Many application fields need land surface temperature (LST) with simultaneous high spatial and temporal resolution, which can be achieved through the disaggregation technique. Most published methods built an assumed scale-independent relationship between LST and predictor variables derived from coarse spatial resolution data. However, LST disaggregation in the heterogeneous areas, especially urban areas, is very difficult to achieve and there are few studies on it. In this article, we propose an adjusted stratified stepwise regression method for temperature disaggregation in urban areas. Landsat Enhanced Thematic Plus (ETM+) data from Shanghai, China, were used to construct remote-sensing indices that are related to LST variance and retrieve LST at 60 and 480 m spatial resolution, respectively. Different stepwise regression models at 480 m resolution were built for two stratified regions according to normalized difference vegetation index (NDVI) distribution, and then each independent variable at 60 m resolution was adjusted to calculate disaggregated LST by considering its relationship with the 480 m resolution image. By using LST retrieved directly from ETM+ band 6 at 60 m spatial resolution as the reference, the proposed method comprising resampling disaggregation, the thermal data sharpening model (TsHARP)/disaggregation procedure for radiometric surface temperature (DisTrad) technique, and the LST-principal component analysis (PCA) regression algorithm were verified and compared. The results show that the temperature distribution estimated using the proposed method is most consistent with that of the reference LST in this heterogeneous study area, and that the precision improves significantly, especially for the low vegetation fraction region.