The mutual importance of anthropogenically and climate‐induced changes in global vegetation cover for future land carbon emissions in the MPI‐ESM CMIP5 simulations

Based on the Max Planck Institute Earth System Model simulations for the Coupled Model Intercomparison Project Phase 5 and on simulations with the submodel Cbalone we disentangle the influence of natural and anthropogenic vegetation changes on land carbon emissions for the years 1850 until 2300. According to our simulations, climate‐induced changes in distribution and productivity of natural vegetation strongly mitigates future carbon (C) emissions from anthropogenic land‐use and land cover change (LULCC). Depending on the assumed scenario, the accumulated carbon emissions until the year 2100 are reduced between 22 and 49% and until 2300 between 45 and 261%. The carbon storage due to climate‐induced vegetation change is generally stronger under the presence of LULCC. This is because the natural vegetation change can reestablish highly productive extratropical forests that are lost due to LULCC. After stopping anthropogenic vegetation changes in the year 2100 the refilling of depleted C pools on formerly transformed land takes (dependent on the scenario) time scales of centuries.


Introduction
Observations exhibit rising global temperatures since the late 19th century [Trenberth et al., 2007]. Climate model projections agree that this temperature rise will continue in the future [Meehl et al., 2007]. One of the most important drivers for this warming is the human-induced growth of atmospheric CO 2 concentrations [Forster et al., 2007;van Vuuren et al., 2011]. In times of increasing industrial CO 2 emissions, atmospheric CO 2 rise depends approximately equally on the oceanic and terrestrial sink capacity [Prentice et al., 2001;Le Quéré et al., 2009;Denman et al., 2007;Sabine et al., 2004]. The ocean carbon buffer capacity is expected to decline over time [Ciais et al., 2013] and the landmasses could turn into a source through changes in vegetation cover and accelerated biochemical processes in soils under the expected warmer conditions [Cox et al., 2000;Jones et al., 2013].
Little is known about the influence of natural changes in the biogeographic distribution of vegetation types on the future land C sink. Most studies focus on the CO 2 fertilization effect due to rising atmospheric CO 2 levels [e.g. Zeng et al., 2004;Thompson et al., 2004;Friedlingstein et al., 2001] or on the effects of anthropogenic land-use and land cover change (LULCC) [e.g. Lawrence et al., 2012;Brovkin et al., 2013a, 2013band Boysen et al., 2014. Even when the response of a changing vegetation cover is included in these model simulations, its impact on the C sink is not separated from other processes like CO 2 fertilization or anthropogenic LULCC. In particular, the impact of natural vegetation changes (caused by rising CO 2 and rising temperatures) on anthropogenic land cover changes has not been assessed by previous studies. The expansion of natural vegetation into a former desert, for example, provides arable land available for land-use (e.g. as crop or pasture) and thus affects decisions of future farmers to expand their agricultural activities, or, in the opposite case, to abandon farm lands. Both the natural vegetation change itself and its effect on LULCC may influence the global C cycle. Conversely, LULCC limits changes in natural vegetation distribution induced by climate change. Accordingly, synergistic effects on the C cycle must be expected. The return of natural vegetation, especially forests, over areas previously managed by humans for agriculture or pasture may also be affected by CO 2 and temperature changes, an aspect that is as yet uninvestigated in models.
The aim of the present study is to quantify the contribution of natural vegetation change and the consequences of interactions between natural and anthropogenic vegetation change for land C fluxes. Indeed, contributions from interactions between those two processes are implicit to many studies on LULCC emissions (e.g., three of the six models participating in the Land-Use and Climate, IDentification of Robust Impacts-Coupled Model Intercomparison Project Phase 5 (LUCID-CMIP5) model intercomparison investigating LULCC effects on climate and the C cycle [Brovkin et al., 2013a[Brovkin et al., , 2013b accounted for natural vegetation changes, while three did not). On the other hand, they are not accounted for in studies such as those applying the widely-used bookkeeping approach [e.g., Houghton et al., 2012] because this approach is by construction not suitable for inclusion of changes in natural vegetation . In this regard, our study contributes to a better understanding of uncertainties involved in estimates of LULCC emissions that are regularly used in Intergovernmental Panel on Climate Change reports to estimate the residual land C sink. It also highlights the complexity that arises in disentangling direct effects of human LULCC activity (changes in managed areas) from indirect effects of human activity: The latter include altered C fluxes on existing (managed and natural) vegetation due to climate and CO 2 changes and changes in C fluxes due to areas undergoing natural vegetation dynamics. Published studies have not been consistent in whether such indirect effects should be attributed to the net land-use flux or to the residual terrestrial sink . Although studies call for a clear separation of direct and indirect effects [e.g., Houghton et al., 2012], this separation is not easily possible, neither in observational data nor in modeling simulations. Our study highlights an additional, largely unrecognized, sources of complexity, namely the mutual dependence of natural and human-induced vegetation dynamics.
Our study is based on simulation results obtained with the MPI-ESM (Max Planck Institute Earth System Model) in the context of CMIP5 (Coupled Model Intercomparison Project Phase 5). We apply three alternative RCP (Representative Concentration Pathway) scenarios for our investigation and perform additional simulations with the land C subcomponent of the MPI-ESM. In section 2 we describe the models employed and in section 3 the experimental designs. Section 4 presents the results followed by a discussion in section 5. The Appendix A provides two auxiliary figures summarizing the results for land C stocks and land emissions from all simulations performed for this study.

