Retrieving atmospheric CO2 concentration from space-based infrared measurements is an ill-posed problem, and the uncertainty of ground properties is an important impacting factor. For the requirement of high precision retrieval of atmospheric CO2, the effects of surface reflectance must be corrected. Thus a surface reflectance correction method is proposed. In this correction method, the selection of non-CO2 absorption (off-line) channel is an important part, so several different types of off-line channels were compared and analyzed. We finally found that averaging all data of multi-channels as off-line channel is the best way, in which the retrieval precision is highest and dispersion is lowest, because the advantage is that averaging many data can reduce random error. The effects of surface reflectance on CO2 retrieval decreased significantly after correction using ratio spectrometry combined with the selected off-line channel, and CO2 retrieval precision improved greatly.