Data and Methods

Tree cover loss.  This data set measures areas of tree cover loss across all global land at 30-meter (m) resolution. The data were generated using multispectral satellite imagery from the Landsat 5 Thematic Mapper, the Landsat 7 Enhanced Thematic Mapper Plus (ETM+), and the Landsat 8 Operational Land Imager sensors. Over 1 million satellite images were processed and analyzed, including over 600,000 Landsat 7 images for the 2000–12 interval and more than 400,000 Landsat 5, 7, and 8 images for updates for the 2011–23 interval. The clear land surface observations in the satellite images were assembled and a supervised learning algorithm was applied to identify per-pixel tree cover loss.

In this data set, tree cover is defined as all vegetation greater than 5 m in height and greater than 30 percent tree canopy density, and it may take the form of natural forests or plantations across a range of canopy densities. Tree cover loss is defined as “stand replacement disturbance,” or the complete removal of tree cover canopy at the Landsat pixel scale (i.e., tree cover from more than 30 percent to about 0 percent). Tree cover loss may be the result of human activities, including forestry practices such as timber harvesting or deforestation (the conversion of natural forest to other land uses) as well as natural causes such as disease or storm damage. Fire is another widespread cause of tree cover loss and can be either natural or human induced. Forest degradation —for example, selective removals from within forests that do not lead to a nonforest state—are not included in the loss data. Therefore, partial reductions in canopy cover (e.g., from 70 percent to 40 percent) are not included in the loss data.

Data for 2011–23 were produced as annual updates, while 2001-2012 were produced as a block as part of the original publication. Recent years of data are also more sensitive to changes due to the incorporation of data from Landsat 8 (2013) and improvements to the method (most notably in 2015). Comparisons between older and more recent data should be performed with caution.

At the global scale, for the original 2001–12 product, the overall prevalence of false positives (detected as tree cover loss but, in reality, is not, also known as commission errors) in this data is 13 percent, and the prevalence of false negatives (not detected as tree cover loss but, in reality, is lost, also known as omission errors) is 12 percent, though the accuracy varies by biome and thus may be higher or lower in any particular location. The model often misses disturbances in smallholder landscapes, resulting in lower accuracy of the data in sub-Saharan Africa, where this type of disturbance is more common. There is 75 percent confidence that the loss occurred within the stated year, and 97 percent confidence that it occurred within a year before or after. The data also does not detect sparse and scattered trees in the agricultural landscape. Additional accuracy assessments for the updated algorithm and additional years of loss data beyond 2012 are not available.

Tree cover loss by dominant driver.  This data set shows the dominant driver of tree cover loss from 2001 to 2023 using the following five categories:

• Commodity-driven deforestation: Long-term, permanent conversion of forest and shrubland to a nonforest land use such as agriculture (including oil palm), mining, or energy infrastructure.
• Shifting agriculture: Small- to medium-scale forest and shrubland conversion for agriculture that is later abandoned and followed by subsequent forest regrowth.
• Forestry: Large-scale forestry operations occurring within managed forests and tree plantations.
• Wildfire: Large-scale forest loss resulting from the burning of forest vegetation with no visible human conversion or agricultural activity afterward.
• Urbanization: Forest and shrubland conversion for the expansion and intensification of existing urban centers. 

For the purposes of statistics generated in the Global Forest Review, commodity-driven deforestation, urbanization, and shifting agriculture with primary forest are considered to represent permanent deforestation, whereas tree cover usually regrows in the other categories (forestry, wildfire, and shifting agriculture outside of primary forests).

The data were generated using decision tree models to separate each 10-kilometer (km) grid cell into one of the five categories. The decision trees were created using 4,699 sample grid cells and use metrics derived from the following data sets: Hansen et al. (2013) tree cover, tree cover gain, and tree cover loss; National Aeronautics and Space Administration fires; global land cover; and population count. Separate decision trees were created for each driver and each region (North America, South America, Europe, Africa, Eurasia, Southeast Asia, Oceania), for a total of 35 decision trees. The final outputs were combined into a global map that is then overlaid with tree cover loss data to indicate the intensity of loss associated with each driver around the world.

