Application of an Integration of Uninformative Variable Elimination and Least-Squares Support Vector Machine for Discriminating Soy Milk Powder Based on Visual Near-Infrared Spectral Calibration: Disc

A novel method which is combination of uninformative variable elimination by partial least squares (UVE) and least-square support vector machine (LS-SVM) was proposed to discriminate soy milk powder. A total of 240 (60 for each variety) samples were characterized on the basis of visual and infrared spectroscopy (VIS-NIR), 160 (40 for each variety) samples were selected randomly for the calibration set, whereas, the remaining 80 samples (20 for each variety) for prediction set. UVE was executed to obtain the stability of each input variables. 327 wavelengths were selected by UVE, and inputted into LS-SVM to build recognition model. The classification rate reached 100%, and the performance was better than common LV-SVM model which was established using the whole spectral wavelengths. The overall results indicated that the proposed method of UVE-LS-SVM based VIS-NIR technology was a powerful way for discrimination of different varieties of soy milk powder.