A new comprehensive index for the assessment of mutton quality deterioration during storage using hyperspectral imaging combined with autoencoder-assisted self-supervised learning and generative network

Mutton is tender and juicy but is prone to quality deterioration during storage. Hyperspectral images (HSI) combined with deep learning has significant advantages in predicting mutton quality deterioration during storage, however, the current models are generally faced with the problems of limited sample size and a individual indicator for evaluating mutton deterioration. In this research, physicochemical indicators of mutton at different storage times were determined, including inosine monophosphate, inosine, hypoxanthine , thiobarbituric acid reactive substances, total volatile basic nitrogen, pH, and color. A comprehensive quality degradation index (CQDI) was constructed using the entropy weight method to integrate these measures. Subsequently, a convolutional neural network (CNN) model was developed to predict the CQDI from HSI data. To optimize the CNN performance under data constraints, we employed autoencoder-assisted self-supervised learning (AE-assisted SSL) and autoencoder-assisted generative networks (AE-assisted GN). The results demonstrated that spectral data predicted the CQDI with higher accuracy compared to individual physicochemical indicators. While AE-assisted SSL degrades CNN model performance, AE-assisted GN notably enhances CNN model performance under small sample scenarios. AE-assisted GN enhances CNN model performance, yielding a 3.73% increase in R 2 p and a 94.56% reduction in RMSE. The research indicated that a comprehensive index approach represented a novel method for evaluating mutton quality degradation during storage, and AE-assisted GN could enhance the performance of prediction models in small sample datasets. • A novel spectroscopic method for meat deterioration assessment. • Improved predictive performance under small-sample conditions. • Interpretation of critical spectral bands influencing model output.