Model Components
The MPI-ESM (Max Planck Institute Earth System Model) encompasses the following submodels: ECHAM6 for the atmosphere [Stevens et al., 2013], MPIOM for the ocean [Jungclaus et al., 2013], JSBACH for the land [Reick et al., 2013], and HAMOCC5 for the marine biogeochemistry [Ilyina et al., 2013]. In the low-resolution version underlying the present study, called MPI-ESM-LR, the resolution of ECHAM6 is set to T63 (i.e., about 2°h orizontal resolution at the equator) with 47 vertical sigma-pressure levels and MPIOM is driven at a resolution of GR15 (a horizontal resolution of about 1.5°) and 40 vertical levels.

Natural Carbon Cycle
The C cycle is fully represented within MPI-ESM by HAMOCC5 and JSBACH but in all simulations considered in the present study the course of atmospheric CO 2 is prescribed, so that ocean and land C dynamics do not influence the atmospheric CO 2 concentration nor each other. Thus, for our purpose of estimating land C fluxes only the C representation of JSBACH is decisive.
JSBACH simulates the land physics and biosphere in response to ECHAM6 climate and in turn affects the atmosphere through surface properties (surface temperature, roughness, and albedo) and by moisture and heat fluxes. In the present configuration the model features 11 plant functional types (PFTs) with 8 PFTs for natural vegetation (tropical and extra-tropical deciduous and evergreen trees, deciduous and raingreen shrubs, and C3 and C4 grasses) and 3 PFTs for agricultural vegetation (C3 crop, C3, and C4 pasture). NPP (net primary production) is allocated to the C pools for the living ("green") and lignified ("woody") parts of plants, as well as to a "reserve" pool for sugars and starches that plants store as energy reserve. Upon leaf shedding, C is transferred from the green and reserve pools to the "green litter" pool while C from the "wood" pool is transferred to the "woody litter" pool. To represent heterotrophic respiration, certain fractions of the litter pools are lost directly to the atmosphere. Other fractions are passed over to the humus pool where they are respired on centennial timescales. The temperature dependence of litter decomposition and soil respiration is described by a Q10 formula [Lloyd and Taylor, 1994], combined with a linear dependency on relative soil moisture content [Knorr, 2000]. The turnover times for woody and nonwoody litter decomposition are about 30 and 2 years, respectively, with exact values depending on the PFT. Besides prescribed minimum fire disturbance rates, fire occurs only when enough aboveground litter is available and when the litter is Global Biogeochemical Cycles 10.1002/2014GB004959 sufficiently dry [Reick et al., 2013]. With decreasing litter humidity the probability of fire increases linearly. The occurrence of windthrow disturbance in the model depends on the occurrence of extraordinarily strong wind speeds compared to a corresponding long-term average.

Natural Land Cover Change
Climate-induced changes in the potential geographic distribution of natural vegetation are simulated on a daily time step by the DYNVEG component of JSBACH [Brovkin et al., 2009;Reick et al., 2013]. The resulting natural biogeography is a consequence of three processes: First of all, the distribution of dry and cold deserts is derived from the survival conditions for plants. Second, the distribution of woody and grass PFTs is calculated with characteristic time scales for establishment and natural mortality and by assuming that upon an increasing level of disturbances (fire and windbreak) woody PFTs loose their structural advantage as compared to grass PFTs. Third, the relative occurrence of PFTs within each of the two classes of woody and grass PFTs is calculated by their relative dominance in net primary productivity. This implies that changes in environmental condition, including, e.g., CO 2 fertilization can affect vegetation distribution. From an ecological point of view this approach is discussed in Baudena et al. [2014]. In the presence of anthropogenic land cover change, DYNVEG does not calculate the actual distribution of natural PFTs but only the potential distribution of natural PFTs that is consistent with climate, which is then modified in a complex scheme by anthropogenic land cover change to obtain the actual land cover distribution.

