Land surface temperature(LST)is an important parameter in the model of energy balance of the earth surface. The enhanced spatial resolution of high temporal resolution of remote sensing surface temperature can be realized by downscaling algorithm, which is of great significance for monitoring the spatial and temporal distribution of the LST. In this paper, Beijing City was taken as the study area, and the LST with 100 m spatial resolution was retrieved by using Landsat8 OLI/TIRS data through improved mono-window(IMW)algorithm,which was used as validation data. Besides,the normalized difference vegetation index(NDVI),normalized difference built-up index(NDBI)and other remote sensing index were calculated and simulated to the spatial resolution of 1 000 m, which was united with the MODIS/LST with the spatial resolution of 1 000 m to be input into the random forest(RF)model to acquire downscaled LST(100 m). Meanwhile, the downscaled results retrieved by RF algorithm were compared with the two commonly used methods of downscaling, multi factor regression method and LST sharpening algorithm based on vegetation index (TsHARP). The results show that, with the simulated Landsat/LST as the data source, the RMSE of downscaling LST retrieved by RF was 2.01 K, and the RMSE was improved by 0.16 K and 0.44 K compared with the multi factor regression method and TsHARP algorithm respectively. For the MODIS/LST, the RMSE of downscaling LST retrieved by RF was 2.29 K, and the RMSE was improved by 0.42 K and 0.50 K compared with multi factor regression method and TsHARP algorithm respectively. For different land surface types, the effects of RF downscaling algorithm are different. The effect of high vegetation coverage area is the best, and the RMSE is 1.81 K. Due to the spatial heterogeneity of the urban surface, the RMSEhas reached a maximum of 2.75 K.