Adaptation of global land use and management intensity to changes in climate and atmospheric carbon dioxide

Abstract Land use contributes to environmental change, but is also influenced by such changes. Climate and atmospheric carbon dioxide (CO 2) levels’ changes alter agricultural crop productivity, plant water requirements and irrigation water availability. The global food system needs to respond and adapt to these changes, for example, by altering agricultural practices, including the crop types or intensity of management, or shifting cultivated areas within and between countries. As impacts and associated adaptation responses are spatially specific, understanding the land use adaptation to environmental changes requires crop productivity representations that capture spatial variations. The impact of variation in management practices, including fertiliser and irrigation rates, also needs to be considered. To date, models of global land use have selected agricultural expansion or intensification levels using relatively aggregate spatial representations, typically at a regional level, that are not able to characterise the details of these spatially differentiated responses. Here, we show results from a novel global modelling approach using more detailed biophysically derived yield responses to inputs with greater spatial specificity than previously possible. The approach couples a dynamic global vegetative model (LPJ‐GUESS) with a new land use and food system model (PLUMv2), with results benchmarked against historical land use change from 1970. Land use outcomes to 2100 were explored, suggesting that increased intensity of climate forcing reduces the inputs required for food production, due to the fertilisation and enhanced water use efficiency effects of elevated atmospheric CO 2 concentrations, but requiring substantial shifts in the global and local patterns of production. The results suggest that adaptation in the global agriculture and food system has substantial capacity to diminish the negative impacts and gain greater benefits from positive outcomes of climate change. Consequently, agricultural expansion and intensification may be lower than found in previous studies where spatial details and processes consideration were more constrained.


| INTRODUCTION
Environmental change will influence future agricultural productivity.
Land use also creates important environmental impacts. For example, 24% of all anthropogenic greenhouse gas emissions (GHGs) in 2010 were associated with agriculture, forestry and other land use , and 11% of anthropogenic CO 2 emissions were associated with land use change (Le Qu er e et al., 2016). Expanding agricultural areas and intensifying production -that is, using more inputs, such as fertilisers, pesticides or water or changes in management practices-can increase GHG emissions, deteriorating soil quality, use scarce water and reduce biodiversity (Cassman, 1999;Johnson, Runge, Senauer, Foley, & Polasky, 2014;Newbold et al., 2015;Smith et al., 2013). Furthermore, land-based mitigation measures may be required to meet current climate change targets (Popp et al., 2017). Understanding how changes in climate, changes in demand for agricultural commodities and land-based climate change mitigation measures will affect the future agricultural and land use system is therefore critical.
Models of the global agricultural system have primarily taken economic equilibrium optimisation approaches, either general (CGE) or partial equilibrium models (Robinson et al., 2014). Due to computational restrictions, these approaches do not typically use high-spatial resolution when choices regarding rates of agricultural areas and intensities are made, instead representing the globe via a small number of regions or agricultural zones. Crop yields achieved with varying intensities of production are represented using different, but stylised approaches (Nelson et al., 2014). Increases or decreases in agricultural areas are also considered, but with increases specified at regional scales as part of the economic production functions, which are subsequently spatially disaggregated. This assumes that land expansion occurs on progressively less productive land but does not closely relate this expansion to physical properties and limitations.
Although downscaling or disaggregating into finer resolution maps is common, nonetheless the optimisation to determine the aggregate land uses within a region (including fertiliser and irrigation rates) has occurred with these aggregate units. An exception to this regional optimisation approach is MAgPIE, which takes a least-cost optimisation approach using gridded yield data from the global vegetation model LPJml (Lotze-Campen et al., 2008). However, even in this case, a location-specific yield response to agricultural input changes is not considered, but rather regional technological change rates are used (Lotze-Campen et al., 2008). Additionally, MAgPIE aggregates global spatial input data to between 100 and 600 clusters with similar crop yields (Dietrich, Popp, & Lotze-campen, 2013;Humpen€ oder et al., 2014;Kreidenweis et al., 2016). Therefore, current global land use models and IAMs do not explore the interactions between agricultural expansion and intensification using crop behaviour from plant-ecosystem process modelling on a spatially disaggregated basis. Furthermore, to date, there has been a lack of focus in global studies on understanding potential adaptation responses to climate change in land use (Berger & Troost, 2013). IAMs have been widely used to investigate land-based climate change mitigation options (Hum-pen€ oder et al., 2014;Popp et al., 2011Popp et al., , 2017Rose, 2014;Wise et al., 2009). While most IAMs represent "top-down" mitigation policies, making the "bottom-up" nature of the adaptation process more difficult to capture (Hertel & Lobell, 2014).
Here, we present initial results from a novel land use model that uses more detailed biophysically derived yield data and responses to inputs, with greater spatial specificity than previously possible. The approach couples a dynamic global vegetation model (LPJ-GUESS; Olin, Lindeskog, et al., 2015;Smith, W€ arlind, et al., 2014) with a new land use and food system model (PLUMv2). PLUMv2 responds to changes in input (yields and demand for commodities) by endogenously adapting land use at high-spatial resolution. Greater demand can be met both by intensification and agricultural expansion (Johnson et al., 2014;Tilman et al., 2011). A further novel aspect of this study is that PLUMv2 does not assume market equilibrium, commodity prices are adjusted to account for over-or undersupply, while trade mechanisms also allow for representation of international tariffs and transport costs. This offers a more accurate representation of the trade-offs, responses and cross-scale interactions that are likely to be important in determining the system dynamics as a whole (Rounsevell et al., 2014). Land use and demand projections from the coupled model system were evaluated against historical data to assess suitability for exploring future scenarios, a task often not conducted for land use models. These coupled models were used to investigate the potential for adaptation to climate change within the agricultural system and possible climate change impacts on land use.