Anthropogenic Land Cover Change
In the simulations presented here anthropogenic land-use change is globally prescribed from the land-use harmonization data set used in CMIP5 ( [Hurtt et al., 2011] see Reick et al. [2013] for a detailed description of this data set 's implementation in JSBACH). JSBACH translates an input sequence of annual maps of so-called land-use transitions between the land cover types primary land, secondary land, crop, and pasture into daily transitions between cover fractions of JSBACH's plant functional types (PFTs). In contrast to an earlier procedure of JSBACH (prescribing the geographic distribution of land cover for each year) transitions between land cover distributions rather than the land cover states themselves are prescribed. Furthermore, this scheme includes gross transitions, i.e., it includes subannual LULCC, which may not affect the net distribution of vegetation within a year but nevertheless causes C emissions. If we consider a situation where transitions within 1 year are such that before and after these transitions the areas of all PFTs are identical, C may still be reallocated in the model because of subannual land-use changes. An example is shifting cultivation which is not covered by the earlier procedure but by the transition scheme. Wilkenskjeld et al. [2014] and Stocker et al. [2014] show that accounting for gross transitions substantially increases the net land-use flux estimate compared to estimates accounting only for net transitions.
Since there are indications that historically pastures were preferentially established on former grasslands [Houghton, 1999;Ramankutty et al., 2006], the implementation of transitions in JSBACH encompasses the so-called "priority" or "pasture" rule [Reick et al., 2013]. Following this rule, pastures expand first into grasslands before other natural vegetation is touched. In contrast, croplands expand into forests and grasslands according to their relative presence. When agricultural lands are abandoned, forests establish first on these areas until the potential forest cover are reached. Afterward, grasses establish on the remaining land.
Besides direct anthropogenic land-use changes JSBACH performs a long-term adaptation of agricultural areas to climatic changes by translocating pastures from potential forest to potential grassland area [Reick et al., 2013]. This seems realistic since the effort for farmers to maintain pasture on potential grasslands is smaller than on potential forest area. This mechanism may also lead to abandonment of agricultural land if climate conditions lead to desertification that prevents growth of vegetation.
During transitions from natural to agricultural land cover types, the aboveground "living" biomass of the shrinking PFT is partly lost to the atmosphere and party distributed to the aboveground and belowground litter pools of the expanding PFT. C stocks in the shrinking PFT within the belowground living biomass or in the litter and humus pools is transferred to the corresponding pools of the expanding PFT according to the converted area.
Harvest rates for agricultural lands are prescribed from the land-use harmonization data set [Hurtt et al., 2011]. If the prescribed harvest C mass exceeds the existing aboveground biomass, only the available biomass Global Biogeochemical Cycles 10.1002/2014GB004959 is harvested. The harvested C is assumed to have a turnover time of months and is therefore put into the fast belowground litter pools.

Land Cover Evaluation
Brovkin et al.
[2009] evaluated the performance of JSBACH for natural vegetation dynamics. They overlaid the simulated tree cover of our model for the present day with the crop extent from land-use reconstructions. The comparison with satellite-based observations shows that global spatial patterns of forests are reproduced by our model. In a similar study Brovkin et al. [2013aBrovkin et al. [ , 2013b compare the year 2001 tree cover and bare ground ensemble mean from three MPI-ESM CMIP5 historical simulations with vegetation data based on satellite observations. The simulated tree cover and bare ground fractions are globally in good agreement with the observations. We therefore have confidence in the model's capability to simulate today's global land cover distribution.

Cbalone
A special feature of JSBACH is its capability to run the whole submodel for land cover change and land C independently from the rest of JSBACH. This submodel is called Cbalone and is forced by data obtained in simulations with the full JSBACH (run, e.g., as part of the fully coupled MPI-ESM). This reduced setup reproduces land cover and C from the full model to numerical accuracy. Cbalone includes the submodels for natural land cover change (DYNVEG), for wind and fire damage as well as for LULCC and is driven by daily time series of NPP, leaf area index, soil temperature, relative soil moisture, and wind speed. These variables are calculated by MPI-ESM for each grid cell. NPP, leaf area index, and soil moisture are further calculated for each PFT as value per square meter and are thus independent of the actual PFT cover fraction, allowing us to combine this output with any land cover distribution in Cbalone. The great advantage of Cbalone therefore is that we do not have to run the full MPI-ESM to separate the contributions of various subprocesses to the results of a full ESM simulation. Instead, it is possible for this purpose to rerun Cbalone with the forcing obtained from the full MPI-ESM simulation but using different submodel configurations, e.g., with and without natural vegetation dynamics or with alternative land-use change scenarios. However, an analysis of this type ignores feedback between land processes and climate, because climate conditions are prescribed from the full MPI-ESM simulation when running Cbalone (for details see section 3.5).

