Key points are not available for this paper at this time.
Global food loss poses critical sustainability challenges, necessitating advanced techniques for food quality assessment and food shelf-life prediction. This review systematically examines deep learning paradigms for food quality monitoring and shelf-life prediction, with emphasis on the relationships among data modalities, model architectures, and application scenarios. Compared with conventional microbial, physicochemical, and sensory methods, which are often destructive, time-consuming and difficult to apply in continuous monitoring, deep learning enables non-destructive and automated feature extraction from images, spectra and multi-sensor data. This review compares recent applications of convolutional neural network, recurrent neural network, long short-term memory network, gated recurrent unit, Transformer, hybrid architectures and transfer learning across fruits and vegetables, meat, aquatic products, eggs, and dairy products. Key findings indicate that multi-modal data fusion, attention mechanisms, data augmentation, and appropriate data partitioning can improve prediction accuracy, robustness, and practical applicability, while explainable model design can enhance the credibility and transparency of model outputs. Persistent challenges remain in data labeling, model generalization, external validation, and deployment under real supply-chain conditions. Future progress in lightweight models, edge computing, soft sensors, digital twins, and AI-agent-based systems is expected to support reliable food monitoring, reduce food waste, improve supply-chain efficiency, and strengthen food safety.
Fu et al. (Mon,) studied this question.