Abstract: The temporal resolution of high spatial resolution remote sensing data can be effectively improved by spatio-temporal fusion of remote sensing data. However, the most widely used spatial and temporal adaptive reflectance fusion model (STARFM) fails to achieve highly accurate prediction effects for areas with abrupt changes at present. Given this, this paper proposed a hierarchical spatial-temporal fusion model (H-STFM). In this model, the target pixels to be predicted are divided into pixels with phenological change and pixels with abrupt changes, which are predicted using linear regression and weighted filtering methods, respectively. Then the prediction results of the two types of pixels are fused using an optimized time weighted function to form the final prediction map. The H-STFM proposed in this paper was qualitatively and quantitatively assessed using two sets of medium-resolution remote sensing images from moderate resolution imaging spectrometer (MODIS) and Landsat satellite. As indicated by the experiment results, H-STFM is significantly superior to STARFM in terms of structural similarity and relative dimensionless global error.