Earth system feedbacks following large-scale tropical forest restoration

To achieve the Paris Agreement requires aggressive mitigation strategies alongside negative emission technologies. Recent studies suggest that increasing tree cover can make a substantial contribution to negative emissions, with the tropics being the most suitable region from a biogeophysical perspective. Yet these studies typically do not account for subsequent carbon cycle and climate feedback processes of large-scale land use change. Here we quantify the maximum potential temperature and CO2 benefits from pantropical forest restoration, including earth system feedbacks, using a fully-coupled, emission-driven 5 Earth System Model (HadGEM2-ES). We perform an idealised experiment where all land use in the tropics is stopped and vegetation is allowed to recover, on top of an aggressive mitigation scenario (RCP 2.6). We find that tropical restoration of 1529 Mha increases carbon stored in live biomass by 130 Pg C by 2100 CE. Whilst avoiding deforestation and tropical restoration in the tropics removes 42 Pg C compared to RCP 2.6, feedback processes mean that carbon in the atmosphere only reduces by 18 Pg C by 2100. The resulting, small CO2 (9 ppm) benefit does not translate to a detectable reduction in 10 global surface air temperature compared to the control experiment. The greatest carbon benefit is achieved 30–50 years after restoration before the Earth System response adjusts to the new land-use regime and declining fossil fuel use. We identify three model-independent key points: (i) the carbon benefit of restoration is CO2-scenario dependent, (ii) in a world that follows Paris Agreement emission cuts restoration is best deployed immediately, and (iii) the ocean carbon feedbacks will reduce the efficacy of negative emissions technologies. We conclude that forest restoration can reduce peak CO2 mid-century, but can only be a 15 modest contribution to negative emissions.

LULCC in RCP 2.6 are small (< 2 Pg C year −1 ) compared to fossil fuel emissions (9.2 Pg C year −1 ). With the widespread implementation of BECCS tropical land use area is expected to increase by 286 million ha (Mha) by 2100 in RCP2.6 (van 90 Vuuren et al., 2011). New bioenergy crops are primarily allocated near existing land use areas (Hurtt et al., 2011;van Vuuren et al., 2011).
This simulation uses carbon emissions from the fossil fuel industry and land use to interactively calculate atmospheric CO 2 rather than prescribing CO 2 concentrations (i.e. it is "emission driven"). The fossil fuel emissions were computed by a RCP 2.6 concentration driven HadGEM2-ES simulation (Liddicoat et al., 2013).
All forcing data, except the scenario related fossil fuel and land use/disturbance mask have been implemented in HadGEM2-ES as described in (Jones et al., 2011). The urban (20 Mha), inland water (330 Mha) and the ice fraction (163 Mha) remain constant throughout both simulations. In addition to the land carbon fluxes E DEFOR and F LA , the carbon cycle in HadGEM2-ES includes the flux from ocean to atmosphere (F OA ), fossil fuel emission (E FF ) and the atmospheric growth rate of carbon (G ATM ) 100 (Equation 1).

Biomass scaling
A doubling in biomass would be caused by either an increase in NPP and no change in litter flux, or no change in NPP and a decline in litter flux, causing no change in soil carbon and R h , meaning the ratio between NPP and R h increases, leading to an increase in tropical F LA also by a factor of two (83.6 Pg C). Scaling this by the ratio between F LA (tropics) and F LA 105 (global) ( Table 1) gives the global F LA from doubling tropical biomass (46.8 Pg C). Scaling by the ratio between F OA and F LA (Table 2) results in a cumulative net flux to the atmosphere of 30 Pg C, equal to 15 (14.2-16) ppm CO 2 (based on 2.00±0.12 Pg C = 1 ppm CO 2 in HadGEM2-ES). This estimate assumes a linear carbon cycle response to increasing biomass and excludes any climate-carbon cycle feedbacks.

