Since the development of remote sensing, studies on urban heat islands (UHIs) have mainly focused on surface UHI (SUHI), which is more than the atmospheric UHI (AUHI). Remote sensing advancements provide strong technical support for SUHI studies. However, the acquisition of air temperature (AT), especially high-resolution large-scale AT data, remains inconvenient. In this study, we selected the Greater Tokyo Area as a case study. We used the extra-trees (ET) model to downscale the AT from 1 km to 250 m based on the regression relationship among AT, digital elevation model (DEM), and land use/land cover (LULC). The downscaled AT results were obtained after residual fitting. Finally, we compared the downscaled AT with meteorological station data to verify accuracy. The results indicated the following: (1) the ET model and the combination of independent variables, including DEM and LULC, could be applied to AT downscaling research; (2) the downscaling accuracy of highly heterogeneous regions was significantly higher than that of highly homogeneous regions; and (3) high-resolution AT data could accurately describe the AUHI of the study area. Finally, an accurate, efficient AT downscaling method based on high-resolution LULC was proposed.