A Clustering-Based Partial Least Squares Method for Improving the Freshness Prediction Model of Crucian Carps Fillets by Hyperspectral Image Technology

Abstract

Hyperspectral imaging technology (HSI) is able to visualize the distribution map of chemicals in samples in combination with a developed prediction model. Generally, prediction models are established based on the spectrum and the chemical reference averagely calculated/measured from the sole area covering the sample. However, uneven chemical distribution is widely observed in the individual sample. The uneven distribution of chemicals may result in the unspecific match between the spectrum and chemical reference, which were averagely achieved from the non-homogeneous sample, leading to low robustness model. The aim of this work was to improve the performance of the freshness prediction models of fillets by eliminating the effect of uneven chemical distribution in each fillet. This study proposed a clustering-based partial least squares (C-PLS) algorithm, which firstly divided a non-homogeneous fillet into several relatively homogeneous sub-pieces using cluster analysis. Spectra and freshness indices were averagely acquired from the sub-pieces respectively, aiming to find a more specific match between the spectra and chemical indices. Compared with the partial least squares regression model, C-PLS model performed a higher coefficient of determination of cross-validation for the prediction of total volatile basic nitrogen (TVB-N), pH, and water holding capacity (WHC) of the fillet, which would be a benefit for precisely monitoring fish quality online.