Machine learning models used in the production of the global maps of annual grassland class and extent for 2000—2022 within the scope of the Global Pasture Wath initiative. The models were trained in scikit-learn using 2.3M of samples, spacetime overlaid with 103 features (GLAD Landsat ARD-2 bi-montly composites, climatic, landform and proximity covariates -- full list in gpw_grassland_rf_input.raster.layers_v1.csv): Binary Random Forest classifier of cultivated grassland vs other land cover Binary Random Forest classifier of natural/semi-natural grassland vs other land cover For each model, Recursive Feature Elimination, Successive Halving hyperparameter tuning and five-fold spatial blocking cross-validation were conducted. The fitted models were compiled to a native C binary using TL2cgen, reducing the prediction time by factor 3. The predictions were computed using Scikit-Map. and computational notebook describing explaining all modeling steps is available in Github. Related resources Maps of dominant grassland:2000-2002 2003-2005 2006-2008 2009-2011 2012-2014 2015-2017 2018-2020 2021-2022 Probability maps of cultivated grassland:2000-2022 (All URLs) Probability maps of natural/semi-natural grassland:2000-2022 (All URLs) Grassland reference samples based on VHR imagery (2000–2022):GeoPackage files Global machine learning models (Random Forest):Parquet and joblib python files Reference sampling design derived by FSCV:GeoPackage and raster files Harmonized reference samples based on existing LULC dataset:GeoPackage and raster files Source code for reproducibility:GitHub release Mapping feedback tool:GeoWiki Data catalogues:OpenLandMap STAC Google Earth Engine Support For questions of bugs/inconsistencies related to the dataset raise a GitHub issue in https://github.com/wri/global-pasture-watch