NIR Model Construction And Application For Diterpene Glycosides Composition In Stevia Rebaudiana

Near infared reflectance spectroscopy (NIRS) has been rapidly developed as a novel physical analysis technique in the late of 1980s in last century. Because the NIRS is nondestructive, fast, cost effective, environmentally safe, and allows the simultaneous estimation of several traits in a unique measurement, this technique was used widely in many areas and considered as international stardards. The diterpene glycosides in Stevia rebaudiana leaves are considered as a potential source of natural non caloric sweeteners and used widely in the food, medicine, cosmetic, etc. The present study aimed for assessing the potential of NIRS technique to estimate the stevioside, rebaudioside A and their total contents in Stevia rebaudiana leaves and to optimize the suitable regression method and variables space to develop a robust and accurate regression-model.A total of 508 samples selected randomly from the individual plants with good agronomic traits from 2008 to 2009. The percentage contens of stevioside and rebaudioside A of the leaf samples were determined by the reference method of HPLC. About 3g leaf powder of each sample was scanned from 400nm to 2498nm at the interval of 0.5nm. The entire spectrum was pretreated with the standard normalized variate, second derivatived and Savitzky-Golay convolution smoothing. For the pretreated spectrum in 350 samples in train set, Monte-Carlo uninformative variables elimination and successive projections algorithm were used to optimize the variable space, reduce the collinearity and overcome the overfitting.Based on the optimized variables space, the prediction model was developed by the insensitive loss function-support vector regression method after the outliers removed by using iterative reweighted least squares support vector regression. The hybrid method is superior to other methods, which has been certificated by the simulation data and the total glycosides content data with the smaller prediction risk and the better generalization. Further more, the extracted feature extracted by partial least squares was used as the inputs to construct the NIR calibration model. It is feasible to determine the stevioside, rebaudioside A and their total contents of them in Stevia leaves with the low root mean square error of prediction, high determination coefficient, and satisfactory residual predictive deviation. By using the developed models to screen the individual plants,133 parental materials and 50 F1 lines with absolutely or relatively high rebaudioside A content were primarily identified and tested subsequently by HPLC. Briefly, the developed model could be directly to predict the diterpene glycosides in Stevia leaves and had good performance in breeding project.