Regional models were created, and training samples allowed for the interpretation of local land uses or management styles. A cell was categorized as commodity-driven deforestation if it contained clearings that showed signs of existing agriculture, pasture, or mining in the most recent imagery (after the tree cover loss occurred) as well as zero or minimal regrowth in subsequent years. Cells were categorized as shifting agriculture if the cell contained clearings that showed signs of existing agriculture or pasture in most recent imagery (after the tree cover loss occurred) as well as past clearings that contained visible forest or shrubland regrowth (gain) in historical imagery spanning 2001–15. Tree crops typically considered as agricultural commodities, such as oil palm, were classified accordingly as commodity-driven deforestation. The forestry class reflects a combination of wood fiber plantations and other forestry activity, including clear-cutting and selective cuts. Cells were categorized as wildfire when large swaths of fire scarring were visible in cleared areas, indicating that the loss event was driven by wildfire. The wildfire class excludes fire used to clear land for agriculture. Cells were categorized as urbanization if the loss of tree cover coincided with visible urban expansion or intensification.

This data set is intended for use at the global or regional scale, not for individual pixels. Individual grid cells may have more than one driver of tree cover loss, with variation over space and time.

Aside from commodity-driven deforestation, urbanization, and shifting agriculture with primary forest, which are assumed to represent permanent conversion from a forest to nonforest state, this data set does not indicate the stability or changing condition of the forest land use after the tree cover loss occurs. The data set does not distinguish between natural or anthropogenic wildfires, but it does distinguish fires for conversion or agricultural activity, which are not included in the wildfire class. Only direct drivers of forest disturbance are considered, and not indirect drivers such a demographic pressures or economic markets.

The accuracy of the data was assessed using a validation sample of 1,565 randomly selected grid cells. At the global scale, overall accuracy of the model was 89 percent, with individual class accuracies ranging from 55 percent (urbanization) to 94 percent (commodity-driven deforestation). The data has been updated since the original publication to include tree cover loss data from 2016 to 2023.

Tree cover gain.  This data set measures areas of tree cover gain across all land globally based on data at 30-meter resolution, displayed as the total area with tree cover in 2020 that did not have tree cover in 2000. The data was developed by Potapov et al. (2022) through the integration of the Global Ecosystem Dynamics Investigation (GEDI) lidar forest structure measurements and Landsat analysis-ready data time-series. The NASA GEDI is a spaceborne lidar instrument that provides point-based measurements of vegetation structure, including forest canopy height at latitudes between 52°N and 52°S globally. The Landsat multi-temporal metrics that represent the surface phenology serve as the independent variables for global forest height modeling with the GEDI data as the dependent reference data. The model was extrapolated to the boreal regions (beyond the GEDI data range).

Tree cover gain is defined as land cover with tree canopy height of at least five meters tall in 2020 but not in 2000 at the Landsat pixel scale. Tree cover gain may indicate a number of potential activities, including natural forest growth or the rotation cycle of tree plantations.

Lower Mekong height and canopy.  This data set measures the annual tree canopy extent and height for the lower Mekong region (Cambodia, Laos, Myanmar, Thailand, and Vietnam) at 30 m resolution for the years 2000–17. The data were generated from the University of Maryland’s Landsat Analysis Ready Data, a time-series data set of 16-day normalized surface reflectance composites, to produce regional woody vegetation structure mapping and change detection. A semiautomatic algorithm was used to map woody vegetation canopy cover and height. It used automatic data processing and mapping using a set of lidar-based vegetation structure prediction models. Any changes in vegetation cover were detected separately and then integrated into the structure time series.

Tree cover change.  This data set measures the net areas of tree cover change (loss or gain) across all land globally, and by country, between 2000 and 2020 based on data at 30-meter resolution. The data was developed by Potapov et al. (2022) through the integration of the Global Ecosystem Dynamics Investigation (GEDI) lidar forest structure measurements and Landsat analysis-ready data time-series. The NASA GEDI is a spaceborne lidar instrument that provides point-based measurements of vegetation structure, including forest canopy height at latitudes between 52°N and 52°S globally. The Landsat multi-temporal metrics that represent the surface phenology serve as the independent variables for global forest height modeling with the GEDI data as the dependent reference data. The model was extrapolated to the boreal regions (beyond the GEDI data range).

Net tree cover change is defined as the difference between tree cover gain and loss (that is, the amount of tree cover gain minus the amount of tree cover loss) between 2000 and 2020. Tree cover gain is defined as land cover with tree canopy height of at least five meters tall in 2020 but not in 2000 at the Landsat pixel scale (30 meters). Conversely, tree cover loss is defined as an area with tree canopy height greater than five meters in 2000 and less than five meters in 2020. Tree cover disturbance, in which tree cover is lost and regrown repeatedly during the 20-year time period, is tracked separately and not considered in the net tree cover change total.

