Land use plays a major role in determining terrestrial–atmosphere mass and energy exchange (Adegoke et al., 2007; Raddatz, 2007), which in turn influences local to global climate (Brovkin et al., 2013; A. D. Jones et al., 2013; Pitman et al., 2009). Despite much recent progress, we still have a limited understanding of how historical land use has affected, and continues to affect, climate (Brovkin et al., 2013; A. D. Jones et al., 2013; Pitman et al., 2009) and carbon (Anav et al., 2013; Arora and Boer, 2010; Houghton, 2010; Houghton et al., 2012; Hurtt et al., 2006; Jain et al., 2013; Jain and Yang, 2005; C. Jones et al., 2013; Smith and Rothwell, 2013), and high uncertainty as to how land use might evolve in the future (Hurtt et al., 2011; van Vuuren et al., 2011a; Wise et al., 2009). Part of the uncertainty in future land use trajectories is due to inherent unpredictability of human actions, and part to the high diversity of potential climate mitigation and adaptation scenarios. Several energy and land strategies have been proposed to mitigate climate change (Rose et al., 2012; P. Smith et al., 2013), and while these have similar overall goals, some strategies will likely compete for land and other resources if implemented simultaneously. For example, afforestation and bioenergy production both aim to reduce atmospheric CO2 concentrations, but both activities require land area, and both strategies would impact crop production and markets through effects on crop area (Reilly et al., 2012).

Reflecting this limited understanding of land use effects on climate and carbon, global climate models (GCMs), and also next generation earth system models (ESMs) that include fully coupled atmosphere–land–ocean carbon cycles, implement a wide range of land use/cover approaches with varying degrees of detail and limited inclusion of managed ecosystems and land use practices (Brovkin et al., 2013; Pitman et al., 2009). The Land Use and Climate, IDentification of robust impacts (LUCID) activity employed seven GCMs to determine whether land use change has significant regional climate impacts and farther-reaching teleconnections due to biophysical changes in land surface. The results for 1972–2002 revealed significant but inconsistent changes in temperature, precipitation, and latent heat in some areas where land use change had occurred. The authors concluded that the model disagreement was due mainly to differences in land use and land cover change implementations and corresponding land cover distributions, with contributions from methodological differences in crop phenology, albedo, and evapotranspiration (Pitman et al., 2009). The environmental factors addressed by LUCID are also key factors for determining carbon uptake by vegetation, and thus it is not surprising that the Coupled Climate–Carbon Cycle Model Intercomparison Project (C4MIP) activity generated ESM projections that range from the land being a carbon source to a large carbon sink by 2100 (Friedlingstein et al., 2006).

To advance the scientific understanding of the effects of land use change on climate, phase 5 of the Coupled Model Intercomparison Project (CMIP5) (Taylor et al., 2012) applied a novel “land use harmonization” approach to produce the required land use change information for all participating GCMs and ESMs. The Global Land use Model (GLM) was used for this land use harmonization to generate the first set of continuous, spatially gridded land use change scenarios for the years 1500–2100 (Hurtt et al., 2011). GLM computes land use states and transitions annually at half-degree, fractional spatial resolution, including secondary land age, area, and biomass, and the spatial patterns of shifting cultivation and wood harvesting (Hurtt et al., 2006). Land use products from GLM have successfully been used as inputs to both regional and global dynamic land models (Baidya Roy et al., 2003; Hurtt et al., 2002; Shevliakova et al., 2009) and fully coupled ESMs (Jones et al., 2011; Shevliakova et al., 2013). The land use harmonization process ensures a continuous transition from the historical reconstructions to the future projections made by integrated assessment models (IAMs).

