Improving the accuracy of quality assessment for salmon (Salmo salar) by multi-source molecular spectroscopy data fusion and machine learning

Molecular spectroscopy has emerged as a vital tool for modern food quality monitoring. However, relying on a single spectroscopy technique for non-destructive food quality testing presents limitations in accuracy and comprehensiveness. Here, a novel approach integrating near-infrared (NIR) and Raman spectroscopy data fusion to simultaneously assess salmon (Salmo salar) quality traits under varying storage conditions was developed. By focusing on critical indicators such as Warner-Bratzler shear force (WBSF) and α-helix content, the research aimed to enhance predictive modeling through machine learning, including partial least squares (PLS) and least squares support vector machine (LSSVM), combined with two-dimensional correlation spectroscopy and principal component analysis (PCA). The study revealed significant positive correlations (p < 0.05) between α-helix content and key quality traits, including water-holding capacity (r = 0.66) and WBSF (r = 0.67). WBSF and α-helix were used as potential biomarkers for evaluating the quality of salmon during storage. For predictive modeling, the low-level data fusion LSSVM (LLDF-LSSVM) model demonstrated superior performance in estimating WBSF values (R²_P = 0.916, RMSEP = 0.201), outperforming mid-level data fusion (MLDF) and single-spectroscopy models. Additionally, the MLDF-PCA-LSSVM model showed excellent performance in predicting α-helix content (R²_P = 0.984, RMSEP = 0.415), enhancing the accuracy of the PCA feature extraction technique for predicting protein secondary structure. By bridging spectroscopy and advanced machine learning, the findings offer novel insights into real-time quality monitoring, supporting enhanced food safety protocols and reducing waste in the seafood supply chain.