Historical maps of land use were based on the History of the Global Environment database (HYDE). HYDE provides long-term historical, spatially explicit time series on a 5 arcmin resolution of population estimates as well as land use reconstructions covering the Holocene period, defined here as 10 000 BCE until the present (Table 1). It is an effort to quantify the agricultural expansion of humankind over time. In principle, HYDE uses a simple approach of combining historical population estimates with assumptions on the trajectory of historical land use per capita. Allocation of land use patterns is steered at the present day by satellite information and UN FAO agricultural land use data (FAO, 2020a), and this is gradually replaced towards the past by a combination of spatially explicit maps such as climate, soil, slope, and neighborhood of rivers and lakes. The latest version (3.2; Klein Goldewijk et al., 2017) presents land use categories such as built-up area, managed pastures and more extensive rangelands, cropland excluding rice, and rice as a separate crop because of its relevancy for greenhouse gas emissions. A distinction was made between irrigated and rain-fed cropland (both for other crops and rice). Besides the baseline reconstruction, two alternative historical land use reconstructions were provided based on uncertainties. For a full description of the methodology, see Klein Goldewijk et al. (2017).

The version of the HYDE 3.2 dataset used for the baseline LUH2 historical product was the 2016_beta_release version, and the version used for the high and low scenarios was the 2017_beta_release_000 version. Data were provided at 5′ spatial resolution every 100 years from 800 to 1700, every 10 years from 1700 to 2000, and then annually from 2000 to 2015. These data were aggregated to 0.25∘×0.25∘ resolution and converted from the absolute area of each grid cell to grid cell fractional area. Data were then linearly interpolated in time to produce annual maps of the fraction of each 0.25∘ grid cell occupied by each of the following land use types: cropland, managed pasture, rangelands, and urban. The ice and water fractions of each grid cell were also taken from the HYDE dataset and were assumed constant over time. By subtracting the land use, ice, and water fractions from each grid cell, the fractions of each grid cell occupied by natural vegetation (either primary or secondary forest or non-forest) were also determined. The HYDE 3.2 dataset also includes a global map that assigns a country code to each terrestrial grid cell at 5′ resolution. This map served as a basis to generate a similar map at 0.25∘ resolution, consistent with the 0.25∘ maps of land use data. In this map every grid cell with an ice / water fraction less than 1.0 was assigned a country code, resulting in a global map containing 199 countries.

The cropland fraction of each grid cell, along with transitions to and from cropland, is further subdivided into five different crop functional types (CFTs): C3 annuals, C4 annuals, C3 perennials, C4 perennials, and C3 nitrogen fixers. For the years 850 to 2015 the CFT fractions of total cropland are primarily based on data from Monfreda et al. (2008), which provide global maps of harvested areas of 175 different crops at 5 min spatial resolution for the year 2000. For use in the LUH2 methodology, these maps were aggregated into five CFT classes at 0.25∘ spatial resolution and then normalized so that all CFT fractions sum to 1 in each grid cell. For grid cells that do not have crop-type data from Monfreda et al. (2008), national crop-type data from the FAO (FAO, 2020a) are used instead (i.e., by aggregating the 169 FAO crop types into the five CFT classes represented in LUH, averaging over all years of FAO data from 1961 to 2013, then assigning the normalized national CFT fractions to any grid cells within each country that did not have Monfreda data). The resulting map of CFT fractions is used for all years 850–2015 to subdivide the gridded cropland fraction and cropland-related transitions into CFT fractions and CFT-related transitions by multiplying the cropland fraction of each grid cell (and the cropland-related transitions to and from each grid cell) by the CFT fraction map. Note that this process includes the inherent assumption that the fraction of a grid cell that was harvested for a crop type (i.e., the Monfreda et al. data, 2008) was roughly correlated with the fraction of the total cropland area that was occupied by that crop type.

For the years 2015–2100, we first identify one or two CFTs in the IAM data that have the greatest global area increase over the 85-year period. We then attempt to follow the gridded changes in the fraction of cropland occupied by those CFTs by first assigning as much of the cropland expansion transitions as possible to the expansion of those one or two CFTs and then, when needed, adding transitions between CFTs to reassign area from CFTs with lower rates of increase (or even reductions) of area in the IAM data to the CFTs with large global increases in area. The result of this process is typically that the global area changes of CFTs in LUH2 tend to follow global area changes of CFTs in the IAM data, not just for the CFTs with the largest area changes, but for others as well. When there were no CFTs with significant changes over the 2015–2100 period, the contemporary CFT ratios were used to disaggregate total cropland area into CFT fractions for all years 2015–2100.

Crop rotations, or the practice of growing a sequence of crops on an agricultural field within or across growing seasons, is a key component of agricultural management and has impacts on overall crop yields, nutrient cycling, fertilizer and water usage, water quality, and biodiversity (Bullock, 1992). An example of such a crop rotation is the corn–soybean–corn rotation practiced extensively in the US Midwest. We generated a national-scale crop rotation dataset for the US to quantify rates of transition from one crop functional type to another and applied those rates to the crop functional types in LUH2. We use the USDA Cropland Data Layer (CDL; Sahajpal et al., 2014) to quantify unique crop rotations for the US from 2012 to 2014 (Sahajpal et al., 2014). Assuming a crop rotation span of 3 years and nearly 100 unique crops in the CDL, we could potentially have 106 unique crop rotations. Empirically, there are close to 100 000 unique crop rotations in the US for that time period. However, by aggregating different crop types to the crop functional types in LUH2 and merging similar rotations, we estimated transition rates between different crop functional types in LUH2 and applied them after all other transitions between land use types had been computed.

Historical information on crop management activities included data on irrigation, flooded agriculture, and industrial nitrogen fertilizer application rates. Data on irrigated area and area of flooded rice were obtained from HYDE. The irrigated fraction of each crop type was computed during the historical period by dividing the HYDE 3.2 irrigated fraction of each grid cell by the HYDE 3.2 cropland fraction of each grid cell. This fraction is then used as the irrigated fraction of each crop subtype.