| Overall coupled model framework
The Lund-Potsdam-Jena General Ecosystem Simulator (LPJ-GUESS; Smith, W€ arlind, et al., 2014)   . LPJ-GUESS produced crop and pasture yield potentials on a 0.5°grid-using a factorial experiment for crops with three fertilisation rates and rain-fed vs. irrigated conditions-using climate forcing scenarios.
PLUMv2 used these yields in combination with scenario data, for example GDP and population, to project land use and management inputs ( Figure 1) (Hickler et al., 2004;Smith, Prentice, & Sykes, 2001;Smith, W€ arlind, et al., 2014). Cropland and pasture were represented in LPJ-GUESS as fractions of land distinct from "natural" vegetation that undergo management and harvest (Lindeskog et al., 2013). Pasture is simulated as a natural grassland but with the addition of an annual grazing "harvest" term. Analogously to natural vegetation, the wide variety of crops planted around the world is simplified into several crop functional types (CFTs). Each CFT was assigned parameters related to plant physiology (e.g. photosynthetic pathway and vernalisation requirements) and management (e.g. fertilisation regime). The LPJ-GUESS crop model includes nitrogen cycling and has been shown to realistically simulate yield responses to nitrogen and CO 2 fertilisation (Olin, Schurgers, et al., 2015).
LPJ-GUESS was used with four CFTs: winter-sown C3 cereals (TeWW), spring-sown C3 cereals (TeSW), C4 cereals (TeCo) and rice (TrRi) (Olin, Lindeskog, et al., 2015). LPJ-GUESS input data and parameterisation details are available in the supporting information, along with information on changes made to crop water demand, soil moisture and irrigation. Potential yields under six alternative combinations of fertiliser and irrigation rates were determined. Three rates of fertilisation were considered: zero fertiliser, 200 and 1,000 kgN/ ha, with each either rain-fed or fully irrigated (i.e. with as much water applied as the plants would take up), with the potential heat units scheme for plant development (Olin, Lindeskog, et al., 2015).
The high-fertilisation rate is substantially beyond that used in practice, but represents a maximum upper limit of achievable yields. Economic considerations are accounted for in the land use optimisation and act to limit the fertiliser modelled as applied.