Overview
Regarding future climate and CO 2 change we aim at quantifying the dependence of the natural land C sink on anthropogenic LULCC and the converse dependence of C emissions from LULCC on changes in natural vegetation. To disentangle these effects we perform simulations with vegetation dynamics and LULCC each switched on and off. Therefore, we have four simulation setups (with each encompassing  Emissions are defined here as differences between two simulations. For example, E NAT + ANT denotes emissions from simultaneous natural and anthropogenic vegetation changes (simulation tNAT_tANT minus simulation cNAT_cANT). The notation used for the various emissions is explained in section 3.1.

Global Biogeochemical Cycles
10.1002/2014GB004959 three future scenarios). These setups are described in the following paragraphs; an experiment overview is given in Table 1. Here, the term "emissions" refers to the difference in land C storage that arises from switching a process on or off. Emissions from climate-induced changes in natural vegetation (E NAT ) are thus calculated from land C storage differences from two Cbalone runs with vegetation dynamics switched on and off. Since these emissions can be derived from two different kinds of model configuration, either including LULCC or not, one can distinguish emissions E cANT NAT and E tANT NAT where we use the upper index to indicate whether LULCC was constant (cANT) or transient (tANT) in the simulations from which the emissions were derived. Analogously, emissions from LULCC (E cNAT ANT , E tNAT ANT ) are calculated from pairs of simulations with LULCC switched on and off, while natural vegetation is kept fixed at the preindustrial distribution (index cNAT) or is transiently changing (index tNAT). Note that the emissions calculated in this way do not include all possible effects of vegetation on the C fluxes but-as intended-only those from the changes in vegetation cover. For example, future land C storage increases also without any vegetation cover change simply because climate is changing in the scenarios and because enhanced atmospheric CO 2 increases plant productivity by CO 2 fertilization (see cNAT_cANT in Figure S1 in the supporting information). These effects are inherent in the simulations but largely cancel since we consider only differences between two simulations that are subject to the same underlying climate change. Climate change thus induces differences in land C storage between simulations only in areas that are concurrently undergoing a (natural or human-induced) change in vegetation cover. For example, an expansion of forest on grassland under climate change will induce a C sink not only because forest, in general, has a higher biomass than grasslands but also because the slower turnover rate of woody material allows the forest to store more C due to CO 2 fertilization as compared to the simulations where grassland prevailed. Pongratz et al. [2014] identified nine different definitions of the net LULCC flux in published literature. The setup used here to derive emissions from LULCC corresponds to their definition "D3." This means that emissions are calculated from simulations with transiently changing environmental conditions identical for the simulation with and without land cover change. The remarks made there on completeness in accounting for all possible C fluxes apply as well to the derivation of the analogous fluxes of natural land cover change analyzed in the present study. In the following, negative emissions represent storage of C at land.

