The growth of grassland in Sanjiang Source Area has an important influence on the development of livestock husbandry and the ecological balance of Sanjiang ecosystem. It is of great importance to estimate grass yield reasonably and accurately. In view of the larger area and the complicated surface structure, this study is aimed at predicting the grass yield by using MODIS product data. The authors built a prediction model of grass yield in Sanjiang by using six kinds of MODIS products (LAI, FPAR, NDVI, EVI, GPP and LST) from April 2009 to October 2009 and, in combination with partial least squares regression (PLS) and multiple linear regression method, accomplished estimation of grass yield by remote sensing. Based on the built model, the authors used the 140 scene data from April to October 2011 for application testing, and then compared the predicting results with standard values which were measured from June to August 2011 in 16 grassland ecological monitoring stations in Sanjiang. The results show that there is a good correlation between grass yield estimated based on the six MODIS products and the measured actual grass yield. A comparison with the result of multiple linear regression shows that the result of PLS has a higher coefficient (R ²≈0.829~0.878) and lower root mean squared error (RMSE≈42.457~93.674 kg·hm ^-2).