physicochemical and VIS/NIR hyperspectral imaging data with machine learning

Abstract

Accurate assessment of pork taste quality remains challenging due to the subjectivity and destructiveness of traditional sensory evaluation methods. This study aimed to develop predictive models for pork sensory attributes using both physicochemical data and hyperspectral imaging (HSI) in combination with machine learning algorithms. Fifty pork shoulder butt samples were analyzed for physicochemical parameters (pH, water holding capacity, meat color, cooking loss, fat content, and glutamic acid) and sensory attributes (fatness, sweetness, saltiness, umami, juiciness, and flavor). HSI data were acquired in the 400–1,000 nm and 900–1,700 nm ranges. Classification models were developed to predict sensory traits and were evaluated using accuracy and F1-score. Models using physicochemical data achieved moderate performance (F1 = 0.60 for saltiness), which improved with quality control-based filtering (F1 = 0.69). HSI models achieved the highest prediction for fatness (F1 = 0.73), and VIP-based wavelength selection improved umami prediction (F1 = 0.65). These findings suggest that combined use of HSI and physicochemical features can support non-destructive, data-driven prediction of pork taste quality. Further integration with metabolomics is recommended to improve prediction of complex traits like umami.