Deep learning-driven hyperspectral imaging for real-time monitoring and growth modeling of psychrophilic spoilage bacteria in chilled beef

Abstract

Owing to the unsound cold chain system in China, chilled beef's quality would be affected by psychrophilic bacteria, resulting in quality deterioration and corruption, which leads to food safety problems. In this study, the growth of Pseudomonas and Lactobacillus in chilled beef was modeled by plate counting method and hyperspectral imaging, while the colony number of each dominant psychrophilic bacteria in chilled beef was determined using a traditional microbiological method. For the spectral data, the competitive adaptive reweighted sampling (CARS) algorithm, variable combination penalty analysis algorithm, successive projection algorithm and iteratively retained information variable were utilized to extract the characteristic wavelengths, and the partial least squares regression (PLSR), Energy Valley Algorithm Optimised Time Convolution Network combined with Multihead Attention Mechanism and stochastic configuration neural network (SCN) were used to predict the content of Pseudomonas and Lactobacillus in chilled beef. For Lactobacillus, the results showed that the prediction based on the Gaussian filtering-PLSR model achieved the optimal modeling (R_c² = 0.7381, R_p² = 0.7101, RMSEC = 0.5802 log₁₀CFU/g, RMSEP = 0.7934 log₁₀CFU/g). For Pseudomonas, the best prediction results were achieved (R_c² = 0.9415, R_p² = 0.8636, RMSEC = 0.7050 log₁₀CFU/g, RMSEP = 1.0546 log₁₀CFU/g) based on the CARS-SCN model. Finally, the growth of Pseudomonas and Lactobacillus was fitted using the Baranyi model, Huang model, and Gompertz model. Rapid nondestructive detection of bacterial content was realized from the hyperspectral data of chilled beef.