Hot spots of primary forest loss.  The emerging hot spots data set identifies the most significant clusters of primary humid tropical forest loss between 2002 and 2023 within each country. The term hot spot is defined as an area that exhibits statistically significant clustering in the spatial patterns of loss. In this analysis, observed patterns of primary forest loss are likely to be attributable to underlying, as opposed to random, spatial processes. 

The emerging hot spots analysis uses the annual Hansen et al. (2013) tree cover loss data set between the years 2002 and 2023, the Turubanova et al. (2018) primary forest extent data set for the year 2001, and the Esri ArcGIS Emerging Hot Spot Analysis geoprocessing tool. The tool uses a combination of two statistical measures: the Getis-Ord Gi* statistic to identify the location and degree of spatial clustering of forest loss and the Mann-Kendall trend test to evaluate the temporal trend over time. The analysis was run for individual countries, and its results are relative to the patterns and amount of loss in each country. It has been updated since the original publication to include the latest tree cover loss data. 

Tree cover loss due to fire.  This data set measures areas of tree cover loss due to fires across all global land (except Antarctica and other Arctic islands) at approximately 30-meter resolution. Tree cover loss is defined, following Hansen et al. 2013, as “stand replacement disturbance,” or the complete removal of tree cover canopy at the Landsat pixel scale. Stand replacement forest fires are defined as natural or human-ignited fires resulting in direct loss of tree canopy cover exceeding 5 meters in height. This can include wildfires, intentionally set fires, or escaped fires from human activities, such as hunting or agriculture. It does not include burning of felled trees, since the direct cause of loss in these cases is mechanical removal. Therefore, trees that are cut down and later burned to clear land for agriculture would not be classified as tree cover loss due to fire in this dataset. It does not include low-intensity and understory forest fires that do not result in substantial tree canopy loss at the scale of a 30-meter pixel.

The data were generated using global Landsat-based annual change detection metrics as input data to a set of regionally calibrated classification tree ensemble models. Tree cover loss due to fire was mapped only within the extent of the global 30-m resolution tree cover loss data set (Hansen et al., 2013). The result of the mapping process can be viewed as a set of binary maps (tree cover loss due to fire vs. tree cover loss due to all other drivers).

Global cocoa, coffee, soy.  For the Global Forest Review (GFR), we use cocoa, arabica and robusta coffee, and soy maps from MapSPAM to assess which crops have replaced forests; the exception is for soy in South America, where higher-resolution and more recent data are available. The MapSPAM data maps crop area for 42 crops in the year 2010 at a spatial resolution of 10 kilometers (km). Physical crop area was used in all analyses, as opposed to harvested area, to account for all land occupied by a specific crop. The data combines country and subnational reported production statistics, an agriculture land cover map, and crop-specific suitability information based on climate, landscape, and soil conditions into a spatial model. Suitable areas for each crop are identified in the MapSPAM data by using existing land resources and biophysical limitations to provide suitable crops areas.  Each 10 km grid cell contains the estimated area of each of the 42 crops, further broken into physical and harvested area of irrigated high input, rain-fed high input, rain-fed low input, and rain-fed subsistence. Whereas high input includes the use of high-yield crop varieties, optimal application of fertilizer, chemical pest disease and weed controls, and might be fully mechanized, low input uses traditional varieties of crops with manual labor and minimal or no applications of fertilizers or pest control measures. Subsistence refers to crop production by small-scale farmers largely for their own consumption under rain-fed and low-input conditions, regardless of the suitability of land. It is assumed to happen more intensively in areas with large rural populations, so rural population density from the Global Rural-Urban Mapping Project (Version 1) helps to further identify subsistence farming.    

Global pasture.  This data set maps global pastureland at a 10 km resolution for the year 2000. In the GFR, we use EarthStat pasture data to assess where pasture has replaced forests; the exception is for Brazil, where there is higher-resolution and more recent data available. EarthStat uses the definition of permanent pasture used by the Food and Agriculture Organization of the United Nations (FAO): “land used permanently (5 years or more) for herbaceous forage crops, either cultivated or growing wild.” Agricultural inventory data from a variety of sources, including country and FAOSTAT data, were modeled onto land-use and land cover maps of agriculture and pasture derived from Moderate Resolution Imaging Spectroradiometer and SPOT imagery. The definition of pasture causes some known inconsistencies because some countries distinguish between grassland pasture and grazed land, but most do not in their reporting. 

