Accurately delineating the spatial distribution of grassland classes across large-scale regions was of critical importance for the sustainable management of grassland resources and the effective protection of grassland ecosystems. The widely used grassland resource maps in China were more than 40 years old, and there was an urgent need to draw new grassland classes distribution maps to reflect the actual distribution of grassland resources more accurately. Nevertheless, in vast and ecologically diverse regions, grassland classes exhibited significant spatial heterogeneity. This complexity, coupled with the limitations of conventional classification techniques that rely primarily on spectral features, posed substantial challenges for precise large-scale mapping, especially in areas characterized by intricate grassland classes distribution. To overcome these limitations, we developed an advanced grassland mapping framework based on the Google Earth Engine (GEE) cloud computing platform. We constructed a comprehensive multi-source feature dataset by integrating Landsat surface reflectance data with a variety of auxiliary features, including spectral indices, texture characteristics, phenological metrics derived from time-series data, and habitat-related environmental variables such as climate, topography, and soil attributes. Using this enriched dataset and following China's grassland zoning scheme, we applied a Random Forest (RF) classifier to generate detailed maps of grassland classes distribution across different grassland zones within Xinjiang. The results of the study revealed several key findings: (1) Incorporating multi-source features markedly enhanced classification performance. Specifically, the inclusion of phenological, topographic, soil, and climatic variables led to increases in overall accuracy (OA) of 25.64%, 27.40%, 23.57%, and 28.73%, respectively, compared to models using spectral features alone. (2) The partitioned modeling strategy with multi-source features outperformed single full-region modeling. This method achieved an OA of 80.86% and a Kappa coefficient of 0.76, representing improvements of 3.51% and 0.05, respectively. (3) Spatially, grassland classes in Xinjiang exhibited a distinct vertical zonality, primarily distributed along mountain systems. With increasing elevation, grasslands transitioned from temperate desert in low-lying regions to alpine steppe in higher altitudes. (4) Temporally, the extent of grassland in Xinjiang has undergone dynamic change of decreasing and then increasing over the past four decades. During the 1980-1990 period, extensive grassland degradation occurred, with many areas converted to barren land. In contrast, from 1990 to 2020, a notable recovery trend was observed, driven largely by vegetation restoration in mountainous regions. Nevertheless, in basin areas, grasslands continue to be threatened by human-driven land use change, particularly conversion to cropland and expansion of bare land. This study elucidated the fine-scale spatial distribution and spatiotemporal evolution of Xinjiang's grassland classes since 1980, providing a scientific foundation for sustainable grassland resource management and ecological conservation.