Desert steppe is the most arid type of grassland. As the transition between grassland and desert, desert steppe constitutes the fragile zone of ecological environment, and it is also the early warning area of climate change and ecosystem evolution. Using unmanned aerial vehicle (UAV) hyperspectral remote sensing technology to extract grassland vegetation types more quickly and accurately is of great significance to the monitoring of grassland ecological security and the rational development of grassland animal husbandry. The HEX-6 eight rotor UAV was utilized, on which the Pika XC2 hyperspectral imager (spectral wavelength: 400~1000nm, spectral resolution: 1.3nm) was mounted to collect remote sensing images of desert steppe in Inner Mongolia, China. The hyperspectral images with a spatial resolution of 2.1cm were obtained by the UAV flying at a height of 30m from the ground. Spectral difference was enhanced by spectral continuum removal transformation and vegetation indices were constructed by the spectra after continuum removal transformation. The step by step band selection method was used to select vegetation feature bands for reducing data dimension. A random forest classification model with 24 variables, including spectral features, vegetation features, terrain features and texture features was constructed and compared with support vector machine (SVM), K-nearest neighbor (KNN) and maximum likelihood classification (MLC). The random forest classification algorithm (SBS_RF) proposed had the best classification effect among the four classification methods. The overall classification accuracy was 91.06%, which was 7.9, 15.61 and 18.33 percentage points higher than that of SVM, KNN and MLC, respectively. Kappa coefficient was 0.90, which was 0.13, 0.23 and 0.26 higher than that of SVM, KNN and MLC, respectively. The results showed that the combination of UAV hyperspectral remote sensing and SBS_RF algorithm provided a technical means for rapid investigation of desert grassland vegetation types and quantitative indicators for grassland ecological monitoring and animal husbandry management.