The land surface temperature (LST) is one of the most important parameters of surface-atmosphere interactions. Methods for retrieving LSTs from satellite remote sensing data are beneficial for modeling hydrological, ecological, agricultural and meteorological processes on Earth’s surface. Many split-window (SW) algorithms, which can be applied to satellite sensors with two adjacent thermal channels located in the atmospheric window between 10 μm and 12 μm, require auxiliary atmospheric parameters (e.g., water vapor content). In this research, the Heihe River basin, which is one of the most arid regions in China, is selected as the study area. The Moderate-resolution Imaging Spectroradiometer (MODIS) is selected as a test case. The Global Data Assimilation System (GDAS) atmospheric profiles of the study area are used to generate the training dataset through radiative transfer simulation. Significant correlations between the atmospheric upwelling radiance in MODIS channel 31 and the other three atmospheric parameters, including the transmittance in channel 31 and the transmittance and upwelling radiance in channel 32, are trained based on the simulation dataset and formulated with three regression models. Next, the genetic algorithm is used to estimate the LST. Validations of the RM-GA method are based on the simulation dataset generated from in situ
measured radiosonde profiles and GDAS atmospheric profiles, the in situ
measured LSTs, and a pair of daytime and nighttime MOD11A1 products in the study area. The results demonstrate that RM-GA has a good ability to estimate the LSTs directly from the MODIS data without any auxiliary atmospheric parameters. Although this research is for local application in the Heihe River basin, the findings and proposed method can easily be extended to other satellite sensors and regions with arid climates and high elevations.