Brazil pasture.  This data set maps annual pasture extent in Brazil at a 30 m resolution. In the GFR, this data is used preferentially over the global pasture map to assess where pasture replaced forests in Brazil. The data were derived from Landsat imagery using automatic, random forest classification. Historical maps from 1985 to 2018 are available from the Image Processing and Geoprocessing Laboratory (Laboratório de Processamento de Imagens e Geoprocessamento; LAPIG) as part of the MapBiomas initiative, but only the 2018 extent was used in the GFR calculations.  

South America soy.  This data set maps annual soy extent from 2001 to 2018. We use this data set for all calculations to assess where soy replaced forests in South America. The data were derived from Landsat imagery to map the harvest season of soy annually from 2001 to 2018. All years were combined to estimate forest loss on land that was eventually used for soy production. 

Oil palm, rubber, wood fiber.  Oil palm, rubber, and wood fiber plantations from the Spatial Database of Planted Trees (SDPT) were used for all calculations to assess where oil palm, rubber, and wood fiber replaced forests. Oil palm is thought to be a comprehensive data set for the year 2015, whereas rubber and wood fiber plantation data is only available for specific countries (for rubber, Brazil, Cambodia, Cameroon, the Democratic Republic of the Congo, India, Indonesia, and Malaysia; and for wood fiber, Argentina, Brazil, Cambodia, China, India, Indonesia, Malaysia, Rwanda, South Africa, and Vietnam). See above for more information about the SDPT. 

Biodiversity intactness.  This data set quantifies the impact humans have had on the intactness of species communities. Anthropogenic pressures such as land-use conversion have caused dramatic changes to the composition of species communities, and this layer illustrates these changes by focusing on the impact of forest change on biodiversity intactness. The maximum value indicates no human impact, whereas lower values indicate that intactness has been reduced. 

The Projecting Responses of Ecological Diversity in Changing Terrestrial Systems (PREDICTS) database comprises over 3 million records of geographically and taxonomically representative data of land-use impacts to local biodiversity.  A subset of the PREDICTS database, including data pertaining to forested biomes only, was employed to model the impacts of land-use change and human population density on the intactness of local species communities. 

First, a relevant land-use map was produced by selecting all forested biomes and each 30-by-30-meter (m) pixel within the biomes was assigned a land-use category based upon inputs from the Global Forest Watch forest change database and a downscaled land-use map.  The modeled results of biodiversity intactness derived from the PREDICTS database are projected onto the land-use and human population density maps, and the final product is aggregated to match the resolution of the downscaled land-use map.  The final output models the impacts of forest change on local biodiversity intactness within forested biomes.

The metric assumes that the biodiversity found in a perfectly intact site is equivalent to the biodiversity that would be present without human interference. Human impacts on biodiversity intactness are quantified through models that extrapolate results from site-specific studies across large areas, and there is always a degree of uncertainty in such extrapolations.

Biodiversity significance.  This data set shows the significance of each forest location for biodiversity in terms of the relative contribution of each pixel to the global distributions of all forest-dependent mammals, birds, amphibians, and conifers worldwide. To calculate it, species that are coded in the International Union for Conservation of Nature (IUCN) Red List of Threatened Species as forest dependent are selected and their distribution maps are clipped by their known altitudinal ranges (note, the altitudinal range for amphibians has not been assessed) using a digital elevation model data set, and overlapped with the layer of forest cover. For each species, the relative “significance” of each forest pixel in their range is calculated as one divided by the total number of pixels of forest in their range. These values are summed for all species occurring within the pixel to give an overall value to the pixel. This metric is also sometimes termed range rarity.