The land use harmonization methodology was designed to satisfy the demands of a broad range of models and to provide a consistent set of land use inputs for GCMs and ESMs. The historical period of the land use harmonization (1500–2005) was based on version 3.1 of the Historical Database of the Environment (HYDE; Klein Goldewijk et al., 2011) and Food and Agriculture Organization (FAO) wood harvest data. For the future period (2005–2100), the land use harmonization process utilized land use data from the four representative concentration pathways (RCPs), each provided by a different IAM. The RCP scenarios were designed to each meet a different radiative forcing target (2.6, 4.5, 6.0, and 8.5 W m-2), and due to differences among the IAMs these scenarios spanned a range of approaches in all sectors, including land use, for meeting the targets (van Vuuren et al., 2011a). As a result, forest cover change varied widely from deforestation to afforestation across the scenarios. Once the land use data were passed through the land use harmonization, each GCM/ESM utilized a unique subset of the harmonized outputs, based on model capabilities, and applied it to a unique set of land use and land cover types (e.g., Lawrence et al., 2012). Although this process was largely successful in enabling the first spatially explicit land use driven climate change experiments, it introduced considerable uncertainty into the climate response for a given RCP in part because of model-specific translation requirements between harmonized land use outputs and GCM/ESM simulated land cover. This uncertainty due to inconsistent land cover distributions among models precluded robust intercomparison of land–atmosphere processes (e.g., carbon uptake, evapotranspiration) because differences among models were dominated by the differences among simulated land cover distributions (Brovkin et al., 2013). As land use and land cover are interdependent, a more detailed specification of the relationship between land use and land cover may reduce uncertainty in earth system simulations such that experiments can focus on land–atmosphere process uncertainty rather than be confounded by inconsistent land use/cover distributions.

Recent analyses of CMIP5 results using prescribed CO2 concentrations have also shown the land ranging from a carbon source to a sink in 2100 for a given scenario (Brovkin et al., 2013; C. Jones et al., 2013). The LUCID activity was repeated for five CMIP5 ESMs and the results demonstrated that large inter-model spreads of key regional land surface variables (temperature, precipitation, albedo, latent heat, and available energy) were still due mainly to differences in land use and land cover change implementations and corresponding land cover distributions. Inter-model spreads of CO2 emissions, however, were attributed mainly to differences in land carbon cycle process parameterizations. As a result, different land cover distributions among the models gave significantly different regional changes in climate associated with land use change, but with insignificant effects on global mean temperature. Furthermore, the range of net cumulative land use change emissions from 2006 to 2100 for RCP8.5 was 34 to 205 PgC, with the high estimate likely due to the combination of relatively high levels of land carbon and the inclusion of all land use transitions rather than just net land use change (Brovkin et al., 2013). Additionally, not all of the models used the GLM wood harvest data, further contributing to the spread of model results. For comparison, estimates of net cumulative carbon emissions during 1700–2000 (1850–2000) range from 138 to 250 PgC (110 to 210 PgC) (Table 3 in Smith and Rothwell, 2013). The differences in land use and land cover implementations are also a main factor in the large spread of 21st century land carbon uptake and of compatible fossil fuel emissions allowable for a given RCP. In fact, the inter-model spreads in land carbon uptake for individual scenarios are greater than the inter-scenario spreads for individual models (C. Jones et al., 2013). It is apparent that further work is needed to resolve inconsistencies among land use and land cover approaches to reduce climate uncertainty, especially for regional impact assessment.

Additional sources of climate uncertainty related to land use are the RCP radiative forcing targets, which include only emissions of greenhouse gases (GHGs) and some aerosols and reactive gases (van Vuuren et al., 2011a). These targets do not include radiative forcing from albedo change or other direct climate effects associated with land use change. In a recent modeling experiment, two different carbon tax policies with dramatically different land use scenarios met the same radiative forcing target (4.5 W m-2) in the IAM used for RCP4.5 but had significantly different radiative forcing in an ESM (difference of 1 W m-2) due to albedo differences between the land use scenarios (A. D. Jones et al., 2013). Likewise, the shared socioeconomic pathways (SSPs) for mitigation, adaptation, and impact studies in the Intergovernmental Panel on Climate Change (IPCC) fifth Assessment Report (AR5) are likely to produce different land use scenarios that meet the same RCP target, but have different radiative forcing in the ESMs due to the direct effects of land use and land cover change on climate. However, one of the goals of the RCP process was to provide a set of radiative forcing targets for ESMs that remains consistent with respect to the diversity of SSPs associated with each RCP target (Moss, et al., 2010). As a result of the wide range of land use and land cover related uncertainties in climate projections, an increased emphasis on land use and land cover dynamics is a high priority for CMIP6 (Meehl et al., 2014).

