Abstract: Due to technical constrains of remotely sensed land surface temperature (LST) data, there is a tradeoff between spatial and temporal resolution, i.e., a high temporal resolution is associated with a low spatial resolution and vice versa. To solve this disadvantage, based on three different conversion models in TsHARP (an algorithm for sharpening thermal imagery) method, this study made improvements by introducing a new parameter, namely, land surface emissivity (ε), and a new conversion model was put forward by replacing the original NDVI with MSAVI. The aims of this study involved the following aspects: firstly, to improve the original low spatial resolution of MODIS LST image data (1 km) to 250 m resolution by applying four models; se-condly, to assess the accuracy of each downscaling model based on qualitative and quantitative analysis with synchronous Landsat 8 TIRS LST data. The results showed that both models could effectively enhance the spatial resolution while simultaneously preserve the characteristics and spatial distribution of the original 1 km MODIS LST image, and also eliminate the “mosaic” effect in the original 1 km image. Statistical results indicated that absolute mean basis error (MBE) in 250 m MODIS LST generated by the four models was less than 1 ℃. Thus, the results were very close to the true surface temperature, indicating that all of the four models were effective and applicable in our study area. Quantitative comparisons of the four downscaled 250 m MODIS LST images and TIRS LST showed that the new conversion model (i.e. model 4) had the lowest RMSE (1.252 ℃), while the original model 1 had the highest RMSE (2.375 ℃). In addition, model 4 had obvious advantages in the description of high temperature areas (i.e. urban area) and low temperature areas (i.e. water bodies).