| Calibration to observed crop yields
PLUMv2 used seven crop types to represent demand for agricultural products, mapped on to the four LPJ-GUESS CFTs (Table 1), with the aim of maintaining realistic physiological and management parameters. A calibration routine was used to translate yields produced by LPJ-GUESS into potential yields for each PLUMv2 crop type, for example, TeSW to pulses. The calibration process was also used to improve the fidelity of LPJ-GUESS yields to observations for crops it was designed to simulate (e.g. TrRi to observed rice yields).
The calibration factors were generated via a slope-only regression between simulated and observed per-area yields for the years 1995-2005 (Table 1). Observed yields for each PLUMv2 crop type were derived from FAO data (FAOSTAT, 2015a(FAOSTAT, , 2015b, except for energy crops, data for which were taken from the Biofuel Ecophysiological Traits and Yields Database (BETYdb, LeBauer et al., 2010). Figure S2 shows the scatter plots comparing the simulated and observed yields for each PLUMv2 crop type, and Table 1 gives the derived calibration factors. The yields used by PLUMv2 were calculated as the product of the calibration factors and associated CFT yield output from LPJ-GUESS.

| Yield potentials in the land use model
The yields available for any combination of fertiliser and irrigation rate were estimated using the calibrated yield potentials at alternate irrigation and fertilisation rates for each grid cell and crop. An exponential yield function for all types of intensity was used that fits the LPJ-GUESS yield potentials provided (see supporting information-Methods for full equations). As well as fertiliser and irrigation rates, the level of management practices was represented by a "management intensity," encompassing activities such as pesticide application rates, reseeding of grassland, controlling of soil pH, for example, through application of lime, and larger stock of machinery or labour. ALEXANDER ET AL.

| 2793
An exponential approach was also used to represent diminishing returns from increasing management inputs. Yield increases from technology change, for example, due to plant breeding, were included as an annual exogenous increment to these yields.  alter as incomes change, with higher incomes being associated with a shift from staples such as starchy roots and pulses, to commodities such as meat, milk and refined sugars (Fiala, 2008;Kearney, 2010;Keyzer, Merbis, Pavel, & van Wesenbeeck, 2005;Tilman et al., 2011;Weinzettel et al., 2013). However, further increases in income tend to lead to lower increases in the rate of consumption (Cole & McCoskey, 2013), while the consumption of the less preferred product, for example, pulses, drops but at a decreasing rate. Both of these observations can be accounted for by the approach, similar to that applied by Tilman and Clark (2014) and Bodirsky et al. (2015); however here, the approach is applied to multiple commodity groups F I G U R E 1 Diagram of main interactions between LPJ-GUESS and PLUMv2 showing the components and flows within the couple models. Data passed between LPJ-GUESS and PLUMv2 are on a 0.5°grid. Water run-off is aggregated to food production units (FPUs), and adjusted for other uses, before being used to constrain water use in PLUMv2. Both models run at annual time steps, with LPJ-GUESS output data averaged over a 5-year period for input into PLUMv2 (see Figure S1 for temporal interaction details) [Colour figure can be viewed at wileyonlinelibrary.com] rather than to calorific intake and aggregate animal product consumption.
Cultural and other variations between countries lead to differences between (a) the consumption implied by the regression relationship from population and income and (b) the observed consumption in the same year. For example, Japan has less meat consumption than the global relationship suggests given its per capita income, but with a high level of fish consumption. The difference between expected and observed consumption rates for each country was calculated in the baseline year of 2010. Under some scenarios, these differences were held constant; under others, an exponential convergence was applied to global dietary patterns as per capita GDP increased (see supporting information-Methods). The historical and projected consumptions plotted against GDP are shown in Figure S3.
First-and second-generation bioenergy demand trajectories were specified exogenously to represent a moderate business-as-usual scenario. Bioenergy demand for food commodities-that is, firstgeneration bioenergy-were modelled from an observed baseline level of demand (Alexander et al., 2015;FAOSTAT, 2015c) adjusted to double by 2030 from the 2010 level and thereafter remain constant. Demand for dedicated energy crops (i.e. second-generation bioenergy) was specified as a global trajectory that increases to 4,000 Mt DM/year by 2100 from 34 Mt DM/year in 2010, in line with the SSP2 demand with baseline assumptions (Popp et al., 2017). Demand for second-generation bioenergy was not associated with individual countries, with all production locations determined endogenously.

