The successful launch of the GF-2 satellite indicates that China's remote sensing satellites have entered the era of high spatial resolution of the sub-meter level. Remote sensing images will play an important role in quantitative inversion, ground object recognition, and change analysis. The accuracy of its atmospheric correction is an important factor that affects its quantitative application. Due to the lack of a short-wave infrared band in GF-2, it is impossible to use a dark pixel method for atmospheric correction. A method of atmospheric correction for the GF-2 image based on a radiation transfer model is proposed. The atmospheric correction coefficient lookup table is established by using 6S (Second Simulation of the Satellite Signal in the Solar Spectrum) radiation transfer model. The aerosol optical thickness (AOT) is retrieved by combining synchronous MODIS image data with an improved dark pixel method. The atmospheric correction parameters are determined to eliminate the influence of absorption and scattering of atmospheric molecules and aerosols in the GF-2 image and to achieve atmospheric correction of GF-2 data. Dunhuang radiation correction field with a flat and uniform surface is selected as the experimental area. The accuracy of the correction results is evaluated by synchronous measured data, and the normalized difference vegetation index (NDVI) before and after atmospheric correction is compared. The results show that the minimum relative error is only 0.9%. The image data after atmospheric correction can accurately reflect the reflection characteristics of ground objects. NDVI after atmospheric correction greatly enhances the contrast of vegetation information and highlights the ability of vegetation information discrimination of the GF-2 satellite sensor.