Spectral analysis has been increasingly applied to estimate plant species diversity through the world, especially for biodiversity field. Although spectral variability hypothesis (SVH) has been widely proved in estimating plant alpha diversity for tropical, temperate, and sub-tropical forests, meadow, steppe and grasslands, however, the performance on beta diversity is still lack. In this study, we measured the hyperspectral reflectances and plant species diversity indices of 270 plots at a fine scale (0.8 meter) in central Hunshandak sandy grasslands of Inner Mongolia, China. 195 plots were used as training data and 75 plots as validating data. Bray-Curtis dissimilarity index (BC), Srensen index (S) and Jaccard index (J) were calculated to indicate actual beta diversity. Based on spectral biological features of different plant species, 164 hyperspectral indices were developed and used to assess plant species beta diversity. Pearson’s correlation analysis and multiple linear stepwise regression were conducted based on sensitive wavebands to produce hyperspectral models. The hyperspectral indices which high Pearson’s correlation coefficients will be remained for further tested. Communities with different coverages and richness were also used to test the robustness of proposed models. By comparing the stability of hyperspectral indices under different communities, the indices with high stability is remained for validation by 75 plots. Results demonstrated that BC, Euclidean distances of first-order derivation values between 400～1 000 nm, and BC of 760～800 nm could accurately estimate species beta diversity. BC can be accurately estimated by hyperspectral indices, since they were both calculated as parameters of the distance between plots. The Jaccard and Srensen indices were hardly estimated, it is hard to find the suitable wavebands or other parameters in spectral data to replace the “common reflectance” between pairwise plots. This study promotes the development of methods in assessing plant species beta-diversity using hyperspectral data.