This data set includes several caveats. There are many ways to define biodiversity significance, and this layer is based on one approach. Only forest-dependent bird, mammal, amphibian, and conifer species were included in the analysis. The individual species range maps upon which this layer is based show distributional boundaries, not occupancy, and so contain commission errors. However, when more than 15,000 species ranges are combined into this single layer, such errors become largely irrelevant. Historical ranges were excluded. Hence, the value of each pixel is related to the global loss of species richness if the pixel is deforested. Locations of high species richness do not necessarily have high scores if most of the species in the location have large global distributions. All species are treated equally, so the evolutionary distinctiveness of different taxa is not considered. When overlaid with maps of forest loss, forest gain is ignored. It is assumed that tree cover gain over the analysis period is unlikely to translate into significant gain in forest-dependent species given the natural time lags in regeneration of forest ecosystems. Finally, the data set provides a broad picture of variation in biodiversity significance of different forests globally. It is not intended to be used in isolation for priority setting or decision-making, for which additional information is typically needed.

Key Biodiversity Areas.  Key Biodiversity Areas (KBAs) are “sites contributing significantly to the global persistence of biodiversity.” The Global Standard for the Identification of Key Biodiversity Areas  sets out globally agreed-upon criteria for the identification of KBAs worldwide. Sites qualify as global KBAs if they meet one or more of 11 criteria, clustered into five categories: threatened biodiversity, geographically restricted biodiversity, ecological integrity, biological processes, and irreplaceability. The KBA criteria can be applied to species and ecosystems in terrestrial, inland water, and marine environments. Although not all KBA criteria may be relevant to all elements of biodiversity, the thresholds associated with each of the criteria may be applied across all taxonomic groups (other than microorganisms) and ecosystems.

The KBA identification process is a highly inclusive, consultative, and bottom-up exercise. Although anyone with appropriate scientific data may propose a site to qualify as a KBA, consultation with stakeholders at the national level (both nongovernmental and governmental organizations) is required during the proposal process.

Over 15,000 KBAs have been identified to date, including Important Bird and Biodiversity Areas, Alliance for Zero Extinction sites, and KBAs identified through hot spot ecosystem profiles supported by the Critical Ecosystem Partnership Fund.

Alliance for Zero Extinction.  A subset of KBAs, this data set shows 587 sites for 920 species of mammals, birds, amphibians, reptiles, conifers, and reef-building corals. The species found within these sites have extremely small global ranges and populations; any change to habitat within a site may lead to the extinction of a species in the wild. To meet Alliance for Zero Extinction site status, a site must

• contain at least one “Endangered” or “Critically Endangered” species;
• be the sole area where an Endangered or Critically Endangered species occurs;
• contain greater than 95 percent of either the known resident population of the species or 95 percent of the known population of one life history segment (e.g., breeding or wintering) of the species; and
• have a definable boundary (e.g., species range, extent of contiguous habitat, etc.).

IUCN Red List of Threatened Species. This data set contains distribution information on species assessed for the IUCN Red List of Threatened Species. The maps are developed as part of a comprehensive assessment of global biodiversity to highlight taxa threatened with extinction and thereby promote their conservation. The IUCN Red List contains global assessments for 105,732 species, with more than 75 percent of these having spatial data. The Global Forest Review (GFR) uses the Asian elephant, Bornean orangutan, Sumatran orangutan and tiger ranges to assess tree cover loss in their habitat ranges, which were mapped in 2020, 2016, 2017 and 2022 respectively. These three species represent endangered, iconic animals of Southeast Asia. Future editions of the GFR will likely include additional iconic species from South America and Africa. 

Erosion risk.Qin et al. 2016, https://www.wri.org/publication/gfw-water-metadata . This data set maps the risk of erosion around the world. Erosion and sedimentation by water involves the process of detachment, transport, and deposition of soil particles, driven by forces from raindrops and water flowing over the land surface. The Revised Universal Soil Loss Equation (RUSLE), which predicts annual soil loss from rainfall and runoff, is the most common model used at large spatial extents due to its relatively simple structure and empirical basis. The model takes into account rainfall erosivity, topography, soil erodibility, land cover and management, and conservation practices. Because the RUSLE model was developed based on agricultural plot scale and parameterized for environmental conditions in the United States, modifications of the methods and data inputs were necessary to make the equation applicable to the globe. Conservation practices and topography information were not included in this model to calculate global erosion potential due to data limitations and their relatively minor contribution to the variation in soil erosion at the continental to global scale compared to other factors. The result of the global model was categorized into five quantiles, corresponding to low to high erosion risks. 

