Enhancement of non-destructive chicken freshness prediction using Vis/NIR spectroscopy through wavelength selection and data augmentation

Abstract

This study aimed to develop a non-destructive prediction model for chicken meat quality using Vis/NIR spectroscopy of drip. Chicken meat was vacuum-packaged and refrigerated at 4 °C for 13 days. Drip samples were analyzed for metabolites and Vis/NIR spectra. Microbial composition changes during storage were evaluated using 16S rRNA. Dominant bacteria, including Carnobacterium, Lactococcus, and Serratia, were identified, which contributed to metabolite changes such as the increase in cadaverine, putrescine, and tyramine and the decrease in inosine monophosphate, tyrosine, and uridine monophosphate (UMP). Prediction models were developed using partial least squares regression (PLSR) to enhance performance through wavelength selection by VIP-PLSR and spectral augmentation, including offset, multivariate normal sampling (MVN), and extended multivariate signal augmentation (EMSA). Optimal models for six metabolites were identified, each demonstrating efficacy through either wavelength selection or augmentation methods. The VIP-PLSR model showed the highest predictive performance for cadaverine (R² = 0.79), putrescine (R² = 0.77), and tyramine (R² = 0.73), while spectral augmentation using the MVN method yield the highest performance for UMP (R² = 0.82). The combination of wavelength selection and spectral augmentation provided superior results for acetate and tyrosine, with EMSA (R² = 0.64) and MVN methods (R² = 0.88), respectively. Therefore, wavelength selection and spectral augmentation could enhance predictive performance of the models.