The accurate assessment of grassland above-ground biomass (AGB) is crucial for the sustainable utilization and protection of grassland resources. In this study, the Gannan prefecture was selected as a representative case study area. The AGB estimation accuracies of 9 vegetation indices were evaluated and the final estimation model were established based on the Sentinel-2 surface reflectance products and field measurements of AGB from 2019 to 2021. The Google Earth Engine (GEE) platform and mathematical statistics method were employed for analysis. Then, the spatio-temporal dynamics of grassland AGB were analyzed in Gannan Prefecture from 2019 to 2021. The results indicated that the R² value between the normalized difference phenology index (NDPI) and AGB was the highest (0.72), followed by normalized difference vegetation index (NDVI) (0.68), and the R² values of the enhanced vegetation index (EVI) (0.37) and different vegetation index (DVI) (0.40) were the lowest, among all vegetation indices. The NDPI was more sensitive to grassland AGB than other vegetation indices. Among the four NDPI based estimation models, the power model showed the best performance in grassland AGB prediction, and the prediction accuracy of AGB in alpine grassland was obviously improved. From 2019 to 2021, the spatial distributions of AGB in alpine grassland were significantly different, and with higher values in the southwestern and lower values in the eastern and northern. The average AGB of mountain meadow was slightly larger than that of alpine meadow and swamp wetland. The AGB of alpine meadow and mountain meadow increased and then decreased, while the swamp wetland continued a decreasing trend over the past three years. Our findings suggest that NDPI can reflect the spatio-temporal dynamics of alpine grassland AGB on a large scale.