Carbon dioxide (CO2) is one of the most important greenhouse gases in the atmosphere. Accurate understanding its content and variation in the atmosphere can provide support for climate change prediction and environmental decision-making. In order to meet the requirement of climate research, the retrieval precision of column-averaged dry-air mole fraction of CO2 (XCO2) observed by satellite should be better than 1%. The greenhouse gases monitoring instrument (GMI) onboard GF-5 used for atmospheric CO2 and CH4 observation adopts a novel spatial heterodyne spectroscopy technology, which can supply hyperspectral resolution. Considering that satellite observations with hyperspectral resolution are easily effected by atmosphere, groundsurface and instrument, a retrieval method special for GMI data is designed in this work. In the method, through the analysis of spectrum information, the information spectrum for atmospheric CO2 retrieval and the reference spectrum for removal of slow frequency interference are constructed firstly, then the measurement vector and simulation vector are rebuilt based on the constructed spectra, and at last atmospheric XCO2 is retrieved by using the optimization estimation method. As the method has the ability of slow frequency interference removal, it makes retrieval of GMI spectrum possible. By using the proposed retrieval method, the GMI observations between August 2018 and March 2019 are retrieved, and the ground-based carbon dioxide measurements from twelve stations of Total Carbon Column Observing Network are used to validate the retrieval results. The validation results show that the algorithm is able to retrieve the GMI observations stably, and the retrieval precision of atmospheric CO2 is high as 0.67%, which is better than the basic requirement of 1%.