A non-destructive detection method of protein and TVB-N content changes in refrigerated and frozen-thawed salmon fillets using fluorescence hyperspectral technology

This study proposed a practical predictive method for non-destructive detection of protein and total volatile basic nitrogen (TVB-N) content in refrigerated and frozen-thawed salmon fillets using fluorescence hyperspectral imaging technology. The spectral and chemical data of salmon fillets were obtained via FHSI and an automatic Kjeldahl nitrogen analyzer, respectively. The research demonstrated that the Long Short-Term Memory (LSTM) deep learning model developed in this study offered superior spectral data analysis compared to traditional machine learning models. Additionally, the Northern Goshawk Optimization algorithm was developed to input optimal hyperparameters into the LSTM model (NGO-LSTM), significantly enhancing prediction accuracy. The NGO-LSTM model surpassed the PLSR model with various preprocessing methods and LSTM models optimized by other algorithms, achieving R2 values of 0.910 for protein and 0.987 for TVB-N. This research established a theoretical foundation for future studies on meat freshness detection. [Display omitted] •A novel method was proposed, using FHSI to detect changes in the quality attributes of salmon fillets.•Characteristic wavelengths revealing changes were identified within the range of 490 nm to 680 nm.•An LSTM deep learning model was constructed to analyze fluorescence hyperspectral data.•Northern Goshawk algorithm was proposed to optimize the model, improving its prediction accuracy.