Retrieval of land surface temperature (LST) is one of the key issues in mid- and thermal-infrared quantitative remote sensing. Compared with thermal-infrared (8~14μm) channel, mid-infrared (3~5μm) channel with the advantages, such as more transparent of atmospheric window and less sensitive to the surface emissivity, might be used to retrieve the LST. A two-channel LST retrieval algorithm, which only uses the night-time data of two mid-infrared channels and introduces the experience of thermal infrared LST retrieval method, has been proposed in this paper. The retrieval accuracies of two channel-combination models, i.e., narrow-channel model (3.929~3.989 μm and 4.020~4.080 μm) and wide-channel model (3~4μm and 4~5μm), were analyzed in detail. The results show that the LST retrieval accuracies of wide-channel and narrow-channel models are approximate ~ 0.5K and ~ 0.3K, respectively. However, compared with the wide-channel model, the noise equivalent temperature difference (NEΔT) and the error of land surface emissivity (LSE) can produce a greater influence on the narrow-channel model.