The fraction of C3 annuals flooded for rice is computed in the historical period by dividing the HYDE 3.2 flooded fraction of each grid cell by the C3 annual fraction of each grid cell (rice is the only C3 annual considered to be flooded in our dataset; non-flooded rice is not explicitly represented here but would be included in the non-flooded C3 annual fraction).

For industrial nitrogen fertilizers, we used a recent global compilation of N fertilizer use for 1961–2011 (Zhang et al., 2015) based on FAOSTAT (FAO, 2020b) as our base dataset. Countries without fertilizer data reported in Zhang et al. (2015) were assigned regional mean values based on the regional grouping of countries defined in Zhang et al. (2015). Fertilizer use between 1915 and 1960 was hindcast using global synthetic N fertilizer use totals from Smil (2001) and was forecast from 2012 to 2015 using an estimate of global industrial N fertilizer use based on data from the International Fertilizer Association (IFA, 2015). Decadal mean N fertilizer rates by crop and country were computed from the Zhang et al. (2015) data and were assigned to the mid-decade year (e.g., the 1961–1970 mean was assigned to 1965). To generate country fertilizer application rates for 2015, which we did not compute as a decadal mean, we assumed that the fertilization rate since 2005 has changed with the same scaling factor across all countries and crop types (as in Zhang et al., 2015). Using the harvested area in 2015 from HYDE 3.2 (see Sect. 2.1), the fertilization rate for country j and crop k in 2015 is determined by Rj,k,2015=Rj,k,2005⋅(F2015,IFA/A2015)/(F2005/A2005), where Rj,k,t is the N fertilization rate by crop type (j) for each country (k) by year (t) (kg N ha-1 yr-1), and At is the global total crop area in year t from HYDE 3.2; F2015,IFA is the global N fertilizer application in 2015 estimated by applying the trend in 2006–2012 from the IFA data to extrapolate to 2015 from 2012, yielding F2015,IFA=115 Tg N yr-1, and F2005 is the global total N fertilizer application estimated as the product of the N fertilizer application rate in 2005 computed from Zhang et al. (2015) and LUH2 cropland area (F2005=94 Tg N, the mean of 2001–2010, as above).

Fertilizer application rates were hindcast from the 1960s to rates for 1950, 1930, and 1915. Synthetic N fertilizer rates in 1915 (and earlier) are set to 0.0 kg N km-2 for all countries and crop types, as this was when the Haber–Bosch industrial process was invented. Using global N consumption data from Smil (2001) for 1950 (F1950,Smil=3.7 Tg N yr-1) and 1930 (F1930,Smil=1.0 Tg N yr-1), as well as crop area from LUH2 (Aj,k,t, see Sect. 2.1), the synthetic N rates by crop and country (Rj,k,t) were estimated for 1950, 1930, and 1915 as follows: Rj,k,1950=Rj,k,1965⋅(F1950,Smil)/Σ[Rj,k,1965⋅Aj,k,1950],Rj,k,1930=Rj,k,1965⋅(F1930,Smil)/Σ[Rj,k,1965⋅Aj,k,1930],Rj,k,1915=0, where the sum is over all countries (j index) and crops (k index). Finally, we generated annual synthetic N fertilizer rate values by country, crop, functional type, and year (Rj,k,t) by linearly interpolating between values for 1915, 1930, 1950, 1965, 1975, 1985, 1995, 2005, and 2015.

We considered shifting cultivation to be a specific land use sequence of clearing, agricultural use typically for 1 to several years, and subsequent abandonment of land to forest (or other natural vegetation) regeneration for 3 years to several decades (“fallow”). While likely widespread in the early millennia of agriculture (Olofsson and Hickler, 2007), more recently it has been restricted to the tropics (Ruthenberg, 1980). We use the recent analysis of the past, present, and future extent of shifting cultivation (Heinimann et al., 2017) to constrain its occurrence in LUH2. Heinimann et al. (2017) based their analysis on the early global map of the distribution of “primitive subsistence agriculture” (Butler, 1980), a visual inspection of the distribution of shifting cultivation based on the 2000–2014 Global Forest Change (GFC) dataset (Hansen et al., 2013) coupled with high-resolution satellite imagery, and an extensive expert survey on regional trends in shifting cultivation, querying lead authors of scientific publications on shifting cultivation over the past decade (Heinimann et al., 2017).

Heinimann et al. (2017) estimated the current area under shifting cultivation (cultivated + fallow) to be about 280 Mha, distributed extensively and heterogeneously across central and tropical South America, tropical Africa, and tropical Southeast Asia (see Fig. 5 in Heinimann et al., 2017). For each 1∘×1∘ grid cell with detected signs of shifting cultivation, they also estimated its level of occurrence, including both active and fallow cropland, aggregated into five classes of the total land area in each grid cell: none (< 1 %), very low (1 %–9 %), low (10 %–19 %), moderate (20 %–39 %), or high (≥ 40 %). They project significant declines in shifting cultivation extent through the 21st century, with losses by the end of the century of more than 80 % in Africa and Latin America and 100 % in Asia, with extent at 1∘×1∘ in remaining areas projected to be low or very low (see Fig. 7 in Heinimann et al., 2017).

We created annual LUH2 shifting cultivation maps by linearly interpolating between the assumed shifting cultivation rates in 1850 and the expert-opinion-based rates of 2010 (Heinimann et al., 2017). The 1850 shifting cultivation rates were assumed to fall in the high category of 70 %. The future shifting cultivation rates were similarly computed by linearly interpolating between the 2010 and the assumed 2100 rates from the expert opinion survey of Heinimann et al. (2017). For LUH2, shifting cultivation involved cropland only (grazing land was included as part of shifting cultivation in LUH1 but not in LUH2). For all grid cells, we used the mid-range of shifting cultivation occurrence (e.g., 5 % for “very low”, 15 % for “low”, 30 % for “moderate”, and 70 % for “high”) and assumed that these fractions also applied to the fraction of cropland involved in shifting cultivation. We also assumed that the residence time for a patch of cropland involved in shifting cultivation was only 1 year. At each time step in our model, we then abandoned the Heinimann et al. (2017) prescribed percentage of total cropland area in the grid cell (e.g., cropland to secondary land) and cleared the same area from natural vegetation (e.g., forest to cropland), with a prioritization of clearing secondary land first unless the available secondary land was less than 10 times the cropland area involved in shifting cultivation (based on an assumption of a 10-year fallow period). The global area of shifting cultivation activity tends to track global changes in cropland area from HYDE 3.2 (Klein Goldewijk et al., 2017, or see Sect. 2.1) and global future cropland area changes from IAMs, although this relationship between cropland area and shifting cultivation area declines over time due to the extent of shifting cultivation declining significantly, especially through the 21st century.