Experiment tNAT_tANT
The first set of experiments analyzed in this study was conducted with MPI-ESM-LR for the Coupled Model Intercomparison Project Phase 5 (CMIP5) following the specifications of Taylor et al. [2012]. These are (in CMIP5 syntax [see Taylor et al., 2011Taylor et al., , 2012) historical_r1i1p1-LR, rcp26_r1i1p1-LR, rcp45_r1i1p1-LR, and rcp85_r1i1p1-LR, of which the first covers the historical period 1850-2005, and the other three follow the specification of the Representative Concentration Pathways (RCP) from 2006 to 2100 and their extensions until 2300 [van Vuuren et al., 2011]. For our purposes we denote this set of experiments as tNAT_tANT since they were performed with transient changes in natural vegetation and anthropogenic LULCC. In the following we provide only a brief description of this group of experiments; more information can be found in Giorgetta et al. [2013].
In all simulations greenhouse gas concentrations are prescribed. While we follow the historical and RCP scenarios for the land-use change forcing (including harvest) until 2100, anthropogenic land-use change after the year 2100 is set to zero to investigate time scales and amounts of C recovery after vegetation relaxation from LULCC; only wood harvest continues at the rates of 2100 to mimic constancy of land-use. The historical_r1i1p1-LR run includes the influence of natural and anthropogenic forcings derived from observations. These forcings encompass orbital parameters, solar irradiance, greenhouse gases, ozone, tropospheric and stratospheric aerosols, and LULCC including wood harvest. The RCP26, RCP45, and RCP85 runs start each in 2006 from the state at the end of the historical experiment. The greenhouse gas forcing of these experiments follows the RCP scenarios developed by Moss et al. [2008] until the year 2100. Afterward, the extended concentrations pathway scenarios of Meinshausen et al. [2011] are applied until the year 2300. The natural forcing includes the same mechanisms as in the historical experiment, except for volcanic aerosols that are set to zero after 2005.
The numbers of the RCPs denote the maximum 21st century top of atmosphere radiative forcing for which they were designed. Accordingly, RCP8.5 is the scenario with the strongest greenhouse gas emissions such Global Biogeochemical Cycles 10.1002/2014GB004959 that the rise in atmospheric CO 2 continues rapidly until 936 ppm in 2100 and rises further to 1962 ppm in 2300. RCP4.5 involves moderate measures against global temperature rise. Here CO 2 reaches about 540 ppm and stays at this level until 2300. The most ambitious scenario is RC2.6: by a mix of technological innovations atmospheric CO 2 reaches a maximum of 443 ppm around 2050, thereafter decreases to a value of 421 ppm in 2100, and finally reaches 361 ppm, which lies below today's CO 2 concentration. In RCP8.5 and RCP2.6 agricultural areas expand until almost all arable land is put under human management. In contrast, RCP4.5 is an ambitious afforestation scenario.

Experiments cNAT_cANT
The cNAT_cANT set of experiments is performed to isolate the C emissions that are induced by the combined natural and anthropogenic land cover changes. In these simulations Cbalone is forced with the output from tNAT_tANT (as described in section 2.6), while the distribution of natural and anthropogenic vegetation (including harvest) are kept constant at their 1850 values for the whole simulation period. The land C pools start from their 1850 conditions and develop exclusively as consequence of the changing atmospheric CO 2 and climate. As shown in Figure 1, the emissions from the combined effect of natural and anthropogenic land cover change (denoted by E NAT + ANT ) are then obtained from the difference in total land C of the two simulation sets tNAT_tANT and cNAT_cANT.

Experiments tNAT_cANT and cNAT_tANT
Two additional sets of Cbalone simulations are conducted to separate the individual impacts of LULCC and changes in natural. Both follow the same general setup as cNAT_cANT, but switch off only land-use change (simulation set tNAT_cANT) or only the change in natural landcover (simulation set cNAT_tANT). Having these additional simulations we can calculate all emissions defined in Figure 1.

Inconsistencies Inherent to the Cbalone Runs
It should be noted that all sets of Cbalone experiments have common climatic forcings, which are derived from tNAT_tANT. As a consequence, climatic feedback due to changes on the land surface are suppressed, and thus, differences in the simulations resulting from biogeophysical effects of natural and anthropogenic land cover change are excluded. However, the resulting inconsistencies between climate and land surface should not principally change the results of the present study and are anyway implicit to the underlying CMIP5 simulations because of the prescribed atmospheric CO 2 levels.

Notation
The notation introduced above for emissions is used analogously also for other quantities: Δf cANT NAT in Figure 4b denotes the difference in plant cover fraction f between simulation tNAT_cANT and cNAT_cANT, and similarly, E cANT NAT denotes the associated change in land C storage C (e.g., Figure 4c).

Overview of Land C Emissions Estimates
The green lines in Figure 2 display the contribution E tANT NAT of natural vegetation changes on total C emissions from LULCC and natural vegetation E NAT + ANT (red lines) for the historical period and the three scenarios.