| Country-level optimisation of land use, livestock production and international trade
For each country and time step, the agricultural land use and level of imports or exports were determined through a least-cost optimisation that meets the national demands for food commodities. For example, an increased national demand for a commodity can be met in three ways-increasing the land area for growing associated crops; increasing the levels of inputs to achieve higher yields, that is, intensification; or increasing the level of net imports, that is, reducing exports or increasing imports. The land use and intensities are spatial (0.5°grid), while the imports and exports rates are national.
Costs were associated with each aspect, using prices in 2010 US$.  F I G U R E 2 Example yield responses to fertiliser and irrigation inputs at 2010, for spring wheat (a) in Aberdeenshire, Scotland (lat: 57°, lon: À2.5°), and (b) maize in Texas, USA (lat: 30°, lon: À96°), at maximum management intensity regionally using the six World Bank regions (World Bank, 2014).
Monogastric livestock was considered to only consume feed, while ruminant livestock nutritional requirements could be met from a mix of pasture and feed, providing the opportunity for intensification by increased feed rates and a substitution between pasture and cropland.
Agricultural land use costs per unit area were calculated from a global base crop cost plus a cost of each of the three inputs considered, that is, fertilisation, irrigation and management intensity. The base costs are a minimum cost to producing that crop. The input costs were all products of the intensity rate and a cost rate. The base crop cost was estimated from a third of the cost per hectare in an intensive production system costs (Alexander & Moran, 2013;SAC Consulting, 2013), assuming reductions in inputs (e.g. seed rate, agrochemical and machinery use) could save cost, but with the implications for yields achieved (yield potentials in the land use model section above). To obtain the maximum possible yields-those output by LPJ-GUESS for a given fertiliser and irrigation rate-the remainder of costs associated with these current intensive production practices were included in the management intensity costs plus additional cost for higher agrochemical usage or machinery use. The base costs for pasture were assumed to be low, representing extensive grazing; intensive systems would include substantial management costs, for example, to represent reseeding to improve pasture yields. The crop costs parameters used are given in Table S2. An range of irrigation efficiencies globally . For each grid cell, the total irrigation water used across crops was constrained by water availability. Each year, the water available for irrigation was determined from the LPJ-GUESS-simulated runoff, assuming water consumption by sectors other than agriculture following Elliott et al. Future water consumption for non-agricultural sectors used SSP2 projections . The water remaining per FPU was allocated equally across the grid cells within each FPU to determine the irrigation water available.
Costs arising from changes in land cover-between natural and agricultural land or between cropland and pasture-were calculated per area converted (Table S3) . Terrestrial protected areas with a WDPA status of "established," specified on a 0.5°grid, were prevented from being converted to agricultural use. China's National Forest Protection Program was implemented as an annual limit to deforestation of 1.1% in these areas (Ren et al., 2015). A minimum natural area fraction was applied to preserve at least a proportion of forest or other natural land cover within each location, where protected areas did not meet this threshold. Expansion of agricultural areas was taken equally from forest and other natural vegetation. Urban and barren (e.g. ice-covered) land areas were constant from LUH2 in 2010 and not available for agricultural land expansion.
The final country-level cost relates to imported and exported agricultural commodities. Within the model, a single global market tariff-free price exists for each commodity and time period. The revenue received for exports was accounted for at this international market price. However, prices of imported commodities were inflated to account for transportation costs, losses during transportation and import trade tariffs (Anderson, Martin, & Valenzuela, 2006).

| Global trade balance and prices
In PLUMv2, as in reality, supply and demand in the global market for each commodity need not be in equilibrium, where over-or undersupply for commodities are buffered through stock variations (FAO-STAT, 2015c). The modelled international market prices for each commodity were adjusted exponentially using market conditions to provide a feedback mechanism (Ghoulmie, Cont, & Nadal, 2005). For example, where larger quantities of a commodity are exported globally than imported, the price for that commodity decreases; this reduces the benefits from its export and reduces the cost of importing it, creating a tendency to correct for the oversupply. The initial prices for each commodity were set exogenously (Index Mundi, 2016) but subsequently adjusted endogenously from the rate of under-or oversupply in the market (see supporting information).
Global stocks for each commodity accommodate periods of overand undersupply and were explicitly modelled. Initial stock levels were derived from FAO Commodity Balance data (FAOSTAT, 2015a, 2015c) following the method of Laio, Ridolfi, and D'Odorico (2016).

