Quantitative determination of TVB-N content for different types of refrigerated grass carp fillets using near-infrared spectroscopy combined with machine learning

Abstract

This research aims to accurately forecast the freshness indicators (TVB-N) of skinned and skinless grass carp fillets by integrating near-infrared spectroscopy (NIR) with machine learning algorithms. By comparing the predictive accuracy of machine learning models for the two types of grass carp fillets, the most effective modeling method is identified. Methodologically, the study first applies orthogonal signal correction (OSC) and the first derivative among other algorithms for spectral pre-processing. Subsequently, competitive adaptive reweighted sampling (CARS), moving window partial least squares (MWPLS), and random frog (RF) are used for the selection of variables. Lastly, partial least squares regression (PLSR), support vector regression (SVR), backpropagation neural networks (BPNN), and particle swarm optimization-enhanced BP neural networks (PSO-BP) are employed to quantitatively analyze the NIR data. The most relevant results reveal that the (OSC+D1)-CARS-PSO-BP model exhibits superior predictive capabilities. Especially when applied to skin-on fish fillets (R2P =0.988, RMSEP=0.092), this model surpasses that for skinless fish fillet data (R2P =0.987, RMSEP=0.096). Therefore, combining near-infrared with machine learning to predict the freshness (TVB-N) of grass carp fillets based on skin-on samples is a more effective non-destructive testing method.