A more consistent and complete land use and land cover coupling between IAMs and ESMs will facilitate more accurate projections of global change scenarios and more robust multi-model intercomparisons of climate and carbon cycle interactions with anthropogenic drivers such as fossil fuel emissions and land use change. These expected outcomes are in line with a primary goal of a scenario-based approach, such as the RCPs, which is “to better understand uncertainties in order to reach decisions that are robust under a wide range of possible futures” (Moss et al., 2010; p. 747). The RCPs were designed to better understand uncertainties in global climate projections by providing distinct scenarios of atmospheric radiative forcing and land use change. Intra-scenario comparison of ESM simulations offers insights to uncertainties in ESM processes, while inter-scenario comparison of ESM simulations offers insights to uncertainties due to a range of possible futures. However, the efficacy of this approach depends on the fidelity of the ESM simulations to the RCP scenarios. Without this fidelity, intra-scenario comparison is not possible, because the ESMs are not simulating the same scenario, and inter-scenario comparison might include futures outside the prescribed range of possibility.

The IAMs projected a complete terrestrial surface (along with ice, rock, and urban) for each given scenario because land use and land cover are interdependent. For example, carbon stocks in various ecosystems might be valued under a carbon price policy, so land cover would need to be determined along with land use. Or a land policy might restrict certain land cover conversions. Within the CMIP5 coupling process, however, GCMs and ESMs determine their own land cover while remaining consistent with the land use harmonization data, thus potentially reducing the fidelity of the full climate simulations to the RCP scenarios. This was a practical design that obviated the redesign of GCM/ESM land use and land cover implementations, but also precluded analysis of the climate impacts of different land cover responses to land use change because such analysis is robust only within a single model where everything but land cover response remains consistent. Another challenge posed by the interdependence of land use and land cover is the implementation of geographic shifts in land cover due to bioclimatic changes. While these shifts are often implemented within ESMs, such shifts are a second-order effect that is superposed upon land use change and might be better implemented as a feedback from ESMs to IAMs to inform land use and land cover projection. Incorporating both land use and land cover into the coupling between IAMs and ESMs is a fundamental step toward realizing the full potential of the scenario-based RCP process.

Our approach to addressing inconsistencies between IAMs and ESMs is to integrate an IAM and an ESM into the first fully coupled model that directly simulates human–environment feedbacks. The resulting integrated ESM (iESM) includes climate feedbacks on vegetation productivity and ecosystem carbon from the Community ESM (CESM) to the Global Change Assessment Model (GCAM) to facilitate land use projection at 5-year intervals. The iESM uses GLM as in the CMIP5 land use harmonization, along with the CESM Land Use Translator (LUT) that converts land use harmonization outputs to CESM land cover and wood harvest area. Our initial iESM simulations showed that time varying factors based on CESM simulated net primary production (NPP) and heterotrophic respiration (HR) were successfully used by GCAM for land use projection. However, these simulations also demonstrated that the large RCP4.5 afforestation signal was not being passed through from GCAM to CESM. GCAM simulated afforestation as a carbon-sequestering strategy to help meet the RCP4.5 target, but this additional forest area was not included in the land use harmonization. As a result, most of this forest area was not included in CESM simulations, both for CMIP5 and in an early version of iESM.

Here we test the feasibility of restoring the lost afforestation signal by using the iESM as a test bed to explore alternative coupling strategies. We focus on modifications to the CESM LUT because initial modifications to GLM did not restore CESM afforestation. One advantage of focusing on a post-land use harmonization approach is that it could be applied to other ESMs independently without changing the land use harmonization product. Section 2 includes model description and experimental design, Sect. 3 presents results and demonstrates that this problem exists in CMIP5, and Sect. 4 discusses the limitations of our current approach and the implications for the CMIP5 archive with respect to land use and climate. We conclude with suggestions for improving IAM to ESM land coupling for future model intercomparisons.