Historical wood harvest in LUH2 is based on national statistics and partitioned into fuelwood and non-fuelwood for 199 countries based on a 1990 country list from HYDE 3.2 (Klein Goldewijk et al., 2017). These national wood harvest statistics are used to solve Eq. (1) and assigned to individual grid cells using the methodology described in Sect. 2.10 and 2.11. For the years 1961–2015 the LUH2 wood harvest data are based on FAO national wood harvest volume data (FAO, 2020c) for both coniferous and non-coniferous round wood, which is combined with wood density values of 0.225 Mg C m-3 for coniferous wood and 0.325 Mg C m-3 for non-coniferous wood (Houghton and Hackler, 2000) to convert volume statistics to mass of carbon harvested. Harvest rates were hindcast to 1920 by interpolating from mean FAO per capita harvest rates from 1961 to 1965 using national population totals from HYDE 3.2 (see Sect. 2.1), as well as national per capita fuelwood (“firewood”) and timber (“sawtimber”) wood harvest totals from 1920 (Zon and Sparhawk, 1923). Note that the Zon and Sparhawk totals for timber consumption include volume of wood for construction, industry, and pulp; so, with firewood, it should be roughly comparable to FAO “total roundwood”.

For the years prior to 1920, national annual per capita wood harvest rates were computed in three different ways for low, baseline, and high LUH2 scenarios, and they use the same national population data from HYDE 3.2 to compute the total national wood harvest (Mg C) per year for each scenario. For the low wood harvest scenario, the national annual per capita wood harvest rates from Zon and Sparhawk (1923) were held constant for all years from 850 to 1920. However, prior to the fossil fuel era, global mean per capita wood harvest was likely significantly higher than in 1920, so for the high scenario we used a national per capita wood harvest demand reconstruction for “fuelwood” and “durable wood” from Kaplan et al. (2017) for the period 850–1800. Per capita wood harvest rates then transitioned linearly from 1800 rates to the 1920 rates of Zon and Sparhawk (1923) to mimic the global shift in energy sources from biomass towards fossil fuels (Smil, 2003). These high and low wood harvest scenarios represented two different extremes in terms of cumulative wood harvested and total area of forests removed. In addition, the high scenario is significantly higher than the LUH1 wood harvest reconstruction. To provide a scenario somewhere between these two extremes, we also generated a baseline wood harvest scenario in which we modified the Kaplan national wood harvest rates from 850 to 1800 by national-scale factors. These scale factors are defined as twice the contemporary FAO national per capita wood harvest rates divided by the national per capita wood harvest rates in 1800 from the Kaplan data, and this definition was determined from analysis of the global time series figure of historical biofuel consumption (Smil, 2003), which shows current global per capita biofuel consumption of around 6 GJ per capita and around 21 GJ per capita in 1800. Reducing the Kaplan wood harvest rates via these scale factors does not imply that the original Kaplan rates are too high; rather, the Kaplan data are likely to be capturing types of wood harvest and related processes that our model does not currently simulate. For years between 1800 and 1920 we linearly interpolate between the modified year 1800 rates from Kaplan and the Zon and Sparhawk (1923) rates in 1920.

For the low and baseline scenarios, the reconstructed national wood harvest data were increased by a slash fraction of 30 % (as in LUH1; Hurtt et al., 2011) to account for non-harvested losses from forests that occur during the wood harvesting process. For the high scenario, we do not add a slash fraction to the data for the years 850–1800 since it is assumed this is already included in the Kaplan data (Kaplan et al., 2017). In this scenario, the slash fraction is linearly increased from 0 % to 30 % during 1800 to 1920 and held constant thereafter.

All national wood harvest totals from FAO and Zon and Sparhawk are assumed to represent the amount of wood produced by each country. In contrast, the data from Kaplan represent the wood harvest demand from each country, although it is assumed that during the years 850–1800 there was limited wood trade in most parts of the world, and hence demand would equal production. In Europe, however, international wood trade occurred during 850–1800 (Kaplan et al., 2017). So, for European countries only, if the available national biomass is not sufficient to meet the national wood harvest demand in a particular year, we seek the unmet demand from other European countries (i.e., increase the wood harvest production in other countries) proportional to the available biomass in each country. From 1500 to 2005, the global cumulative total wood harvest in the baseline scenario was 190 Pg C, including slash (Fig. 2), compared with 142 and 381 Pg C in the low and high scenarios, respectively.

