[Geographical origin discrimination of Auricularia auricula using variable selection method of modeling power]

Near infrared (NIR) spectroscopy combined with variable selection method of modeling power was investigated for the fast and accurate geographical origin discrimination of auricularia auricula. A total of 240 samples of auriculari auricula were collected in the market, and the spectra of all samples were scanned within the spectral region of 1100-2500 nm. The calibration set was composed of 180 (45 samples for each origin) samples, and the remaining 60 samples were employed as the validation set. The optimal partial least squares (PLS) discriminant model was achieved after performance comparison of different preprocessing (Savitzky-Golay smoothing, standard normal variate, 1-derivative, and 2-derivative). The effective wavelengths, which were selected by modeling power (MP) and used as input data matrix of least squares-support vector machine (LS-SVM), were employed for the development of modeling power-least squares-support vector machine (MP-LS-SVM) model. Radial basis function (RBF) kernel was applied as kernel function. Three threshold methods for variable selection by modeling power were applied in MP-LSSVM models, and there were the values of modeling power higher than 0.95, higher than 0.90, and higher than 0.90 combined with peak location (0.90+Peak). The correct recognition ratio in the validation set was used as evaluation standards. The absolute error of prediction was set as 0.1, 0.2 and 0.5, which showed the wrong recognition threshold value. The results indicated that the MP-LS-SVM (0.90+Peak) model could achieve the optimal performance in all three absolute error standards (0.1, 0.2 and 0.5), and the correct recognition ratio was 98.3%, 100% and 100%, respectively. The variable selection threshold (0.90+Peak) was the most suitable one in the application of modeling power. It was concluded that modeling power was an effective variable selection method, and near infrared spectroscopy combined with MP-LS-SVM model was successfully applied for the origin discrimination of auricularia auricula, and an excellent prediction precision was also achieved.