基于高光谱成像技术的冷鲜猪肉水分和脂肪含量同步快速检测OA

This study aims to develop a rapid method for determining the moisture and fat contents in chilled pork using visible-near-infrared(VIS-NIR)/NIR hyperspectral imaging technology.Traditional laboratory methods were used to measure the moisture and fat contents of 128 chilled pork(longissimus dorsi muscle)samples.Hyperspectral data were collected in the VIS-NIR(388-1 045 nm)and NIR(930-1 710 nm)ranges.Three prediction models were constructed using partial least squares regression(PLSR),one-dimensional convolutional neural network(1DCNN),and recurrent neural network(RNN)based on the full-band spectra and were compared.The effects of different spectral data preprocessing methods:Savitzky-Golay smoothing(S-G),S-G first derivative(S-G 1'),and S-G second derivative(S-G 2')and different feature wavelength extraction methods:successive projections algorithm(SPA)and two-dimensional correlation spectroscopy(2DCOS)on the prediction accuracy of the PLSR model.Among the three models,the PLSR model was selected considering its stability as the best model for predicting the moisture and fat content in chilled pork.For both VIS-NIR and NIR ranges,the model based on the raw data was more accurate than those based on S-G,S-G 1'or S-G 2'preprocessed data.To simplify the process,the raw NIR data were selected for modeling.Compared to the full-band PLSR models,the PLSR models built using the feature wavelengths showed slightly reduced predictive performance.Under NIR,the optimal moisture prediction model,built using S-G,SPA and PLSR,with coefficient of determination of prediction of 0.71,performed marginally better than did the full-band model,indicating that feature wavelength extraction can,in certain cases,improve model construction.