Integrating EEM fluorescence spectroscopy with machine and deep learning models for multi-output prediction of fish freshness

• a multi-task model achieved fish freshness assessment. • The PCA-RF model performed best in predicting freshness than others. • CNN-assisted EEM to achieve precise classification of fish freshness classes. • The key wavelengths of both fishes by gradient integration algorithm were similar. • Fine-tuning of models between the two kinds of fishes proved to be applicable. A precise and real-time assessment of the freshness of aquatic products during storage can effectively reduce economic losses and food safety problems. This study developed a novel multi-output freshness monitoring model integrating excitation-emission matrix (EEM) fluorescence spectroscopy with two machine learning (ML) and one deep learning (DL) approach to simultaneously predict freshness indices and classify freshness stages. We Utilized bighead carp ( Hypophthalmichthys nobilis ) andrainbow trout ( Oncorhynchus mykiss ), to predict TVB-N, K value, and TVC freshness indices via PCA-RF regression, achieving R² = 0.85-0.98 and RMSE = 0.23-3.63 significantly outperforming PARAFAC-SVR/RF and PCA-SVR; classify freshness into fresh, sub_fresh, and rot stages using a custom convolutional neural network (CNN), attaining 93.84% for bighead carp and 90.47% for rainbow trout; validate model generalizability through transfer learning, demonstrating that fine-tuned species-specific CNNs achieve cross-species classification accuracy exceeding 88.40%. This approach circumvents manual feature engineering via end-to-end EEM-CNN processing. Furthermore, we established the feasibility of transferring models trained on one species to another through fine-tuning. This study provides a model for precise quality monitoring in the fish supply chain and further applications of real-time freshness monitoring and risk assessment.