The iESM and CMIP5 land cover area discrepancies (Figs. 5–7) result from a gap in the original CMIP5 land coupling design that allows inconsistent forest area and land cover type definitions across models (Fig. 2), along with different underlying carbon cycles. The land use harmonization was, however, ambitious and largely successful in developing consistent land use definitions and data without requiring extensive redevelopment of land use and land cover components of all participant models (Hurtt et al., 2011). As our study attests, such redevelopment is challenging and model-specific, but might be required for ESMs to adequately simulate the IAM-prescribed anthropogenic drivers and their corresponding effects on carbon and climate. Thus, while this is a specific case, the lost iESM afforestation signal is instructive of the shortcomings of the CMIP5 design and the restoration of this signal offers insights into improving land use and land cover coupling for model intercomparisons.

A primary challenge for improving the CMIP5 land coupling is to increase the amount of specific land cover information being shared between IAM (and historical) scenarios and ESMs. For CMIP5, the land use harmonization was designed to harmonize land use data between models, and as such GLM did not receive forest area or any other land cover information from any of the IAMs (Masui et al., 2011; Riahi et al., 2011; Thomson et al., 2011; van Vuuren et al., 2011b). Thus, at the first coupling step, scenario-prescribed land cover associated with any IAM policy that valued carbon within unmanaged ecosystems (e.g., grassland, wetland, forest) was lost. While GLM does, however, keep track internally of forested and non-forested land (according to its own definition of forest, which likely differs from those within IAMs and ESMs), the output land use harmonization product includes only cropland, pasture, primary, and secondary land areas and transitions, and the age and biomass density of secondary land (and harvest areas, carbon amounts, and transitions, which we do not address here). As each ESM characterizes the land surface by its own suite of vegetation and management types (Brovkin et al., 2013), additional land use and land cover information could be lost in the second coupling step between GLM and the ESMs. For example, some ESMs were able to use the primary, secondary, and transition information, but they might have been applying this information to different land covers than those used by GLM, thus introducing a second shift away from the original IAM scenario. Our specific case demonstrates an even greater inconsistency due to the use of only cropland and pasture information. GCAM has 17 crop types (the CMIP5 version had 10) and seven managed and unmanaged land cover types while CESM has 16 PFTs, only one of which is a crop type. The LUT algorithm uses only the GLM cropland and pasture area information to adjust PFTs because CLM does not keep track of primary versus secondary land. The resulting spatial pattern of non-crop PFTs is determined by the existing PFT distribution and CESM's internal representation of potential vegetation cover (Lawrence et al., 2012; Ramankutty and Foley, 1999). An additional source of error that we did not investigate here is the relationship between individual PFTs and land cover types that may comprise several PFTs (e.g., forest land may consist of 60 % trees and 40 % grass).

Due to the lack of a prescribed land cover input associated with the land use input, forest area changes in CESM (and iESM) are effectively residual changes that are only indirectly linked to GCAM forest area through changes in cropland and pasture areas. The LUT calculates cropland area changes first and pasture area changes second (Figs. 3 and 4). In CMIP5 CESM simulations, cropland area changes cause non-crop PFTs to be added or removed in proportion to their potential or existing grid-cell fractions, respectively. Pasture is more complicated because it is not tracked as such: pasture is not a single PFT and its changes are represented as changes in herbaceous and tree PFTs. Specifically, tree PFTs are removed when pasture is added, and non-crop PFTs are added in proportion to their potential vegetation grid-cell fractions when pasture is removed (Lawrence et al., 2012). This residual PFT determination, combined with independent and unique forest definitions across GCAM, GLM, and CESM, causes the bulk of prescribed afforestation to not appear in the CESM land surface. As a direct consequence, CESM grass area (and shrub area to a lesser extent) increases while GCAM pasture decreases dramatically (Fig. 5). CESM has this same limitation for all four RCP scenarios, and the other CMIP5 ESMs implement similar inconsistencies to varying degrees due to the lack of specific vegetation types in the land coupling between IAMs and ESMs. For example, Davies-Barnard et al. (2014) recently reported that the HadGEM2-ES RCP4.5 forest area increased 11 % from 2005 to 2100, while the GCAM forest area increased by 24 %. Additionally, the GCAM 2005 forest area was 41.1 M km2, the GLM 2005 forest area was 39.9 km2, but the MPI-ESM 2005 forest area was about 24 M km2. As a result, the 35 % increase in MPI-ESM RCP4.5 forest area by 2100 (Wilkenskjeld et al., 2014) was still only 77 % of GCAM's afforestation. It is apparent from these inconsistencies that interdependent land use and land cover need to be faithfully transmitted from IAMs to ESMs to robustly simulate the effects of prescribed scenarios on the earth system.

