Fast and nondestructive methods for assessing plant community species diversity are topics of great scientific interest and concern to ecologists worldwide. Hyperspectral data, which have the advantage of hundreds of spectral wavebands and multiple spatial scales, have huge potential for assessing plant species diversity. However, identifying the most sensitive wavebands is still a challenge. In this study, we measured the hyperspectral data and plant diversity indices of 120 samples from sandy grasslands in central Hunshandak Sandland, Inner Mongolia, China. Ninety plots were used as training data and thirty plots as validating data. After pre-processing, sensitive wavebands were selected using Pearson's correlation analysis, principle component analysis (PCA), and experienced selection. Multiple linear stepwise regression (MLSR) was conducted based on sensitive wavebands to produce hyperspectral models. The results demonstrated that the regression models based on PCA bands could accurately estimate the Pielou (r=0.65^{* *}), Simpson (r=0.49^{* *}), and Shannon-Wiener indices (r=0.40^*). Communities with different coverages were also used to test the robustness of proposed models based on PCA bands. We propose that the Simpson, Shannon-Wiener, and Pielou indices, widely used as indicators of plant species alpha diversity, can be precisely estimated by hyperspectral indices at a fine scale. Community complexity and coverage can substantially affect the accuracy of estimating plant diversity. First-order derivations of vegetation reflectance can reduce environmental noise, water absorption disturbance, and reflect differences among species, hence greatly improving the estimation accuracy. This study promotes the development of methods in assessing plant community diversity using hyperspectral data. Based on the results of this study, future study topics are also suggested.