Hyperspectral remote sensing is an important technique to fulfill real-time monitoring for crop growth status based on its superior performance in acquiring vegetation canopy information rapidly and non-destructively. The objective of this study was to establish the best simulating accuracy and adaptability of wheat fraction of absorbed photosynthetically active radiation (FPAR) estimation model based on wheat canopy hyperspectral reflectance with different nitrogen or phosphorus application rate treatments, and to improve the forecast precision of the FPAR estimation model at different growth stages of dryland wheat on the Loess Plateau. The experiments were carried out during 2009—2014 at Northwest A&F University, Yangling, China. Different winter wheat (Triticum aestivum L.) varieties with stronge or weak drought resistance were chosen for the treatments in different years, nitrogen and phosphorus treatments included five nitrogen fertilizer application rates (0, 75, 150, 225 and 300kg/hm2 pure nitrogen, expressed as N) and four phosphorus application rates (0, 60, 120 and 180kg/hm2P2O5,expressed as P), the FPAR and canopy hyperspectral reflectance of different varieties and fertilizer treatments were monitored at jointing, booting,heading, grain filling and maturity stage, respectively. Then FPAR monitoring models at different growth stages of winter wheat were constructed by using correlation analysis, regression analysis and other methods. The results showed that the FPAR of wheat increased with nitrogen and phosphorus application rate increasing in different growth stages, there were significant differences among test cultivars. A good correlation relationship was presented between FPAR and canopy spectral reflectance at 670, 850 and 960nm, and the sensitive band of the FPAR occurred mostly within visible and near-infrared spectrum. The correlations between soil adjusted vegetation index (SAVI), red edge normalized difference vegetation index (NDVI705), enhanced vegetation index (EVI), difference vegetation index (DVI) and ratio vegetation index (RVI) to FPAR were significant, and the range of correlation coefficient was 0.818~0.942 at jointing, booting, heading, filling and maturity stages. Monitoring models based on SAVI, NDVI70, EVI, RVI and RVI produced better estimation for FPAR at different growth stages, and the determination coefficients (R2) were 0.854, 0.888, 0.811, 0.844 and 0.911, and the standard errors (SE) were 0.054, 0.032, 0.044, 0.047 and 0.044, accordingly. Meanwhile, comparing the predicted value with measured value to verify reliability and applicability of monitoring model, result showed that the relative errors (RE) between measured value and predicted value were 14.1%, 17.4%, 12.8%, 18.8%, 10.7%, and the root mean square errors (RMSE) were 0.139, 0.146, 0.136, 0.158, 0.130, respectively. Therefore, it was suggested the vegetation indices of SAVI, NDVI705, EVI, RVI and RVI was the most suitable model for monitoring winter wheat FPAR at jointing, booting, heading, filling and maturity stages, respectively, and there was higher prediction precision with different vegetation indices in monitoring FPAR of winter wheat at different growth stages, and different N and P rates. These conclusions had important implications for the large areas FPAR monitoring of winter wheat in the Loess Plateau. Meanwhile, there was a higher prediction accuracy of monitoring model based on the different vegetation indices at different growth stages.