Even partial restoration of the lost afforestation signal in iESM demonstrates the potentially dramatic effect on global carbon and climate of using IAM land cover and land use information in ESMs. As soon as 25 years after the initial increase in forest area, and with only 66 % of GCAM's afforestation area, the NEWLUT has a significant impact on global carbon balance (Fig. 9). The assumption that forest exclusively replaces abandoned cropland and pasture in GCAM's land use projection (Figs. 6–8) sets the upper limit for CESM because there is no other information to constrain forest area, and may be applicable only to the RCP4.5 scenario. Although this limits NEWLUT to including only two-thirds of the total afforestation, adding more forest area to CESM would be arbitrary without additional land cover information. Nonetheless, the increased afforestation in NEWLUT results in an increase in net land carbon uptake over the OLDLUT case due to a sustained increase in average annual land carbon uptake after 2020 (Fig. 9). As a result, the NEWLUT simulation increases vegetation carbon gain by 19 PgC and decreases atmospheric CO2 gain by 7.7 ppmv from 2005 to 2040 in comparison to OLDLUT (Fig. 10). The NEWLUT simulation also decreases soil carbon gain by about 1.5 PgC over this period (data not shown).

Simple linear extrapolation of the iESM vegetation carbon gain and atmospheric CO2 gain from 2005 to 2100 increases these changes to approximately 52 PgC and 21 ppmv, and extending CESM forest area to match GCAM total afforestation could potentially increase these changes to 88 PgC and 36 ppmv in 2100. These are rough estimates that use 2005 as a starting point to reduce the high slope associated with the initial increase from 2015 to 2020, and also assume that additional forest area continues to gain carbon for 60–80 years after it is established. Regardless of the absolute accuracy of these extrapolations, the potential gain in vegetation carbon alone for CESM with full afforestation is on the order of estimates of net cumulative land use change emissions during 1850–2000, which range from 110 to 210 PgC (Table 3 in Smith and Rothwell, 2013). For comparison, the range of CMIP5 vegetation carbon stock gains for RCP4.5 is about 50 to 300 PgC from 2005 to 2100, with most gains being less than 150 PgC and relatively linear (Fig. 2 in C. Jones et al., 2013). An increase in gain of 88 PgC would dramatically shift CESM vegetation carbon dynamics in relation to the other ESMs. The corresponding 36 ppmv decrease in atmospheric CO2 is nearly one-third of the difference between the prescribed 2100 concentrations of the RCP4.5 (∼ 540 ppmv) and RCP2.6 (∼ 420 ppmv) scenarios (Fig. 1 in C. Jones et al., 2013). More importantly for CESM's ability to robustly simulate the effects of the RCP scenarios on the earth system, the prognostic CESM atmospheric CO2 concentration in 2100 for RCP4.5 is 610 ppmv (Keppel-Aleks et al., 2013), and a decrease from 610 to 574 ppmv has an approximate decrease in radiative forcing of 0.33 W m-2, which is non-trivial with respect to the 4.5 W m-2 target. While these carbon cycle changes in the CESM component of iESM may have a significant effect on climate, it is important to note that the carbon cycle effects of afforestation in CESM are not identical to those in GCAM or GLM because these three models have different biogeochemistry and vegetation models. These differences in carbon cycles, however, do not obviate the need for making both land cover and land use consistent between IAMs and ESMs in order to best match the prescribed radiative forcing scenario.

