Fusion of near-infrared and Raman spectroscopy with machine learning strategies: Non-destructive rapid assessment of freshness and TVB-N value prediction in Pacific white shrimp (Litopenaeus vannamei)

Total volatile base nitrogen (TVB-N) is a key indicator of shrimp freshness. Nevertheless, traditional detection methods are cumbersome, time-intensive, and destructive. Here, a rapid and non-destructive method based on near-infrared (NIR) and Raman spectroscopy for the assessment of TVB-N content in Litopenaeus vannamei was proposed. A TVB-N content prediction model was constructed based on three machine learning methods (Convolutional Neural Network, Extreme Learning Machine, and Backpropagation) combined with low-level and mid-level data fusion strategies. After Savitzky-Golay (SG) smoothing preprocessing, the NIR model with SPA feature extraction (coefficient of determination for prediction, R²p = 0.864) outperformed the Raman model with GA feature extraction (R²p = 0.784), with both being the optimal feature-level prediction models for their respective spectra. Furthermore, the combination of mid-level data fusion strategy and the Extreme Learning Machine model resulted in the best prediction performance, with R²p and root mean square error of prediction (RMSEP) values of 0.986 and 0.677 mg/100 g, respectively. Additionally, the feature-level fusion models optimized by the feature selection algorithm showed R² values all exceeding 0.96. These results demonstrate the complementary advantages of NIR and Raman spectroscopy for non-destructive, real-time freshness monitoring of shrimp using portable instruments.