The spatial patterns of forest transitions, particularly those related to wood harvesting, were constrained by the Landsat-based gridded forest loss observations from Hansen et al. (2013). This product consists of global 30m grids of tree canopy cover for the year 2000 and gross forest cover loss and gain for the 2000–2012 time interval mapped using the entire global Landsat data archive (although only the forest loss data were used within LUH2). Within this dataset, forest was defined using a single tree canopy cover threshold to match the global forest extent provided by the FAO FRA report (FAO, 2000). Cumulative forest area was estimated by summing pixels with different tree canopy cover. Then the threshold was selected that most closely enabled a match to the total world forest cover for the year 2000, which is 4085 million ha, according to FAO data. A threshold of 28 % tree canopy cover produced 100.5 % of the FAO forest area. This threshold was used to define forest area for the year 2000 at 30 m spatial resolution. Gross forest cover loss was reported only within areas covered with forest in the year 2000. Gross forest cover gain was mapped independently outside areas forested in the year 2000 and represents a gain of tree canopy cover to 30 % or higher from non-forest state. The global maps of forest extent and change were then aggregated to the same spatial resolution and format as the LUH1 datasets (0.5∘×0.5∘ fractional). To aggregate the data to the 0.5∘ grid, the area of each class was computed within each grid cell, and then the class area percent of total cell area was calculated. The 0.5∘ product shows percent forest cover for the year 2000 and percent gross forest cover loss and gain during the 2000–2012 time interval. The 0.5∘ product was later downscaled to 0.25∘ for consistency with the new LUH2 spatial resolution. A very simple downscaling method was employed that kept the fraction of forest area (or forest loss) equal within each 0.25∘ grid cell inside the 0.5∘ grid cell.

The resulting map of forest loss was used within LUH2 as part of the algorithm for determining the spatial pattern of forest loss from wood harvesting. However, it should be noted that the Landsat-based forest loss maps differ from the LUH2 forest loss maps in multiple ways, including definitions of “forest” (i.e., tree canopy cover vs. biomass density), whether or not a single grid cell can contain both forest and non-forest (LUH2 grid cells are either potentially forested or potentially non-forested), and whether or not the forest loss includes natural disturbances such as fires (LUH2 forest loss results only from land-use-related changes). As a result, the match between these products is not perfect, and the Landsat-based forest loss data are used as a guide to improve the LUH2 forest loss patterns rather than a hard constraint on those patterns.

To discriminate forested land from non-forested land and to convert quantities of harvested wood in biomass units into harvested area, information was needed on the historical distribution of forests and aboveground carbon stocks. As no complete global, gridded, historical record of these quantities was available, a simple empirically based global terrestrial model was used to provide a consistent set of both global forest cover and carbon stocks. Estimates of ecosystem properties were based on an updated version of the MIAMI-LU ecosystem model (Hurtt et al., 2002, 2006, 2011). Miami-LU was driven by the empirically based Miami model of net primary production (Leith, 1972), which has integrated sub-models of plant mortality and disturbance. The model tracked subgrid heterogeneity resulting from land use changes in a manner similar to the more advanced Ecosystem Demography (ED) model (Hurtt et al., 1998, 2002; Moorcroft et al., 2001).

Miami-LU was run globally at 0.5∘×0.5∘ resolution for a spin-up period of 500 years using data from the Multi-Scale Synthesis and Terrestrial Model Intercomparison Project (MsTMIP) (Wei et al., 2014). These data are a combination of climatologies from the Climate Research Unit and National Centers for Environmental Protection, and they have a global 0.5∘×0.5∘ climatology with a 6-hourly daily time step from 1901 to 2010. MIAMI-LU outputs were subsequently downscaled to 0.25∘×0.25∘ resolution to match the remaining LUH2 inputs (downscaling simply assigned all 0.25∘×0.25∘ grid cells the same fraction value as the 0.5∘×0.5∘ grid cell they were contained within). Aggregated globally, the net primary production (NPP) estimate from Miami-LU was 63 Pg C yr-1. This fell within a range of NPP estimates from various global biogeochemical models, ranging from 40 to 81 Pg C yr-1 (Cramer et al., 1999). Miami-LU estimated a global stock of potential plant carbon of 718 Pg C (Fig. 3). This fell within a range spanning 557 Pg C (Kucharik et al., 2000) to 923 Pg C (Sitch et al., 2003), with a more recent estimate of 772 Pg C (Pan et al., 2013). The total potential aboveground carbon stock was 563 Pg C. To differentiate forest from non-forest areas, a definition based on potential aboveground standing stock of 2 kg C m-2 was used (Hurtt et al., 2002, 2006, 2011). Each grid cell was thus identified as potential forest or potential non-forest based on potential biomass, providing a static map that is used for the entire time period from 850 to 2100. Using this definition, 48.8×106 km2 of the land surface was classified as potential forest. For comparison, potential forest area based on the BIOME model was estimated at 60×106 km2 (Klein Goldewijk, 2001). Finally, Miami-LU was also used to estimate the recovery of carbon stocks on secondary lands by tracking the mean age of secondary land in each grid cell, although this does not explicitly account for the full age distribution or the potential effects of land degradation, management, or pollution that may have occurred.

For 2015–2100, we use land use and wood harvest information from eight different marker SSP–RCP scenarios derived from five different Integrated Assessment Models (Riahi et al., 2017). These marker scenarios were prioritized as input to CMIP6 climate model simulations by ScenarioMIP. They are fully described elsewhere (O'Neill et al., 2016; Riahi et al., 2017), and their main features are summarized below and in Table 2 in the order described in O'Neill et al. (2016).

Harmonization of inputs involved minimizing the difference between the end of the historical reconstruction and the beginning of future projections, as well as preserving as much information on the future from IAMs as possible. Five different IAMs provide future land use, wood harvest, and management data using a variety of variables and units at different spatial and temporal resolutions (Table 2). Prior to harmonization, inconsistencies in definitions, resolutions, and other factors resulted in significant discrepancies. The spread of global cropland values from the IAMs in 2010 was 5 % of the historical reconstruction values in that year, and the spread of global pasture values from the IAMs in 2010 was 23 % of the historical values. Gridded values had even larger discrepancies, differing by as much as 100 % from the historical values. After harmonization, these inconsistencies were eliminated by design of the harmonization methodology. Since some IAMs did not simulate built-up area or urban spread, and for consistency of urban land definitions across all scenarios, the IMAGE model provided land use inputs for built-up area in all scenarios (Doelman et al., 2018). Also, since the REMIND–MAgPIE model did not compute wood harvest amounts, these were provided for the SSP5-8.5 and SSP5-3.4OS scenarios from analogous scenarios computed by GCAM.

