Abstract. Grassland aboveground biomass (AGB) is a critical component of the global carbon cycle and reflects ecosystem productivity. Although it is widely acknowledged that dynamics of grassland biomass are significantly regulated by climate change, in situ evidence at large spatiotemporal scales is limited. Here, we combine biomass measurements from six long-term (> 30 years) experiments and data in existing literatures to explore the spatiotemporal changes in AGB in Inner Mongolian temperate grasslands. We show that, on average, annual AGB over the past four decades is 2,561 ka ha−1, 1,496 kg ha−1 and 835 kg ha−1, respectively, in meadow steppe, typical steppe and desert steppe in Inner Mongolia. The spatiotemporal changes of AGB are regulated by interactions of climatic attributes, edaphic properties, grazing intensity and grassland type. Using a machine learning-based approach, we map annual AGB (from 1981 to 2100) across the Inner Mongolian grassland at the spatial resolution of 1 km. We find that on the regional scale, meadow steppe has the highest annual AGB, followed by typical and desert steppe. During 1981–2019, the average annual AGB generally exhibited a declining trend across all the three types of grassland. Under future climate warming, AGB in the study region could continue to decrease. On average, compared with the historical AGB (i.e., average of 1981–2019), the AGB at the end of this century (i.e., average of 2080–2100) would decrease by 14 % under RCP4.5 and 28 % under RCP8.5, respectively. The decreases in AGB under warming show large disparities across different grassland types and future climate change scenarios. Our results demonstrate the accuracy of predictions in AGB using a machine learning-based approach driven by several readily obtainable environmental variables; and provide new data at large scale and fine resolution extrapolated from field measurements.