| Spin-up and spatial clustering in PLUMv2
PLUMv2 was initialised with GDPs, populations, net imports and demand from FAOSTAT (2015a, 2015c), and land covers from Land To test the demand projection approach, the FAOSTAT (2015a, 2015c) data were divided into a time series for calibration ) and a time series for benchmarking . The demand regression relationships were derived from the calibration data as and country-level demand (Figure 3). At 2010, the largest global percentage difference was seen in ruminants, a commodity group in which demand increased by 60% globally between 1990 and 2010, with the projections 15% higher than the FAO values (FAOSTAT, 2015a(FAOSTAT, , 2015c. Monogastric livestock was the only commodity with a larger growth, increasing by 78%, but here, the PLUMv2 projections were 1% lower globally in 2010 than the FAO value. This may indicate a shift in animal product preference from ruminant products to monogastrics between the time periods of the split data sets. Nonetheless, the modest level of these differences and the ability to reproduce the patterns of country and global changes in demand for the validation period suggest that the demand projection approach is adequate for the purposes of the land use modelling exercise being conducted. However, one limitation is that the approach assumes a continuation of the relationship between income and food demand and therefore does not account for potential future changes or transformation in food preferences (Alexander, Brown, et al., 2017;Stehfest et al., 2009).

| Land use benchmarking
The land use results were benchmarked by initialising the model at The impact of parameter uncertainty on the historical model results was tested using a stochastic approach. Uniform distributions of model parameters (Table S3)  Given the agreement between agricultural expansion and intensity with the historical estimates as well as the concurrence of spatial distributions, the modelling was considered appropriate for exploration of future scenarios.

| Scenario descriptions
The aim of the scenario design was to explore the adaptation of the   scales (Bala et al., 2007;Gibson et al., 2011;Malhi et al., 2008). Possible future policies for avoiding deforestation or existing policies that provide a level of non-spatially specific protection are not contained in this protected area database, and therefore are not included in these results. For example, economic approaches that incentivise a reduction in deforestation  were not represented. If such polices were included, deforestation and conversion into pasture or cropland may be reduced in the associated areas. If land use change was to be avoided in this way, other consequences in the model would arise through the indirect effects caused by the displacement of production from areas no longer entering agricultural use (Popp, Humpenoeder, et al., 2014). The indirect effects could include the expansion of cropland and pasture in other less protected areas, increases in intensity on existing agricultural land or most likely a combination of both. Similar adjustments in results could occur if the cost of conversion from forest to agricultural land cover were increased, with a greater conversion of other land covers to agriculture combined with higher intensity of production. Greater competition for land could also arise from climate mitigation policies, for example, supporting bioenergy use or afforestation to provide a terrestrial carbon sink (Albanito et al., 2016;Popp, Rose, et al., 2014). Investigating land use outcomes and displacement effects under climate mitigation policies were out of scope for this study, but could be addressed in future work. agricultural land may be more constrained and costly than during the historic period. Further scenario development, for example, using the SSP framework, and uncertainty analysis would help understand this more fully, but was out of scope for the analysis presented here.
Our results assume technology change in plant breeding, which provides an annual increase in yield above that which can be achieved by increasing intensification (the central parameter value used was 0.2%, Table S3). Higher rates of technology improvements-which could be achieved by, for example, the introduction of genetically modified or gene edited organisms-would reduce the expansion of agricultural land or inputs. Conversely, if technology improvements were not able to achieve these gains, then more land and other inputs to agricultural production would result.