The first step in harmonizing inputs was to convert the IAM data into a standardized format for comparison with the historical product. Future land use data were aggregated into the fractions of each grid cell occupied by total cropland, total grazing land (the sum of managed pasture and rangeland), urban land, and natural vegetation (the sum of primary and secondary forest and non-forest) annually at 0.25∘×0.25∘ resolution. Future data on irrigation and flooded areas were standardized into national totals. Future wood harvest data were standardized into a total national wood harvest demand in megagrams of carbon per year (Mg C yr-1), as was the fuelwood component of that national wood harvest, either by aggregating gridded wood harvest data into national totals or by disaggregating regional wood harvest data using the ratio of national to regional wood harvest from the end of the historical period (i.e., 2015). Wood harvest data that were provided in volume units (m3) were converted to biomass (Mg C) using a conversion factor of 0.2688 Mg C m-3. A 30 % slash fraction was added to the wood harvest scenarios. Future fertilizer rates were standardized into national fertilizer application rates in kilograms of nitrogen per hectare per year (kg N ha-1 yr-1) per crop functional type. For future scenarios with only regional data, all countries within a region were assigned the same regional rates. When gridded future fertilizer application rates were available these were also used in LUH2 and were standardized into annual rates per crop type (kg N ha-1 yr-1) at 0.25∘×0.25∘ resolution. For SSP4-3.4 and SSP4-6.0 (both from GCAM), the fertilizer rates for the GCAM crop types misccrop and palmfruit were used as estimates of fertilizer rates for C3 perennials, sugarcrop and biomass rates were used as estimates for C4 perennial rates, oilcrop and misccrop rates were used for C3 nitrogen-fixing crops, rice and wheat were used for C3 annuals, and corn was used for C4 annuals.

Although the IAM land use data were generally in good agreement with end-of-historical-period values at the global scale, there were still significant differences both globally and spatially, particularly for pasture, which has less consistent definitions across models (Fig. 4). To address this issue, we applied IAM-based annual changes in land use sequentially to the spatial pattern of land use at the end of the historical reconstruction. Annual future changes in cropland, grazing land, and urban land were computed and aggregated to 2∘×2∘. These changes were then applied to the 2∘ aggregated cropland, grazing land, and urban land from the previous time step, starting with the end of the historical period (i.e., 2015). When it was not possible to apply the annual change within a 2∘ grid cell due to lack of available land to expand into or lack of cropland, grazing, or urban land to abandon, the unmet changes were applied in neighboring 2∘ grid cells, starting with immediate neighbors and then radiating outward. The harmonized grids of cropland, grazing land, and urban land were then disaggregated into 0.25∘×0.25∘ grids according to the following method: when disaggregating decreases, the percentage change in each land use state was computed and then applied to all underlying 0.25∘ land use fractions; for increases in cropland, grazing, or urban land, the needed change was applied across all underlying 0.25∘ grid cells and was weighted by available land in each grid cell. Figure 5 shows how well the IAM 2015–2100 changes in cropland and pasture fractions are retained in the harmonized data, which increases markedly with decreased spatial resolution. For wood harvest, analogous methods were applied.

After the harmonization of total cropland, grazing land, and urban land, cropland and grazing areas were further disaggregated into underlying subtypes. Assignment of future crop functional types were based on fixed contemporary Monfreda–FAO proportions and adjusted to match IAM-specific information as needed. For grazing land, a pasture–rangeland mask was generated for 2015 (and held constant for all years) to subdivide future total grazing land into the two grazing subtypes. For new grid cells projected to be converted to grazing land in the future, national ratios were used.

Next, management data were harmonized by applying analogous algorithms to sequentially apply projected changes in managed area and rates to the pattern at the end of the historical reconstruction. Annual changes in national irrigated areas were computed and then applied to the previous year's gridded irrigation fractions for all crop types, first increasing irrigated area on grid cells with existing irrigation, and then adding any additional needed irrigated area equally to all nonirrigated cropland grid cells within each country. Annual national percentage change in flooded area was computed, and this percentage change was applied to all grid cells that have a nonzero flooded fraction in the previous time step. Any resulting fractions that are greater than 1 are reset to 1. Finally, annual national percentage changes in fertilizer rates per crop type are computed. These national percentage changes are applied to the previous year's gridded fertilizer rates for all grid cells within each country. In an effort to ensure that the final (year 2100) gridded fertilizer rates closely approximate the future IAM fertilizer rates, there are a few exceptions to this method, which are based on simple assumptions that aim to keep the LUH2 rates from remaining too low or becoming too large when compared to the IAM gridded rates. First, the gridded fertilizer rates are held between 0 and 500 kg N ha-1 yr-1. Then, for grid cells with fertilizer rates below 1 kg N ha-1 yr-1 in the previous time step and with an increasing national percentage change in fertilizer rates, the actual gridded IAM fertilizer rates for the next time step are used instead of the computed LUH2 rates. Also, if gridded fertilizer rates increase between time steps and are above the gridded IAM fertilizer rates, the gridded fertilizer rates for the next time step are held constant at the current LUH2 gridded rates. Finally, if the gridded LUH2 fertilizer rates are less than 80 % of the IAM gridded fertilizer rates and the national percentage change in fertilizer rates is positive, a small additional increase (1 % of the total current difference between IAM gridded rates and LUH2 gridded rates) is added to the LUH2 fertilizer rates.

While it is primarily a land use dataset, LUH2 also provides a simple estimate of forest cover change. For IAM future scenarios with positive forest cover gain (SSP1-2.6, SSP2-4.5, SSP1-1.9), an algorithm was developed to match the spatial pattern of forest gain from IAMs, preserve existing harmonized land use transitions, and be implemented relatively easily in ESMs. For each scenario, a supplementary file was created with a data variable called added_tree_cover. The variable specifies the added tree cover that needs to be planted in each grid cell each year to better represent the corresponding IAM added tree cover estimates. For the other IAM scenarios that are not affected by this issue, added_tree_cover values are set to zero. To produce these datasets, the spatial patterns of differences in forest cover between LUH2 and each corresponding IAM were computed annually for 2015–2100. For each year and each grid cell, if the difference could be met on LUH2 classified non-forest land, that difference was noted as added_tree_cover in the new file. If the gain could not be met on the non-forest area, the change was applied to nearby cells up to four grid cells away.

