The selection of wavelength points and modeling methods is very important for the establishment of predictive analysis model in near infrared spectroscopy. On the basis of the traditional correlation coefficient method, a correlation coefficient threshold optimization method based on the genetic algorithm is proposed. In the method, the determination coefficient is maximized so as to find the optimal threshold. A radial basis neural network is used to establish a calibration model and an orthogonal least square method is used to select the center for modeling. The new method is used to predict the dimethy carbonate content in gasoline and its result is compared with the experimental result of the partial least square method. The result shows that the new method has a better prediction accuracy. Its determination coefficient is up to 0.9993.