Quantitative estimation of the spatial and temporal dynamics of the fractional cover of photosynthetic vegetation (fPV) and non-photosynthetic vegetation (fNPV) in semi-arid grasslands is critical for understanding grassland conditions such as vegetation abundance, drought severity, fire fuel load, stocking rate, and disturbance events and recovery. It is also important for scientific grassland resource management. Over the past several decades, remote sensing has become an important tool for estimating the fractional cover of vegetation, which is a key descriptor of ecosystem function. However, most efforts have been devoted to the estimation of fPV rather than fNPV, although the latter is equally important, especially in arid and semi-arid ecosystems. This study describes a linear unmixing approach for estimating fPV and fNPV in the Xilingol steppe region with hyperspectral and field investigation data. Five Hyperion images acquired on April 4, May 20, July 27, August 30, and November 15 in 2012 and a field-measured spectral library were utilized to explore the spectral feature space of fPV and fNPV in order to validate the feasibility of a linear unmixing model. This model is based on two complementary spectral indices of vegetation that have been used in remote sensing analyses to discriminate green and dry vegetation from soils: the Normalized Difference Vegetation Index (NDVI) and the Cellulose Absorption Index (CAI). Different end-member extraction methods, including the Minimum-Volume Enclosing (MVE) method, the Pixel Purity Index (PPI) method, and a field measurement method, were adopted to retrieve the end-member values of photosynthetic vegetation, non-photosynthetic vegetation, and bare soil, respectively, from NDVI and CAI. Then, the influence of end-member extraction on the accuracy of the fPV and fNPV estimation was evaluated through comparison with field-measured fPV and fNPV values acquired from classifications performed on fisheye photos (N=52). Subsequently, the optimum unmixing strategy was utilized to retrieve the temporal dynamics of fPV and fNPV in a fenced area, where the grassland was not influenced by human activities, so that the usefulness of these fractional coverage indices could be validated by checking their consistency with the phenology of natural grassland. The result shows that the linear unmixing model based on NDVI and CAI was effective for estimating fPV and fNPV in the Xilingol steppe region. The NDVI-CAI feature space follows a triangular distribution, where the three vertexes represent photosynthetic vegetation, non-photosynthetic vegetation, and bare soil, meeting the essential requirements of the linear unmixing model. The estimation accuracy was different for the different end-member extraction methods. The MVE-based estimation had the highest accuracy, with estimated accuracy of 91.2% and 67.91% for fPV and fNPV, respectively, followed by the PPI-based estimation (91.0% and 59.5% for fPV and fNPV, respectively) and the field-measurement-based estimation (86.2% and 56.7% for fPV and fNPV, respectively). In general, the estimation accuracy was higher for fPVthan fNPV, and the field-measured end-member performed worse than the image end-member, which was probably due to the inconsistency between the field-measured spec and the Hyperion spec. Additionally, the temporal dynamics of fPV and fNPV were confirmed to be consistent with the phenological seasonal change in natural grasslands. Therefore, the method proposed here can be used to monitor the temporal and spatial variations of fPV and fNPV in semi-arid grasslands.