In order to overcome the influences of multi-collinearity and absorbance non-linearity in near-infrared spectroscopy on predicting soil total nitrogen content, the gray correlation-extreme learning machine method was used to select the combination wavebands with good prediction capability to establish high precision prediction model for soil total nitrogen content. First, the first derivative spectra was used to get the sensitive spectrum area. And then the grey correlation sensitive wavelength selection method was used to select wavelengths which were respectively 1007, 1128, 1360, 1596, 1696, 1836, 2149 and 2262nm. Finally, by using the above sensitive wavelengths as input data, a soil total nitrogen prediction model was established based on the method of extreme learning machine and multiple linear regression. As a comparison, while using the traditional correlation analysis method to select the sensitive wavelengths, the results showed that R2c of the soil total nitrogen forecast model established by using gray correlation-extreme learning machine was 0.9134, and the prediction R2v was 0.8787. Its accuracy was higher than that of the traditional modeling method. It indicated that the gray correlation-extreme learning machine method had more obvious advantages especially in the prediction of low soil total nitrogen content.