MODIS FAPAR products ranging from 2001 to 2010, which cover the whole basin of Heihe, were selected as the study objects. Combined with the land cover classification products in the same periods, MODIS FAPAR time series data of different species were decomposed into tendencies, cyclical fluctuations and residuals. Auto Regression models were then established for tendencies and cyclical fluctuations and Kalman filtering was used to remove the noises. At last, FAPAR time series with high quality were obtained. On this basis, the FAPAR variation differences of different vegetation types in upstream, midstream and downstream of Heihe Basin are analyzed. A region with different phase change characteristics was chosen to predict FAPAR value of each pixel at some point. Results show that FAPARs of different species own obvious seasonal variations. Because of different climates in different basins, FAPARs of same species also have slight differences. In addition, the proposed method can be used for different vegetation types and prediction results are quite similar to MODIS FAPAR values, with an error of about 3%.