Abstract A long-term dataset of aboveground net primary productivity (ANPP) for global natural grasslands is essential for carbon dynamics modeling and sustainable land management. However, existing datasets are limited: they often fail to separate above- and below-ground productivity or reflect only post-disturbance conditions. To address these gaps, we developed a gridded annual ANPP dataset using machine learning, spanning historical (1958–2023) and future (2015–2100) periods. Historical ANPP data were derived from TerraClimate at 1/24° spatial resolution, while future projections came from CMIP6 models under SSP245 and SSP585 scenarios at 1/2° resolution. Our model performed robustly (R 2 = 0.675 ± 0.009), showing temporal and spatial reliability through cross-validation with published products. Notably, systematic ANPP underestimation occurs in high-productivity regions (>700 g m −2 ) due to sparse field observations, so values in these areas should be interpreted with caution. Our dataset provides a spatially explicit baseline of climate-driven productivity, supporting precise evaluation of human impacts on grasslands and informing adaptive management under climate change.
Chen et al. (Fri,) studied this question.