[Application of PCA-SVR to NIR prediction model for tobacco chemical composition]

Near infrared diffuse reflectance spectra of 50 tobacco samples were pretreated with PCA. The calibration models of determination of the main components in tobacco were developed with support v ector regression (SVR). The models weretested with leave-one-out (LOOCV) method and optimized with parameters of kernel function, penalty coefficient C and insensitive loss function. The root mean square errors (RMSE) with leave-one-out cross validation of the optimal models of nicotine, and total sugars, reductive sugar, and total nitrogen were 0.313, 1.581, 1.412 and 0.117 respectively. Based on the comparison of RMSE of the SVM model with those of the partial least square (PLS), multiplicative linear regression (MLR) and back propagation artificial neuron network (BP-ANN) models, it was found that the SVR model was the most robust one. This study suggested that it is feasible to rapidly determine the main components concentrations by near infrared spectroscopy method based on SVR.