10.1002/2014GB004959
Since E tANT NAT is negative throughout the simulation period, these natural changes have in the past-and might also in the future -mitigate emissions caused by LULCC. This mitigation effect is quite substantial, as also the comparison with the separate contribution from changes in anthropogenic vegetation E tNAT ANT (blue lines) shows. For example, for the year 2100, the (cumulative) contribution from changes in natural vegetation is about one third of the LULCC contribution for RCP26/RCP85 and over one half for RCP45 (Table 2). However, natural and anthropogenic vegetation do not develop independently under the common climate. Accordingly, the sum of the individual contributions is not equal to the total effect. This is demonstrated by comparing E tANT NAT þ E tNAT ANT (black lines) with the total emissions E NAT + ANT in Figure 2 (red lines): The black lines show simply the sum of the single contributions (green plus blue lines), while the red lines show the combined effect when both natural and anthropogenic vegetation changes are simulated simultaneously (tNAT_tANT minus cNAT_cANT). Obviously, the black and red lines do not coincide for the future predicitions due to synergistic effects: As the two individual effects are calculated here under the presence of the other process, they cannot be separated fully from each other: the dynamics of natural vegetation also includes their impact on LULCC and vice versa.
There is a surprisingly strong difference in strength of the synergies between the different scenarios: Whereas for RCP26 and RCP45 the synergies lead to only slightly higher cumulative C emissions (~5 Gt C and~24 Gt C in 2300); they even dominate cumulative emissions for RCP85 in 2300 (~156 Gt in 2300). The reasons for this different behavior will become clear after the more detailed discussion below on the mutual influences between changes in natural and anthropogenic vegetation. Figure 3 shows the C emissions from changes in natural vegetation (E cANT NAT , E tANT NAT ) in the various simulations. Except for a short period around 1900 they are negative because of expansion of natural vegetation under the increasing CO 2 concentration and associated warming (Figure 4a). To get a geographical impression of the regional origin of these emissions, Figures 4b and 4c show the differences Δf cANT NAT and ΔC cANT NAT in vegetation cover and land C (compare section 3.6) that appear from the climate-induced changes in natural vegetation at the end of our historical plus RCP45 simulation (year 2300). The corresponding patterns for the RCP26 and RCP85 scenarios look very similar with respectively weaker and stronger changes (not shown). In regions with generally high vegetation coverage (e.g., inner tropics, midlatitudes, and SE Asia) there is not much potential for further increase in vegetation coverage; accordingly, in such regions the induced change in vegetation coverage is small. In contrast,  in the high latitudes conditions for vegetation growth strongly improve due to polar warming (compare Giorgetta et al. [2013]) resulting in a significant increase in vegetation coverage.

C Emissions From Changes in Natural Vegetation 4.2.1. General Trends
A Comparison Figures 4b and 4c reveals that the geographical distribution of total vegetation cover changes alone cannot fully explain the corresponding change in land C storage: Besides regions where the expectation is met (mostly in the high latitudes) the figures show also regions with a decrease in land C despite an increase in vegetation coverage (e.g., in N-E America, S-Africa, and S-E Asia). A more detailed analysis reveals that the C reductions in these areas are caused by a shift from more productive to less productive PFTs under the changing climate, namely, from forest to grasslands in the particular regions mentioned. Nevertheless, global emissions are negative ( Figure 3) because they are dominated by the boreal expansion of vegetation. However, the above findings illustrate that the amount of negative C emissions does not only depend on vegetation expansion and CO 2 fertilization but also on PFT changes.

Modifications From Changes in Anthropogenic Vegetation
In the presence of anthropogenic vegetation changes the expansion of natural vegetation is slightly weaker (cf. Figure 4a), but the effect on land emissions surprisingly shows the opposite sign: The curves for E tANT NAT in Figure 3 lie below the curves for E cANT NAT , which implies that the presence of LULCC enhances the global C uptake from changes in natural vegetation. Thus, even when the absolute C storage on land is always lower in the simulations with LULCC, the relative storage due to changes in natural vegetation is higher with LULCC. This seems counterintuitively as under LULCC the spreading of the natural vegetation is partly transformed into managed PFTs and we would therefore expect a lower C storage than without LULCC. However, it turns out that under LULCC the dynamic vegetation also reduces LULCC C loss as anthropogenic PFTs can now shift from potential forest to potential grass area. In the following this mechanism will be explained in detail.
Besides leading to a small reduction in total vegetation extent (Figure 4a) we find that LULCC influences the natural PFT changes in such way that more productive PFTs are established. The shift in PFT distribution under the presence of LULCC is depicted in Figure 5 exemplarily for RCP85. Without LULCC (cANT) grasslands expand substantially under the warming climate, partly at the cost of extratropical forest and shrubs. As expected, with LULCC (tANT) parts of the grasslands are transformed to pasture and crop. While extratropical forests are significantly diminished in the absence of LULCC, their extent remains almost unchanged when LULCC is accounted for (see Figure 5). It turns out that in the presence of LULCC the natural extratropical forest reduction (as it is visible without LULCC) is compensated by less extratropical forest reduction as consequence of LULCC. This is because the submodel for long-term adaptation of agricultural areas (section 2.4, third paragraph) gets active (which represents the reaction of farmers to changes in natural vegetation) when the natural vegetation dynamics in JSBACH is switched on (tNAT_tANT). This long-term adaptation can lead to a slow geographical shift of pastures toward lands where without anthropogenic disturbance grasslands would occur. Associated with this shift, formerly deforested extratropical forests can reestablish until their natural potential is reached. In the concrete tNAT_tANT simulation extratropical forests can strongly reduced until the year 2100 because of the agricultural expansion. At the same time natural grasslands strongly expand in the course of the simulation even after 2100. The emergence of additional grasslands causes pastures to shift towards these areas and extra-tropical forests strongly reestablish (as far as their potential occurrence is not limited by their reduction due to the natural vegetation change). As a result, former reductions of extratropical forests due to LULCC are reversed with the spreading of potential grasslands. Figure 5 shows the cover fraction changes with and without LULCC for RCP85. The differences in productivity of the PFTs with and without LULCC (brown and green bars in Figure 5) are depicted in Figure 6. As expected the productivity increase of PFTs with a larger extent in the presence of LULCC (extratropical forests, shrubs, pastures, and crops) is higher than the productivity loss of PFTs with a smaller extent in the presence of LULCC (tropical forests and grasses). The productivity increase of extratropical forests is the main reason for the stronger C storage due to natural vegetation changes under LULCC.

