The invention discloses a method for reducing that scale of surface temperature space. At first, that method quantitatively analyzes the surface temperature and the surface parameters including the impervious surface coverage, vegetation coverage, soil coverage, NDVI, NDBI, MNDWI, DEM, and the correlation between building density and its spatial distribution difference, Then the regression model of low spatial resolution land surface temperature products and related land surface parameters is established by using machine learning stochastic forest algorithm, land surface temperature with highspatial resolution can be predicted by combining the land surface parameters with high spatial resolution, Then the area-to-point Kriging interpolation method of geostatistics is used to reduce the residuals of the random forest regression model to improve the spatial resolution of the residuals, Finally, the high spatial resolution stochastic forest regression model and the surface-to-point Kriging interpolation residuals are added to generate high resolution and high precision surface temperature products to make up for the lack of spatial resolution of the existing surface temperature products.