In addition to the eight future scenarios for the period 2015–2100, the LUH2 dataset also includes extensions for the years 2100–2300 for three of the harmonized future land use forcing datasets for use in long-term climate stabilization experiments. By design, in these extensions, all land use states and management variables are held constant at year 2100 values for the years 2100–2300. As a result, almost all transitions between land use states are set to zero, with the exception of crop rotations and shifting cultivation, which continue at their year 2100 rates, and wood harvest, which uses the year 2099 national wood harvest demands for all years from 2100 to 2299. These extensions to future scenarios are available for SSP1-2.6, SSP5-3.4OS, and SSP5-8.5.

The annual gridded land use states are aggregated to annual global values by multiplying the grid cell land use fractions by the grid cell area and summing over all grid cells (Fig. 7). The 12 land use states represented in the LUH2 dataset can be further aggregated into the five broader land use categories of total cropland (the sum of all five crop types), total grazing land (the sum of managed pasture and rangeland), primary land (the sum of primary forest and primary non-forest), secondary land (the sum of secondary forest and secondary non-forest), and urban land. Historically, the area of cropland increased at an accelerating rate from 1.7×106 km2 in 850 to 4.3×106 km2 in 1800 and 15.9×106 km2 by 2015 (Fig. 7). Grazing lands increased more rapidly from 3.3×106 km2 in 850 to 9.2×106 km2 in 1800 and to 32.8×106 km2 by 2015. Urban increased from 0 in 850 to 0.6×106 km2 by 2015. See also HYDE 3.2 on the historic trends of cropland and pasture (Klein Goldewijk et al., 2017). During the historical period (850–2015 CE), primary land area decreased from 125×106 to 50.1×106 km2 (44 % of which is forested), while secondary land increased from 0 to 30.4×106 km2 (approximately 49 % of which is forested); note that by definition LUH2 initializes secondary land area to zero in 850 CE. The new land use history reconstruction derived here generally compared favorably to prior reconstructions (Hurtt et al., 2006, 2011) and other references across a range of important diagnostics (Table 3), albeit at higher spatial resolution and with more process detail.

For the future, all eight scenarios projected increases in global cropland area, while six projected grazing land decreases (SSP4-RCP6.0 from GCAM and SSP3-RCP7.0 from AIM projected grazing land increases). The global and regional trends of agriculture and land use in these eight projections are described in detail in Popp et al. (2017), and underlying drivers of these land use dynamics have been identified in Stehfest et al. (2019). For nonagricultural land, six out of eight scenarios projected large increases in wood harvesting, which contributed to large increases in secondary area and corresponding reductions in primary area by 2100. In 2100 global cropland ranged from 17.8×106 km2 (SSP1-RCP2.6 from IMAGE) to 29.1×106 km2 (SSP4-RCP3.4 from GCAM). As shown in Table 4 and Fig. 15 (panel a), for six out of eight scenarios the dominant crop functional type in 2100 was C3 annuals, with C4 perennials (for biofuels) as the dominant crop functional type in 2100 for the remaining two scenarios (SSP4-RCP3.4 from GCAM and SSP5-RCP3.4OS from REMIND–MAgPIE). Global grazing land in 2100 ranged from 25.4×106 to 35.5×106 km2, with the majority of that coming from rangeland (Table 4). Secondary land in 2100 ranged from 36.5×106 to 44.5×106 km2 (Table 4). In all cases, approximately half of all secondary land was forested, and the estimated mean age of secondary forest ranged from 58 to 74 years. Added tree cover data layers were computed to match the forest tree cover gains of the SSP1-2.6, SSP2-4.5, and SSP1-1.9 scenarios and were able to capture > 80 % of the global afforestation signal in the IAM scenarios. Extensions to the year 2300 were computed for the SSP1-2.6, SSP5-3.4OS, and SSP5-8.5 scenarios and by design did not change the gridded or global cropland, grazing land, or urban land areas. However, due to wood harvesting and shifting cultivation continuing at their end-of-century rates, the area of secondary vegetation continued to grow, and the area of primary vegetation continued to decline in these extensions. By 2300 the global secondary vegetation area in these extension scenarios ranged between 46.3×106 and 51.2×106 km2, while the global primary vegetation area ranged between 28.6×106 and 33.0×106 km2.

Gross transitions (the sum of the absolute value of all land use transitions) are a measure of all land use change activity. In general, the annual gross transitions tend to increase through time, beginning at 2×105 km2 in 850 and increasing to 1.86×106 km2 in 2000 (Table 3). The differences between the historical period low, baseline, and high scenarios in LUH2 (computed using three different HYDE land use reconstructions and three different national wood harvest reconstructions) prior to 1920 are primarily due to the differences in rates of wood harvest between those three scenarios. After 1920 the three LUH2 historical scenarios share the same wood harvest reconstruction and their associated gross transitions are very similar. In the future scenarios, gross transitions mostly increased and by 2100 ranged from 2.0×106 to 4.8×106 km2 (Table 5).

Net transitions measure only the net changes into land use (excluding wood harvest on secondary forests, shifting cultivation, and other agricultural land abandonment that is offset by land conversions to agriculture). Net transitions increase from 2×104 km2 in 850 to 2.3×105 km2 in 2000 (Table 3). The net transitions across all three historical LUH2 scenarios (low, baseline, and high) are very similar at most time points. The LUH2 historical scenario shows a significant reduction in transitions to pasture around 1950–1960, with implications for carbon investigated separately (Ma et al., 2020). In the future, net transitions range from -1.1×105 to 1.6×105 km2 in 2100 (Table 5).

