Deep learning algorithm-assisted non-destructive detection of TBARS values of salmon flesh using multi-modal molecular spectra fusion

Abstract

This study presents a deep learning framework for the non-destructive assessment of lipid oxidation in salmon flesh, quantified by thiobarbituric acid reactive substances (TBARS), under diverse storage conditions (−20, 0, 4, 20 °C, and dynamic temperatures). By integrating multi-modal molecular spectroscopy (near-infrared and Raman) with a convolutional neural network (CNN), the research achieved enhanced accuracy in predicting TBARS values compared to partial least squares regression (PLSR) and long short-term memory (LSTM) models. The CNN achieved superior performance with R²ₚ = 0.866 and RPD = 4.136. Its end-to-end architecture and spectral data fusion improved prediction robustness and enabled real-time, non-invasive monitoring of salmon quality. This approach addresses industry demands for efficient cold-chain quality assessment, offering a scalable solution to optimize freshness evaluation and consumer satisfaction. Overall, the methodology demonstrates the strong potential of combining multi-modal spectroscopy with deep learning to advance food quality analytics in dynamic storage environments.