Integrating transfer learning and spectroscopy for enhanced pork spoilage assessment using correlation analysis

Accurate Total Viable Count (TVC) detection is vital for food quality monitoring. In this study, we investigated the feasibility of using visible near-infrared (VNIR) spectroscopy (400–1000 nm) combined with transfer learning (TL) to track the chemical spoilage of pork. The base models developed using the full band for pork TVC, total volatile basic nitrogen, pH, and color showed predictability; the correlation coefficient of prediction set (RP) for all models ranged from 0.821 to 0.916; and the root mean square error of prediction set (RMSEP) of the TVC model was 0.617 (lg CFU/g). A correlation analysis of the different indexes of pork was carried out to optimize the TVC calibration model. Different TL methods for TVC optimization were designed. The results showed that multiple correlation chain stacking-partial least squares performed best with RP, RMSEP, and the relative percent deviation of 0.947, 0.425 lg CFU/g, and 2.355, respectively, the RMSEP of TVC was reduced by 31.12 % as compared to the base model. This study demonstrated the possibility of combining the VNIR spectroscopy system with TL to monitor the degree of meat's chemical spoilage. •Transfer learning combined with visible near-infrared spectroscopy to enhance food safety monitoring.•Multi-index correlations of pork spoilage processes analyzed.•Multiple correlation chain stacking-partial least squares algorithm optimized model to significantly improve prediction performance.•Total colony count detection model with high prediction accuracy.