Different implementations of land cover and land use among IAMs and ESMs also reduce the fidelity between RCP scenarios and their associated effects on the earth system. Figure 8 shows that most of the additional forest area in NEWLUT occurs on grassland and shrubland, and that these lands generally coincide with areas of limited potential forest. The OLDLUT could not add forest area where no potential forest area exists, and the rate of forest carbon accumulation is constrained by environmental conditions. GLM also limits forest area and growth based on potential forest and environmental conditions, but with a different growth model and map of potential forest area than used by CESM. On the other hand, GCAM afforestation is a strategy to expand forest area for carbon sequestration, and assumes that it is cost effective to use agricultural inputs (e.g., water, fertilizer) to achieve the expected forest growth. This disagreement among the three models hampers communication of forest area changes and contributes to the differences in forest area among the models, both in CMIP5 (Fig. 5) and in the iESM (Figs. 6 and 7). Nonetheless, sharing forest area between GCAM and GLM does improve the fidelity between GCAM and GLM's forest area changes (Figs. 5 and 6). GLM and CESM do not simulate agricultural inputs for forests, yet the NEWLUT can simulate most, but not all, of the prescribed afforestation (Figs. 6 and 7) by adding forest area based on GCAM's cropland and pasture changes, rather than on potential forest area. The additional forest might not grow as well in CESM as in GCAM, but the CESM forest productivity is fed back to GCAM for subsequent land use projections, so environmental restrictions on forest growth will influence future land use and land cover. This feedback does not, however, fully compensate for the lack of bioclimatic or agricultural input availability constraints on GCAM's land use projection, which might contribute to an overly optimistic afforestation projection. More generally, this feedback mechanism opens a path for more robustly simulating interdependent land use and land cover through incorporation of potential bioclimate-driven geographic shifts in land cover. ESMs could estimate bioclimatic drivers or geographic shifts for given land use/cover scenarios, and then feed this information back to the IAMs for incorporation into land use/cover projection. Implementing such a feedback for scenario-based simulations would consolidate land use/cover determination into internally consistent modules within the IAMs, thereby increasing fidelity between the scenario-prescribed land surface and the one used by the ESMs.

We have focused on understanding the effects of mismatched land cover areas on global simulations, rather than on mismatched carbon cycles, because the spatial distribution of land cover and land use is a scenario-determined boundary condition for ecosystem-specific processes such as biogeochemical dynamics. For global simulations this boundary condition is generally provided by historical data and IAMs, and, as we have shown, a mismatch in this boundary condition causes CESM to simulate non-scenario effects on carbon and climate (due to a non-scenario land surface), rather than the scenario-driven effects of the land surface prescribed for meeting the RCP4.5 target. Mismatched carbon cycles among IAMs and ESMs, on the other hand, along with differences in atmospheric radiation code, will preclude exact matches in radiative forcing for a given RCP scenario, but should not cause significant deviations among models in the carbon and climate effects of a given scenario. While we plan to completely reconcile land use and land cover inconsistencies within the iESM by implementing a single carbon cycle with consistent land surface characterization among the components, it is not desirable, nor feasible, for all IAMs and ESMs to have the same biogeochemistry and vegetation growth components. For example, a diversity of terrestrial models can help characterize uncertainty in global simulations. This uncertainty, however, is most useful if these models simulate the same spatial distribution of land cover and land use change. Therefore, iESM redevelopment that ensures land use and land cover consistency between GCAM and CESM could provide a template for improving the fidelity between IAM scenarios and ESM simulations in the next CMIP. In fact, land cover information is currently planned to be included in the CMIP6 land coupling, along with a more extensive land use model intercomparison project (Meehl et al., 2014).