Understanding the Emission Differences Between the Scenarios
In general, C emissions in Figure 3 are negative because of expansion of natural vegetation providing both a direct sink to C in the buildup of C stocks as compared to nonvegetated land and an additional C sink in response to CO 2 fertilization. The increase in strength of the cumulative sink seen in E cANT NAT (brown lines) of RCP85, the scenario with strongest climate change and CO 2 rise (the CO 2 rise being 5.0 (2.5) times the rise in RCP26 (RCP45)), slows down around 2070 because grasslands start to expand strongly at the cost of extratropical forests. For RCP45 the curve of E cANT NAT decreases stronger than the one for RCP26 because natural vegetation spreading is higher. As explained above, E tANT NAT (green lines) lies for all RCPs below E cANT NAT (brown lines) since LULCC leads to more productive PFTs. However, for RCP85 the difference between brown and green lines is large (as compared to this difference in the other RCPs) because with LULCC (green) formerly reduced extratropical forests are reestablished so that the slow down appearing without LULCC (brown) is overcompensated. Figure 7 shows the land C emissions caused by anthropogenic LULCC. In all scenarios LULCC increases land C Figure 6. Difference in net primary productivity (NPP) due to the presence of LULCC (tANT) at the end of our historical plus RCP85 simulation (average over all simulation years). This is the relative change in average NPP associated with the difference in bar height in Figure 5 (green bars minus brown bars). More precisely, we calculated here (using our condensed notation and n as the number of PFTs) NPP tANT NAT; i À NPP cANT NAT; i = X n i NPP tANT NAT; i , accordingly, a value of "1" means 100% of the overall increase due to LULCC. emissions. Due to the strong deforestation and the well-filled C pools its emissions are highest in the RCP85 scenario. Due to the strong reforestation in the RCP45 land-use scenario, emissions in RCP45 are even lower than those in RCP26. After 2100 the accumulated C emissions diminish. This is because land-use change ends in 2100 by definition of the RCPs and extended scenarios. Before 2100 land-use change has depleted in particular the C pools of the living parts of vegetation. Under ongoing land-use change LULCC emissions are only partly compensated by refilling depleted C pools, but after 2100, in the absence of land-use change, only the refilling continues so that accumulated emissions decrease. It should be noted that reforestation goes along with the removal of living vegetation. This would be different for afforestation by abandonment of land-use, where C pools would not be touched during that land-use change. Depending on the scenario, land-use change affects different vegetation types and is happening under different climates. Accordingly, this refilling of C pools takes more or less time. Only for RCP85 global C stocks reach equilibrium regionally (not shown) and globally (Figure 7) by 2300, whereas for the other scenarios at that time the refilling is not fully completed.

Modifications From Changes in Natural Vegetation
In the presence of changes in natural vegetation, emissions from LULCC are weaker (E tNAT ANT < E cNAT ANT , compare Figure 7 and Table 3) due to the strong productivity increase of natural vegetation when it occurs in combination with LULCC. This is the same effect as explained in section 4.2, just from the LU emission perspective (black lines in Figure 3 are the same as black lines in Figure 7). Note that the comparison of LULCC emissions with other studies (cf. Table 3 and the LUCID study of Brovkin et al. [2013aBrovkin et al. [ , 2013b) reveals that land-use emissions are quite high in our model as compared to others. Houghton et al. [2012] reviewed so far the processes included and ignored in estimating global emissions from LULCC by model simulations. Changes in natural vegetation distribution and performance were not mentioned as missing. And indeed, the review addressed only the historical period for which the impact from such additional human-induced changes is presumably minor (E tNAT ANT and E cNAT ANT are almost identical until 2030 in Figure 7). But as our study shows, such changes may gain importance under future climates: For emissions it makes a difference how productive the replaced vegetation is, and this depends on climate either directly through vegetation productivity (via temperature, precipitation, and CO 2 ), or because of climate-induced shifts in natural vegetation that cause land-use change to replace a structurally different vegetation type (e.g. forest instead of grass).

