Land surface temperature (LST) plays a very important role in obtaining the thermal information of land surface. LST has been widely used to detect the agricultural drought, monitor evapotranspiration, and estimate net radiation, sensible and latent heat flux, surface urban heat island intensity, precipitable water, urban-induced surface moisture and surface runoff. However, due to technical constraints of remotely sensed LST data, there is a trade-off between spatial and temporal resolution: a high temporal resolution is associated with a low spatial resolution and vice versa. Improvement in the spatial resolution of LST image will be greatly important when the revisit time of satellite remains the same. To solve this disadvantage, the aims of this study include the following aspects: firstly, to improve the original low spatial resolution of the Moderate Resolution Imaging Spectroradiometer (MODIS) LST image data (1km) to 500m resolution by applying three relationships between NDVI and LST in TsHARP (An algorithm for sharpening thermal imagery) method; secondly, to assess the accuracy of each downscaling model based on qualitative and quantitative analysis with synchronous Enhanced Thematic Mapper (ETM) LST data; thirdly, to examine whether the TsHARP downscaling method is applicable and valid in special and particular land cover types in China. The northwest of Shandong province and south of Hebei province in China were selected as the research area. Then we utilized the TsHARP method to downscale original 1km LST image (taken on 9^th, September 2004, provided by MODIS sensor) to 500m LST image, given that MODIS can offer very high temporal resolution image. In order to validate the simulated 500m LST image, we compared the simulated LST_MODIS,500m image with the synchronous 500m ETM LST image generated from 500m ETM band 6 radiance by Sobrino technique, and the 500m ETM band 6 radiance image was upscaled from the original 120m ETM band 6 radiance image. Qualitative comparison results show that LST_MODIS,500m images visually capture more spatial details than LST_MODIS,1km and resemble more that of LST_ETM,500m. In the downscaled 500m MODIS LST image, boundaries between different land covers are highlighted. When compared to 500m ETM LST, statistical results indicate that the Mean Absolute Error (MAE) in 500m MODIS LST generated by the downscaling techniques ranged from 0.536℃ to 0.752℃. Quantitative comparisons of the three downscaled LST_MODIS,500m images and LST_ETM,500m image show that Root Mean Square Error (RMSE) are 0.786℃, 1.002℃ and 0.754℃, respectively, with method 3 (i.e. vegetation fraction is linearly correlated to 4th power of LST) corresponding to both the lowest MAE (0.536℃) and RMSE (0.754℃). The variance (S²) of method 2 (i. e. vegetation fraction is linearly correlated to LST) is the highest, which shows that method 2 captures more spatial details of LST_ETM,500m. Considering the mean values of different methods, the mean of method 2 is also the most consistent one to the mean of the LST_ETM,500m. As for the LST accuracy, an RMSE of 0.754℃ or 1.002℃ is regarded as reasonable and acceptable, given the change in the spatial resolution of the MODIS LST and the potential sources of error due to different sensor properties. Above all, all of the three methods were found to substantially achieve good performance and produce 500m LST results with the high accuracy and improve the spatial pattern of the 1km MODIS LST. The downscaling techniques were proved effective and applicable in our study area.