Land surface temperature (LST) is a key parameter to describe the process of surface-atmosphere interaction and reflect the surface condition. However, due to the constraints of the imaging conditions of the thermal infrared sensor, there is a contradiction in the spatial and temporal resolution of the remote sensing images. This contradiction prevents the LST data being fully utilized. The LST downscaling algorithm can effectively resolve the contradiction of spatiotemporal resolution, and obtain LST images with higher spatiotemporal resolution. The LST downscaling model is gradually studied from the global model to the local model. In the current studies, the linear relationship between the surface temperature and the auxiliary parameters is considered, however, the local nonlinear relationship is ignored. This study proposed a new algorithm based on the non-linear geographically weighted regression (NL-GWR) model to downscale the LST. This study selects the proper research area,and selects Normalized Difference Vegetation Index (NDVI), Normalized Difference Build-up Index (NDBI) and digital elevation model (DEM) respectively as an auxiliary parameters,this study proposed an LST downscaled model to downscale spatial resolution of the moderate resolution imaging spectroradiometer (MODIS) LST data from 1 000 to 100 m, and analyzed the LST downscaling results, and the results of downscaling based on geographic weighted regression (GWR) model and NL-GWR model were compared and analyzed. The experimental results show that under the same parameters, the NL-GWR model considering the nonlinear relationship is better than the GWR linear model, and the lowest root mean square error (1.96 ℃) and mean absolute error (1.63 ℃).