In order to accurately obtain land surface temperature (LST) from stratospheric platform, a simulation data set was established based on the SeeBor atmospheric profile training dataset and MODIS split window channel parameters. Firstly, the effectiveness of the general split window algorithm obtaining surface temperature at stratospheric altitude was analyzed; then, the sensitivities of several key parameters such as observation height, emissivity, water vapor content, noise equivalent temperature difference (NETD), and spectral response drift were analyzed. The LST simulation shows a theoretical error of 0.185 K, in which the emissivity and NETD dominate inversion error. Finally, the deviation constraint demand of different influencing factors is analyzed. Results show that if the uncertainty of emissivity and NETD can be controlled within 1% and 0.4 K, respectively, with spectral response drift within ±3 nm, the temperature inversion accuracy can be better than 2 K.