Grassland is an important natural resource; however, in recent years, due to the effects of climate warming and human disturbances, grassland degradation has become more severe. Consequently, the role of hyperspectral remote sensing in extracting ground object information and monitoring degradation has become increasingly significant. Taking Xinghai County as an example, this paper utilizes hyperspectral images from the China-Brazil Earth Resources Satellite (CBERS) 01, in conjunction with community survey data from sample plots. It extracts pixel-scale grassland information via spectral feature analysis and mixed spectral decomposition, performs grassland vegetation classification mapping, and classifies the level of grassland degradation based on the proportion of high-quality forage and vegetation. The main conclusions are as follows: (1) In analyzing the spectral characteristics of typical grassland vegetation, low-coverage grasslands can be identified through the reflectance of chlorophyll-sensitive bands (559 nm and 670 nm) in the visible light spectrum. In the near-infrared spectrum, the spectral curve exhibits significant fluctuations. Reflectance characteristics at 1122 nm and 1660 nm can be used to differentiate shady slope vegetation, linear-leaved Kobresia grassland communities, and crops from other types of vegetation. The first-order derivative transformation and envelope removal transformation enhance the spectral characteristic differences of various vegetation types in the “red edge” and “red valley” regions, respectively; (2) By employing the Mahalanobis distance method and correlation analysis, the reflectance spectra for 12 dominant grass species, including Alpine Songgrass, Dwarf Songgrass, and Wolf Poison Grass, were optimized and de-redundant, identifying 8 sensitive bands (507 nm, 670 nm, 713 nm, 765 nm, 1257 nm, 1324 nm, 1459 nm, 1929 nm) extracted. (3) Comparing the three hyperspectral remote sensing classification methods (spectral angle mapping, spectral information divergence and decision tree), it was found that the decision tree method based on spectral feature analysis had the highest overall classification accuracy (77.56%), and achieved ideal classification results at the grassland community level; (4) Multiple endmember spectral mixture analysis (MESMA) was used to study the mixed pixel decomposition of grassland communities, and the results showed that MESMA had good accuracy for spectral mixture decomposition between vegetation and bare soil (RMSE of 0.16), but poor performance for mixed decomposition between fine forage, weed grass, and bare soil (RMSE of 0.28); (5) Two evaluation indicators of grassland degradation, namely “vegetation ratio” and “excellent forage ratio,” were proposed, and the grassland degradation was evaluated hierarchically, and the results showed that the proportion of moderate grassland degradation in the study area reached 56.25%.