Discussion and Conclusions
That geographically identical land-use changes that lead to different emissions under different climates have been well recognized by Pongratz et al. [2014]. Insofar, our study is only exemplifying this effect in more detail. But our study also demonstrates that the presence or absence of land-use change impacts the climate-induced land sink (i.e., negative emissions) arising from natural vegetation (compare E tANT NAT and E cANT NAT in Figure 3). Concerning the emissions from land-use change, the picture is not that uniform. The cumulated land-use emissions (E tNAT ANT in Figure 7) increase during the historical period. For RCP85 and RCP26 they continue to increase until 2100 but decrease afterward, because, with land-use change being stopped in 2100 the depleted C pools are refilled and the assumed agricultural reorganization toward using the extending natural grasslands instead of forests for cattle grazing dominates. In contrast, for the reforestation scenario RCP45 cumulated emissions already stall during the 21st century before they decrease after 2100. In section 3.5 we discussed the inconsistencies between climate forcing and land surface conditions within our model runs that arise from our particular simulation setup. These inconsistencies could only be avoided by using the full MPI-ESM instead of only the submodels for C allocation and vegetation dynamics. It is unlikely that these inconsistencies crucially distort our results. Brovkin et al. [2013aBrovkin et al. [ , 2013b show for MPI-ESM-LR that there are nearly no significant changes from LULCC in annual land surface air temperatures. Significant changes in land surface albedo, available energy, and latent heat fluxes are also found to be small.
Our model results strongly depend on the representation of natural and anthropogenic land cover changes and the associated C fluxes but unfortunately evaluation of these factors is difficult. As shown in Table 3 our C emissions from anthropogenic LULCC are high as compared to other models. One explanation is that the dynamic vegetation overestimates the spacial extent forests and associated with this the corresponding C stocks [Goll et al., 2015]. A further explanation is that slash-and-burn and harvest emissions are-in contrast to some other models-represented in our LULCC scheme. Another kind of shortcoming in our simulations concerns the representation of afforestation. In our model afforestation leads to the recovery of natural (primary) forests instead of anthropogenic (secondary) forests (like commercial forests, palm oil plantations, or orchards). Since anthropogenic forests can include a lot of bioenergy with low C stocks the C storage in plants and soils of our model might be overestimated. Forest density and productivity might be overestimated in our model but the fast turnover rates are represented by wood harvest.
As a consequence of the above-mentioned caveats our attention should focus more on the evaluation of JSBACH and on the implementation of additional anthropogenic PFTs. Nevertheless, even though simulated C emissions are difficult to evaluate, our study indicates that the effects discussed here are still likely to be important under future climates. So for example, we think that our statement that natural vegetation dynamics clearly mitigates LULCC emissions still holds although C emissions are model dependent. For all RCPs the climate-induced change in natural vegetation strongly mitigates (cumulated) LULCC emissions (depending on the RCP and the year considered; compare, e.g., E NAT + ANT and E tNAT ANT in Figure 2). This is largely caused by warming in the high latitudes whereby large areas get available for vegetation (Figure 4b), especially for forest and, to a minor extent, grasses. These forests are highly productive, inducing a carbon sink that counteracts emissions from LULCC.
In conclusion, we find that land-use change and climate-induced changes in natural vegetation strongly influence each other and cannot be regarded as independent processes (compare E NAT + ANT and the sum of E tNAT NAT and E tNAT ANT in Figure 2). However, the mechanism causing this strong interdependence in our model significantly depends on the assumed agricultural reorganization associated with what we call the "pasture rule" (the preferential use of natural grasslands for pastures and the long-term shift of pastures toward potential grassland). Although plausible, it is not clear how farmers in the future organize their land-use. Simply ignoring human reactions on a strongly changing natural environment under climate change is equally problematic; even more so as our study indicates that the decision of farmers how to rearrange their land-use spatially will be of importance.

Appendix A
Throughout the main text of our paper, only "emissions" are discussed, i.e., differences in land C storage between two simulations. For completeness, we show here in Figure S1 our simulation results for land C storage found in the individual simulations. Moreover, the discussion in the main text concentrated on cumulative emissions. Once more for completeness, we show in Figure S2