Urban watersheds.McDonald and Shemie 2014, http://water.nature.org/waterblueprint/#/intro=true . Urban watershed boundaries for 530 cities, mapped as part of the Urban Water Blueprint project, including 33 megacities with more than 10 million people. According to United Nations population data for 2018, there were 33 cities with a population greater than 10 million people in 2018. Due to the availability of watershed boundary data, 32 of these cities are included in the Global Forest Review

Ecozones.FAO 2012, http://www.fao.org/3/ap861e/ap861e00.pdf . This data set depicts major ecozones, including boreal, temperate, tropical, and subtropical regions. 

Peatlands.For more information about peatlands, see Global Forest Watch, https://gfw.global/37Pfnpw . This data set shows peatlands in Indonesia greater than five meters in depth. 

Indonesian forest moratorium.For more information about the moratorium, see Global Forest Watch, https://gfw.global/3oxnjBJ . Data set indicating the area of Indonesia’s moratorium against new forest concessions, designed to protect Indonesia’s peatlands and primary natural forests from future development. In May 2011, the Ministry of Environment and Forestry put into effect a two-year moratorium on the designation of new forest concessions in primary natural forests and peatlands. This moratorium is designed to allow time for the government to develop improved processes for land-use planning, strengthen information systems, and build institutions to achieve Indonesia’s low-emission development goals. The moratorium, made permanent in 2019, is part of Indonesia’s pledge to curtail forest clearing in a US$1 billion deal with the Norwegian government. 

Rural complex.Molinario et al. 2015, https://doi.org/10.1088/1748-9326/10/9/094009 . This Democratic Republic of the Congo (DRC) land-use and land cover data set depicts core forest, forest fragmentation, and the rural complex, a land-use mosaic of roads, villages, active and fallow fields, and secondary forest that we use as a proxy for shifting cultivation in the DRC. This is separate from shifting cultivation identified in the drivers of deforestation data set and is only used in the DRC-specific analysis from the Forest Extent Indicator.  

The data set was created by characterizing forest clearing using spatial models in a geographical information system, applying morphological image processing to the Central African Forests Remotely Assessed (Forets d'Afrique Central Evaluee par Teledetection; FACET) product. This process allowed for the creation of maps for 2000, 2005, 2010, and 2015, classifying the rural complex and previously homogenous primary forest into separate patch, edge, perforated, fragmented, and core forest subtypes. 

Countries.See GADM, https://gadm.org/ .  This data set shows political boundaries, including country, provincial, and jurisdictional administrative units. 

ESA CCI Land Cover.See ESA CCI, http://maps.elie.ucl.ac.be/CCI/viewer/download/ESACCI-LC-Ph2-PUGv2_2.0.pdf This data set maps annual global land cover at 300-meter resolution for the year 2020. The annual maps are derived from a baseline land cover map based on the Medium Resolution Imaging Spectroradiometer (MERIS) Full Resolution and Reduced Resolution archive from 2003 to 2012. Land cover changes are detected based on the Very High Resolution Radiometer (VHRR) time series from 1992 to 1999, SPOT-Vegetation time series between 1999 and 2013 and PROBA-Vegetation for years 2013, 2014 and 2015. The baseline land cover map is then backdated and updated to produce annual maps. For the Global Forest Review, it is used to delineate urban, grassland and agricultural land, on which trees outside forests are quantified.

SBTN Natural Lands.Mazur et al. 2023, https://sciencebasedtargetsnetwork.org/wp-content/uploads/2023/05/Technical-Guidance-2023-Step3-Land-v0.3-Natural-Lands-Map.pdf This data maps natural and non-natural lands for the year 2020 at 30-meter resolution based on Accountability Framework Initiative (AFi) definitions and operational guidance. This data was created by WRI in collaboration with WWF and Systemiq as part of the Science Based Target Network‘s guidance for setting land science-based targets. The Natural Lands map was created by combining the best available global spatial data on land cover and land use into a single, harmonized map using a series of overlays and decision rules. Both global and regional/local data was used, and where available, regional/local data was given priority. The global binary map was independently validated by the International Institute for Applied Systems Analysis (IIASA) using a random sample of 4,730 points and has an overall accuracy of 91.6 percent. Individual class accuracies show that the map misclassifies 6 percent of the natural points as non-natural, and 18 percent of the non-natural points as natural. Limitations include definitional, temporal and resolution inconsistencies due to the combination of data from various sources. Due to a lack of global remote sensing data on pasturelands, gridded livestock densities from the UN Food and Agriculture Organization (FAO) were used as a proxy for non-natural short vegetation.