Timely, accurate and rapid prediction of grain yield is of great significance for guiding agricultural production and formulating national food policy. A split-plot design experiment, taking different rice varieties and nitrogen application levels as experimental factors, was conducted to measure the canopy spectral reflectance at rice jointing stage, booting stage and heading stage. By selecting the best band combination with the highest correlation with yield, 12 vegetation indexes composed of the best band combination were calculated, and a rice yield prediction model based on the combination of single vegetation index and multi vegetation index was established. The results showed that at booting stage, there were a significant negative correlation between the original spectral reflectance of rice canopy and yield in the band of 401~723 nm, the correlation between each vegetation index and yield reached a very significant level. The rice yield prediction model based on single vegetation index has the highest precision (R²=0.436, RMSE=874.57 kg/hm²) in the linear model at booting stage, and the best vegetation index is the normalized vegetation index (RDVI), and the model expression is y=7.7E+05×RDVI_{(455, 456)}+1.1E+04. The multi vegetation index yield prediction model based on stepwise regression also showed the best performance at booting stage (R²=0.443, RMSE=861.81 kg/hm²), the optimal vegetation index was ratio vegetation index (RVI), soil regulated vegetation index (SAVI) and optimal vegetation index (VIopt), and the model expression was y=1.8E+05×RVI_{(1661, 1687)}-2.1E+05×SAVI_{(1235, 1268)}+5.3E+04×VIopt_{(2260, 2215)}-3.4E+05. In general, the fitting accuracy and prediction effect of the multi vegetation index yield prediction model are better than those of the single vegetation index yield prediction model, and the simulation effect at booting stage is the best.