An open global reference dataset for natural and improved grasslands and its application for validation and area estimation

Grasslands support livestock production, biodiversity, carbon storage and numerous other ecosystem services, yet their global monitoring is constrained by a lack of harmonized reference data that distinguish natural and semi-natural grasslands from cultivated systems. Here we present an openly available global reference dataset for grassland characterization, area estimation and map validation (Lesiv et al., 2026, doi: 10.5281/zenodo.23368916). The dataset comprises 50,045 probability-sampled locations distributed across the global land surface and interpreted for 2020 by trained experts using the Geo-Wiki platform. Interpretation used very-high-resolution satellite imagery, Sentinel-2 imagery and Landsat and MODIS vegetation time series along with historical imagery and geotagged photographs where available. Each location was interpreted at least once, while locations with a higher likelihood of grassland occurrence received two independent annotations. The probability-sampling design supports sample-based estimation of classification accuracy and grassland area within design-based inference. Validation of the 2020 Global Pasture Watch version 1 (GPW v1) product yielded overall accuracies of 73.3 % and 66.8 % (standard errors of 0.2 %) for two definitions of cultivated grasslands (“narrow”, which excluded grazed semi-natural grassland and open woodland and “broad”, which included these classes). Natural/semi-natural grasslands were mapped more accurately than cultivated grasslands under both definitions. Cultivated grasslands showed substantial commission errors under the narrow definition and substantial omission errors under the broad definition, with classification difficulties particularly evident in Africa and Asia. Estimated total grassland area was approximately 3.30 billion ha. Under the narrow definition, the distribution to natural/semi-natural and cultivated grasslands was estimated to be 2.97 and 0.33 billion ha, respectively, compared with 1.89 and 1.41 billion ha under the broad definition. Estimated total grassland area was approximately 3.30 billion ha. The dataset provides a transparent and reusable benchmark for grassland map validation, sample-based area estimation, machine-learning development and assessments of grazing systems and grassland-related environmental change.