| Limitations of the approach
PLUMv2 is not constrained to reproduce initial land covers used in the calibration process. Imposing such a constraint could lead to rapid changes in initial simulation years. Therefore, the approach of finding a stable state in proximity to a calibration data set, comprising land covers as well as national production, consumption and international trade data, was preferred. No data on fertiliser or irrigation use were provided to the model calibration, in part due to a lack of suitable available data, and therefore, the initial fertiliser and irrigation rates were derived endogenously during the calibration process. The PLUMv2 simulation calibration outcomes in 2010 are close to the historical estimates ( Figure 6), making this potential difference of minor importance in the future simulation results. There are greater differences in the benchmarking runs starting in 1970 for cropland and irrigation water use (Figure 4). For example, cropland in 1970 was 65 Mha (5%) lower, and irrigation water use 300 km 3 (20%) lower, than historical estimates (FAOSTAT, 2017;IFA, 2017), although the high uncertainty in these estimates complicates any benchmarking. Although the benchmarking process produced a reasonable fit to observed aggregate global outcomes and land cover distribution from LUH2, discrepancies were noted. The explanations suggested above for these differences-for example, influence on land use change in proximity to existing infrastructure, imperfect protected area enforcement, and effects of bilateral trade agreements between countries-could be implemented to test the outcome from altering these assumptions.
The demand projections assume a continuation of historical income-demand relationships and thus do not consider possible alterations in dietary preferences, for example, towards lower meat consumption for both health and sustainability reasons (Stehfest et al., 2009). Furthermore, there was no price elasticity of demand, and so the types and quantities of commodities demanded do not alter in response to price changes, but only population and per capita incomes. Given the objective to investigate adaptation in response to alternative climate futures, we believe such assumptions are acceptable. However, to investigate other scenarios, for example, which include dietary trend adjustments, other assumptions and approaches would be required.
Soil degradation-including from erosion, compaction, sealing and salinisation -was not included in the modelling conducted. Agricultural land lost to degradation between 2000 and 2030 was projected to be 30-87 Mha (Lambin & Meyfroidt, 2011), with 7.5% of grassland degraded because of overgrazing (Conant, 2012), while erosion degradation can lead to compensatory benefit at the site of deposition (Lal, 2001). Changes in soil pH resulting from excessive nitrogen fertilisation were also not considered. Continued land degradation increases the pressure on land, but is perhaps smaller in magnitude than other drivers considered, for example, socio-economic and climate changes. Nonetheless, it would be advantageous to include the effect of soil degradation within models such as LPJ-GUESS and PLUMv2.

| FUTURE RESEARCH
This study applied newly coupled models to study the response to climate changes for a single fixed socio-economic scenario. Further work is required to explore the response to alternative socio-economic conditions, for example, using the SSPs, and to a range of potential climate change mitigation measures, for example, bioenergy and measures to reduce deforestation and increase afforestation.
There are also important aspects of crop response to climate change, such as heat stress and CO 2 fertilisation, which are currently the subject of high uncertainty and merit further investigation. A key aim of the coupled LPJ-GUESS and PLUMv2 modelled framework was to allow the feedback for land use change on climate as well as the climate impacts on land use, to be considered. Further work is planned to continue model development and to integrate these feedbacks, using a climate emulator (IMOGEN; Huntingford et al., 2010), to study the response in a fully couple climate, vegetation and land use modelled system.
The results suggest that the global agriculture and food system has the capacity to potentially diminish the negative impacts and take greater advantage of the more positive outcomes of climate change through adaptation, for example, by changing crop types, management practices or shifting cultivated area. These adaptations are spatially specific, given geographic variability in climate change impacts on agricultural production. Outputs from models projecting future land uses without accounting for detailed spatial-, crop-and input-specific factors may therefore be biased towards overestimating land use impacts under a changing climate. To quantify this potential bias, further work is required to establish the extent modelled land adaptation is affected by the level of detail in the representation of spatial and input factors. The results found here suggest that increased intensity of climate forcing reduces the inputs required for food production, largely due to the fertilising and enhanced water use efficiency effects of elevated atmospheric CO 2 concentrations. However, achieving this requires substantial shifts in the global patterns of intensity of production, with greater inputs required in Africa and South America, and reductions in North America and Western Europe. Such changes in land use and management intensity have consequences for other ecosystem services, and thus, the apparent resilience in the food system indicated by this study may lead to degradation of other ecosystems.

ACKNOWLEDG EMENTS
The research was supported by the UK's Global Food Security Programme project Resilience of the UK food system to Global Shocks