Methane (CH4) ranks as the second-largest greenhouse gas. Although satellite remote sensing is widely employed to track CH4, there is a disparity between the column concentrations obtained from satellite retrievals and the near-surface methane concentration. At present, the research on converting satellite-derived column concentrations into near-surface methane concentrations focuses on exploring the influence of satellite column concentrations and multi-source geographical factors. However, this approach overlooks critical spatial and temporal information, resulting in inaccurate estimation results. In this paper, we introduce an approach to estimating methane concentration near the surface through the application of two widely used machine learning techniques. Our input variables comprise column concentration data obtained from the Sentinel-5 Precursor, along with meteorological, spatial, and temporal factors. The experimental results reveal that (1) the LightGBM model which incorporates spatio-temporal data outperforms the XGBOOST model in accuracy on the test set, with a cross-validation correlation coefficient of 0.9, a root mean square error of 25.96 ppb, and a mean absolute percentage error of 0.79%; (2) estimation precision has improved for these two machine learning models that consider spatio-temporal information as model input features on the test set; and (3) these two machine learning algorithms that consider spatio-temporal information exhibit superior spatial extrapolation.