To visualize the magnitudes of transitions between variables, we present chord diagrams indicating the average net transitions occurring annually for 850–1849, 1850–2015, 850–2015, and 2015–2099 for all future scenarios amongst all the major land use categories (Fig. 8). Each arc in a chord diagram represents the average annual area transitioning from one land use to another. The color of the arc represents the land use category from which transition to a different category occurs. For example, in Fig. 8 the arc in light green represents the transition from cropland to other categories. Transitions involving croplands and secondary forest lands dominate land use transitions in all three historical scenarios. The dominant land use transition is secondary forest lands to croplands, and it ranges from nearly 6×104 km2 yr-1 in the low historical scenario to 8×104 km2 yr-1 in the baseline scenario and 1×105 km2 yr-1 in the high scenario when averaged from 850 to 2015. Cropland abandonment activities are also significant, with nearly 1×105, 1.4×105, and 1.7×105 km2 of croplands transitioning annually to secondary lands (both forested and non-forested) in the low, baseline, and high LUH2 historical scenarios, respectively (averaged over the entire historical period). On an annual basis, the transitions to and from croplands and secondary lands are generally the same in all three LUH2 historical scenarios.

LUH2 historical results were compared to multiple diagnostics (Table 3). Almost all metrics are within or very close to published reference ranges. These metrics show that 65 % of the secondary land increase between 1700 and 2000 is forested, and 93 % of US forests in the year 2000 are on secondary land. Global natural vegetation in biodiversity hotspots in the year 2005 is estimated as 1.6 % of the land surface (compared with the reference value of 2.3 %). The mean age of secondary land can be calculated for each grid cell and aggregated to a global mean age. For the first several hundred years of the simulation the global mean secondary age grew with time due to primary land being used for land conversion and wood harvesting more often than secondary land (which was initialized to have zero area). Around 1700–1800, existing secondary land was used more often for new land conversions and wood harvesting, and the global mean secondary age started to decrease with time. The median age of secondary forests in the year 2005 is 42 years and is 43 years in the year 2015 (compared with the reference range of 30–40 years). The high scenario had the highest secondary mean age because it had a larger secondary land area, which allows secondary land to be used less frequently for wood harvesting and land conversions. Conversely, the low scenario had a lower secondary mean age than the baseline scenario. The overall land area impacted by human land use in the year 2000 is 59 % of the land surface. The global area of secondary land increase between 1700 and 2000 is estimated as 13.2×106 km2, with 10.4×106 km2 of that area forested and 2.8×106 km2 non-forested.

Cumulative clearing for cropland and pasture between the years 1500 and 1990 resulted in 251 Pg C of wood being removed (compared with a reference range of 121.9 to 356.3 Pg C). Total wood harvest over this period was 170 Pg C, 132 Pg C of which was from direct wood harvest and 38 Pg C was included from agricultural clearing. In the year 2000, an estimated 0.32×106 km2 of agricultural land was involved in shifting cultivation (compared with a reference value of 0.3×106 km2). Potential forest area was 47×106 km2 compared to a reference value of 52×106 km2, and in the year 2015 global forest area was estimated at 37×106 km2 compared with a reference range of 32–41×106 km2. In the year 2000 global wood harvest was 1.29 Pg C, 0.71 Pg C of which was for fuelwood. Global synthetic fertilizer usage in the year 2012 was 106.6 Tg N yr-1 (compared with a reference value of 100 Pg C), and the global area of irrigated cropland in 2003 was 2.51×106 km2 (compared with a reference value of 2.77×106 km2). In 2004, the area of cropland (primarily corn) used for biofuels was 0.03×106 km2 compared to the reference value of 0.033×106 km2. Total potential plant biomass on all lands was 718 Pg C (compared with a reference range between 557 and 923 Pg C), while total plant biomass in 2005 was 434 Pg C (compared with a reference value of 393 Pg C). Plant aboveground biomass on pantropical forested lands between the years 2007 and 2008 was 184 Pg C (compared with a reference range between 188 and 229 Pg C), and total plant biomass on forested lands in 2005 was 395 (compared with a reference value of 363 Pg C). In addition, the cumulative loss of aboveground biomass resulting from land use transitions (i.e., the sum of all losses) is an important metric of the gross effects of land use on the terrestrial carbon cycle and rose from 0 Pg C in 850 to 5.6×104 Pg C in 2015. Similarly, the cumulative net loss in aboveground biomass is the difference between the estimated aboveground biomass, including land use, and the estimated biomass of potential vegetation; it includes both the losses of aboveground biomass due to land use and the gains due to regrowth. During the historical period the global cumulative net loss of aboveground biomass carbon increases monotonically from nearly zero in 850 to around 310 Pg C in 2015. The low, baseline, and high historical scenarios all give similar global estimates of this metric; the high scenario gives the highest estimates, which is presumably due to the high historical wood harvest in this scenario.

In the future scenarios secondary land increases between 6.0 % and 13.27 % across the years 2015 to 2100, with between 48.9 % and 72.8 % of that increase being on potentially forested land (Table 5). The median age of secondary forest in the year 2100 ranges between 58 and 74 years. The global area covered by natural vegetation in the biodiversity hotspots ranges between 0.57 % and 1.08 % of the land surface. Wood clearing for cropland and pastures across the years 2015 to 2100 removes between 44 and 88 Pg C of aboveground biomass, whereas direct wood harvest removes between 93 and 148 Pg C of aboveground biomass. Global wood harvest in the year 2100 ranged between 0.9 and 1.87 Pg C, the fuelwood component of which was between 0.15 and 0.88 Pg C. Total forest area change between 2015 and 2100 ranged from a decrease of 5.1×106 km2 to an increase of 3.42×106 km2, resulting in a global forest area in 2100 of between 32.1 and 38.1×106 km2. Global fertilizer use in the year 2100 ranged between 110 and 240 Tg N yr-1, while the global irrigated area in 2100 ranged between 2.6 and 4.1×106 km2. Land flooded for rice in 2100 ranged from 0.23 to 0.96×106 km2, and cropland used for growing biofuels in 2100 ranged from 0 to 18×106 km2. Total biomass of natural vegetation on forested lands in 2100 ranged between 290 and 391 Pg C, between 170 and 239 Pg C of which is aboveground biomass on pantropical forested lands. In 2100, the global cumulative net loss of aboveground biomass carbon ranges widely across scenarios from 320 to 385 Pg C.

