Rapid determination of chemical compositions in chicken flesh by mining hyperspectral data

Abstract

Hyperspectral data in 900–1700 nm range was mined for rapid and non-destructive determination of chicken chemical compositions. Hyperspectral images of chicken flesh samples were acquired and spectral data within the hyperspectral images were extracted, preprocessed to relate to moisture, protein and ash contents by partial least squares (PLS). The results showed that PLS model built with Savitzky-Golay convolution smoothing (SGCS) spectra (SGCS-PLS model), RAW spectra (RAW-PLS model), normalization (NOR) spectra (NOR-PLS model) had the best performance in predicting moisture (RP = 0.925, RMSEP = 0.391%), protein (RP = 0.926, RMSEP = 0.589%) and ash (RP = 0.938, RMSEP = 0.114%) contents, respectively. Four different methods including regression coefficients (RC), stepwise regression (SR), successive projections algorithm (SPA) and competitive adaptive reweighted sampling (CARS) were then applied to select optimal wavelengths for simplifying the original PLS models. As a result, 23 and 15 optimal wavelengths were respectively selected by CARS method from SGCS spectra (for moisture) and NOR spectra (for ash) and the original PLS models were optimized (CARS-SGCS-PLS model, CARS-NOR-PLS model) with RP of 0.932, 0.932 and RMSEP of 0.376%, 0.120%, respectively. Thirteen optimal wavelengths were selected by RC method from RAW spectra (for protein) and RC-RAW-PLS model was constructed with RP of 0.923 and RMSEP of 0.588%. F test and t test results further verified the statistical soundness of the three best optimized models. The chemical maps of moisture, protein and ash were finally generated by transferring the models into each pixel of images of chicken samples to visualize the chemical distribution. The overall study indicated that hyperspectral data combined with PLS algorithm could be potentially used for predicting the chemical compositions of chicken flesh.