We have identified the lack of specific land cover type information being shared among GCAM, GLM, and CESM in the iESM as the primary cause of CESM having very little afforestation and effectively no change in herbaceous PFT area in contrast to GCAM's large RCP4.5 afforestation and corresponding pasture reduction. Initial efforts to fix this problem through GLM modifications and the sharing of forest area between GCAM and GLM improved only the fidelity of forest area changes between GCAM and GLM. We then focused on modifying the algorithm that translates GLM land use harmonization outputs to CESM PFTs. While these land use translator modifications have been successful at capturing two-thirds of GCAM's RCP4.5 afforestation signal and corresponding reductions in herbaceous PFT area, they are not sufficient to completely overcome the limitations imposed by not passing specific land cover types from GCAM through to CESM. These modifications are also specific to the GCAM RCP4.5 scenario, and might need to be altered for the other RCP scenarios. Furthermore, we have not addressed the lack of constraints on GCAM forest area expansion, nor mismatches between land cover and PFT definitions. Nonetheless, this partial restoration of afforestation has a significant impact on iESM's global carbon cycle through increased vegetation carbon and decreased atmospheric CO2 concentration.

The iESM framework follows the CMIP5 land coupling design, and as such we have characterized a major gap in this design that precludes accurate translation of projected IAM land surface scenarios to ESMs by focusing only on land use such as cropland and pasture (albeit successfully), and not including specific land cover types such as forest, grassland, and shrubland. The relationship between land use and land cover is handled uniquely by individual ESMs, which means that the effects of scenario mismatch will be model-specific and more relevant for some RCPs than others. The resulting land cover discrepancies are likely most pronounced for the large RCP4.5 afforestation signal, which was greatly reduced in the CMIP5 CESM and HadGEM2-ES (see Davies-Barnard et al., 2014) simulations, but could also arise for other large land cover changes such as the extensive deforestation of RCP8.5. As total land area is conservative, errors in the distribution of one land cover are complemented by errors in the distributions of other land covers. In GCAM's RCP4.5 scenario, pasture decreases over the 21st century, but the CMIP5 CESM runs have increasing grass and shrub areas over the same period. It is very important that the land use and land cover changes (which determine land use change emissions and the total capacity for vegetation carbon assimilation) match between the IAMs and ESMs because the CMIP5 experimental design is predicated on the fidelity between IAM scenarios and ESM simulations such that they have similar, specific radiative forcings for a given scenario, including CO2 emissions from land use change (Moss et al., 2010). Furthermore, future radiative climate targets are likely to include the biogeophysical forcings of land use change because it has been shown that the modeled climate system is sensitive to changes in these forcings due to the spatial distribution of land use and land cover change (Brovkin et al., 2013; A. D. Jones et al., 2013; Pitman et al., 2009), making it imperative that IAM and ESM land use and land cover distributions match as closely as possible. Maintaining the diversity of global biogeochemical and vegetation models also calls for GCMs and ESMs to match historical and projected land cover and land use distributions as closely as possible, so as to isolate carbon cycle contributions to uncertainty from contributions due to differences in land use and land cover. Fortunately, our results indicate that it might be possible to adjust land cover in other CMIP5 models to better match RCP4.5 afforestation and the corresponding climate scenario, while still using the standard land use harmonization data.

We conclude that the land coupling between IAMs and ESMs for future model intercomparisons needs to ensure greater consistency in land cover and land use among the models in order to realize the full potential of scenario-based earth system simulations. In short, the models need to agree on the actual land area and the annual spatial distribution of major (non-)vegetation land covers and land uses. In other words, the ESMs need to simulate the same basic land surface as prescribed by the IAM-generated RCP scenarios. To achieve the required consistency, we suggest that the next CMIP land coupling design provides land cover and land use information, and a standard mapping between land cover and plant functional types. Fortunately, this is an emerging priority for the CMIP6 Land Use Model Intercomparison Project (LUMIP, http://www.wcrp-climate.org/index.php/modelling-wgcm-mip-catalogue/modelling-wgcm-mips/318-modelling-wgcm-catalogue-lumip, http://www.wcrp-climate.org/wgcm/WGCM17/LUMIP_proposal_v4.pdf). The following gridded data with fractional shares within grid cells are specifically recommended:

Annual land cover states with complete, contiguous spatial coverage within grid cells. Land cover needs to include at least the basic categories of cropland, grassland, shrubland, woodland, forest, and other (bare/sparse, ice, urban, water). This will allow consistency in major (non-)vegetation types for model intercomparison (with the “other” category having fixed area). The “other” categories could also be separated out for models that can use them, and in preparation for changing their areas also.