Regional results for the historical period, averaged for each century, are shown in Table 6. In each region or continent, secondary land, gross transitions, and net transitions all tended to increase with time. Secondary land, along with both gross and net transitions, was highest in Eurasia and Africa. Mean regional secondary land area was 8.47×106 km2 in Eurasia and 6.01×106 km2 in Africa in the 1700s and increased to 12.4×106 km2 and 6.82×106 km2 in Eurasia and Africa, respectively, in the 1900s. Gross transitions peaked in Eurasia in the 1800s at 660×106 km2 yr-1, while net transitions peaked in Eurasia in the 1900s at 121×106 km2 yr-1. After 1700, secondary age tended to decrease with time for most regions, although it has held relatively constant over the last 3 centuries for both Africa and Oceania. The range of secondary mean age in the 1900s was between 52 and 289 years. In 1850 there are large areas of cropland in the eastern USA, Europe, India, and China, as well as large areas of primary land worldwide with the exception of Europe, northern Africa, and the Middle East (Fig. 9). By 2015 cropland areas have expanded throughout Africa and the Americas as well, primary land is lost in large areas of the eastern USA, Africa, Europe, India, and China, and the mean secondary age is lower in most locations (Fig. 10).

Regional results are also averaged for the period 2000–2099 for each future scenario (Table 7). Across all scenarios, there were only small differences in regional secondary areas (3.8–4.5×106 km2 for North America, 2.0–3.0×106 km2 for South America, 17–18×106 km2 for Eurasia, 9.2–11×106 km2 for Africa, and 0.7–0.87×106 km2 for Oceania), with SSP1-1.9 having the highest secondary area on each continent. Secondary land area was highest in Eurasia and Africa for all scenarios. Regional secondary age also did not vary significantly across scenarios; the SSP5-8.5 scenario had the highest secondary age for all regions except Oceania (67 years for North America, 49 years for South America, 209 years for Eurasia, 70 years for Africa, and 50 years for Oceania), and the SSP4-3.4 scenario had the lowest secondary age for most regions (60 years for North America, 45 years for South America, 197 years for Eurasia, 69 years for Africa, and 48 years for Oceania). Secondary age was highest in Eurasia for all scenarios. Gross transitions were highest in Eurasia in seven out of eight scenarios (with Africa the second highest) and highest in Africa in one scenario (with Eurasia the second highest). The highest overall rate of gross transitions was 1936×106 km2 yr-1 in Eurasia in the SSP5-3.4OS scenario, but comparable rates of gross transitions were also observed in Eurasia and/or Africa in the SSP4-3.4, SSP4-6.0, SSP3-7.0, and SSP5-8.5 scenarios. Net transitions were largest in Africa in all scenarios (between 34 and 143×106 km2 yr-1) and lowest in Oceania in seven out of eight scenarios (and negative in six of those), with South America having the lowest net transitions in the remaining scenario. The SSP4-3.4, SSP4-6.0, and SSP3-7.0 scenarios had the highest rates of net transitions overall at 143×106, 133×106, and 133×106 km2 yr-1, respectively.

Large-scale spatial patterns are similar across most scenarios in the year 2100 (Figs. 11–14), with the trends of increased cropland area in South America, continued loss of primary land worldwide and particularly in Africa, and continued reduction of mean secondary age. Analogous mapped results for Tier 2 scenarios are provided in the Appendix.

During the historical period, the use of synthetic nitrogen-based fertilizer on croplands was zero until the early 20th century. After 1950 fertilizer usage started increasing rapidly, and by 2015 global synthetic nitrogen fertilizer usage was 112 Tg N yr-1 (4150 Tg N cumulatively from 1915 to 2015; none prior to 1915), with the majority of this being applied in cropland-dominated locations including North America, Europe, India, China, and Southeast Asia. The eight harmonized future scenarios show a range of potential nitrogen futures; all except one scenario (the SSP5-8.5, which does increase but then falls again to close to current year values) project an increase in global nitrogen fertilizer usage. The range of harmonized global nitrogen fertilizer values in 2100 is between 110 and 240 Tg N yr-1, with the total cumulative use of synthetic nitrogen fertilizer from 2015 to 2100 between 9840 and 14 800 Tg N (Fig. 15b).

The global area of irrigated cropland increased steadily throughout the historical period and was around 2.7 million km2 in 2015. The spatial patterns of this irrigated area show that the majority of global irrigation occurs in India and China, with other significant areas in the USA, Europe, the Middle East, and Southeast Asia. Six out of eight future scenarios project the global irrigated area to remain steady or even decrease slightly, whereas two future scenarios (SSP3-7.0 and SSP5-8.5) show large increases in global irrigated area. The range of values across all future scenarios in 2100 is between 2.6 and 4.1 million km2 (Fig. 15c).

The global use of cropland area for purpose-grown biofuels was very low prior to the year 2000 when a small amount of first-generation biofuel production began (such as corn or sugarcane). In the future scenarios the fraction of cropland area grown for first-generation biofuels was held constant, although underlying changes in cropland area resulted in some small increases or decreases in the total area of first-generation biofuels. Second-generation biofuel area (such as miscanthus or switchgrass) expanded in each of the future scenarios, assumed to start from zero in 2015. Five of the eight scenarios (SSP1-1.9, SSP1-2.6, SSP4-3.4, SSP5-3.4OS, and SSP4-6.0) all showed significant increases in the area of second-generation biofuels, while the remaining three scenarios have very little growth in this land management type. By the year 2100, global areas of biofuel crops ranged between 0 and 18 million km2, and maps of the spatial distribution of total biofuel area (both first- and second-generation biofuels) show the dominant locations to be the USA, Europe, China, non-Amazonian Brazil, and Argentina. Large expansion of secondary biofuels primarily occurred in Southeast Asia, eastern Europe, the former USSR, and the Middle East (Fig. 15d).