Historical grassland cover change is vital for global and regional environmental change modeling; however, in China, estimates of this are rare, and therefore, we propose a method to reconstruct grassland cover over the past 300 years. By synthesizing remote sensing-derived Chinese land use and land cover change (LULCC) data (1980–2015) and potential natural vegetation data simulated by the relationship between vegetation and environment, we first determined the potential extent of natural grassland vegetation (PENG) in the absence of human activities. Then we reconstructed grassland cover across western China between 1661 and 1996 at 10 km resolution by overlaying the Chinese historical cropland dataset (CHCD) over the PENG. As this land cover type has been significantly influenced by anthropogenic factors, the data show that the proportion of grassland in western China continuously decreased from 304.84 × 106 ha in 1661 to 277.69 × 106 ha in 1996. This reduction can be divided into four phases, comprising a rapid decrease between 1661 and 1724, a slow decrease between 1724 and 1873, a sharp decrease between 1873 and 1980, and a gradual increase since 1980. These reductions correspond to annual loss rates of 7.32 × 104 ha, 2.90 × 104 ha, 17.04 × 104 ha, and −2.37 × 104 ha, respectively. The data reconstructed here show that the decrease in grassland area between 1661 and 1724 was mainly limited to the Gan-Ning region (Gansu and Ningxia) and was driven by the early agricultural development policies of the Qing Dynasty. Grassland was extensively cultivated in northeastern China (Heilongjiang, Jilin, and Liaoning) and in the Xinjiang region between 1724 and 1980, a process which resulted from an exponential increase in immigrants to these provinces. The reconstruction results enable provide crucial data that can be used for modeling long-term climate change and carbon emissions.