Restoration impacts on land cover 2006-2100 CE
In the control simulation (control), broadleaf forest declined globally by 107 Mha from 2006-2100 CE and by 213 Mha in the tropics. In the restoration simulation (restore), ending human land use led to an increase in broadleaf forest of 671 Mha globally, and 572 Mha in the tropics (Fig. 1A). Unsurprisingly, the largest differences between control and restore in all five plant functional types (PFTs) used in the model (Cox, 2001) are located in the tropics (Table A1). Anthropogenic disturbance 115 increases in control over the 94 years, as do C 3 grasses (and crops) and to a smaller extent shrubs, while tree PFTs and C 4 grasses decline. By contrast, in restore anthropogenic disturbance of 1529 Mha in the tropics abruptly ends, with shrubs 4 https://doi.org/10.5194/bg-2020-432 Preprint. Discussion started: 30 November 2020 c Author(s) 2020. CC BY 4.0 License.   increase in needleleaf trees in the tropics is located around the edges of the Amazon and Congo basin in restore. Shrubs also substantially increased in the tropics by 409 Mha in restore compared to control (Table A1). These increases in tree and shrub cover in restore compared to control in the tropics, were at the expense of C 3 and C 4 grasses (including crops).
Abruptly stopping anthropogenic land use in the tropics led to vegetation changes outside the tropics. Broadleaf trees declined, by 8 Mha, mostly in the high latitudes of North America and East Asia, being replaced by needleleaf trees. Shrubs 130 declined in total, particularly along the eastern edge of the Arabian peninsula. C 3 grasses declined overall, with a complex spatial pattern: increases in Australia, southern Africa, and central Asia, but almost 100% decreases in western Asia and the mid-western U.S. C 4 vegetation increases in Australia and the mid-western U.S., leading to a net increase in C 4 grasses in the extratropics. Overall, tropical land use change altered patterns of vegetation cover globally by 2100 CE.
The final land use change is deforestation. Overall global deforestation emissions, E DEFOR , correspond to the episodes of 135 land use expansion in RCP 2.6, with deforestation halted in the tropics only in restore. The pattern is reflecting waves of deforestation until 2035 CE, after which emissions decline rapidly and remain low for the rest of the century (Fig. 2). In control, these deforestation emissions are largest in the tropics, particularly the Amazon, central Africa, and south-east Asia ( Fig. A1). In restore, tropical deforestation is halted, but extratropical deforestation occurs at a modest level until 2035 CE under RCP 2.6. Overall, global cumulative E DEFOR are 16.1 Pg C (control) and 6.5 Pg C (restore), resulting in an emission 140 reduction of 9.6 Pg C from halting deforestation alone.

Land carbon response to tropical restoration
The terrestrial carbon cycle responds to the tropical land cover change with an increase in net carbon uptake, driven by an increase in biomass in the tropics. Overall, there is a decline in net primary productivity (NPP), but there is also a stronger decline in litterfall (flux of dead plant matter) into the soil. The flux of carbon into the atmosphere from soil respiration is 145 6 https://doi.org/10.5194/bg-2020-432 Preprint. Discussion started: 30 November 2020 c Author(s) 2020. CC BY 4.0 License. smaller than the uptake through NPP, thereby creating a net carbon sink (Table 1). NPP declines globally at first due to the change from grassy vegetation and crops to trees in restore, and secondly due to the lower atmospheric CO 2 concentrations in restore relative to control. The subsequent decline in dead plant matter leads to lower soil carbon stocks in restore compared to control, which ultimately leads to lower soil respiration ( Table 1).
The flux of carbon from the atmosphere into the living biomass carbon pool is modulated by the NPP of the different 150 vegetation types. The global trend, variability, and the differences in both simulations are driven by tropical NPP, although extratropical NPP also declines after ∼40 years (Fig. A2A). Taken together this results in lower cumulative global NPP at the end of restore compared to control ( Table 1). The main reasons for the greater NPP decline in restore is that grassy vegetation has a two to three times greater NPP per unit area than tree PFTs in HadGEM2-ES (Table A2), and a lower CO 2 fertilisation from lower atmospheric CO 2 in the second half of the century.

155
The carbon flux of dead plant matter into the soil carbon pool (F SC ) is determined by a temperature and PFT-specific turnover rate, as well as the areal extent and biomass of each PFT. Trends in F SC have a similar spatial and temporal pattern to NPP ( Fig.   3B; Table 1). In control, global F SC increases before it slowly declines from 2040 CE onward. In restore, F SC initially drops, . before it increases until 2040 CE and then again declines over the second half of the simulation. In both simulations global F SC is primarily modulated by the tropics (Fig. A2B). Outside the tropics F SC gradually increases until 2030 CE, after which 160 it remains constant in control but slowly declines in restore. The difference in cumulative global F SC by 2100 CE is driven by its large decrease in the tropics and further amplified by a more modest decline in the extratropics (Table 1). In particular the decrease in tropical F SC in restore is caused by the decline in grassy vegetation. At the same time lower overall production in extratropics leads to decreased F SC in restore.
Finally, some vegetation is converted via deforestation, 458 Mha, as specified in RCP 2.6 in control, but only the 172 Mha 165 specified in the extratropics by RCP 2.6 in restore. This global carbon flux from deforestation is largely transferred into the wood product pool (F WP ), and so shows the same pattern as deforestation (Fig. 3C c.f. Fig. 2A).
The net change in living biomass carbon is the difference between NPP and the combined fluxes into the soil carbon pool and the wood product pool (∆Biomass = NPP − F SC −F WP ). Globally, in control, ∆Biomass slowly shifts from net-zero change to a small loss by the end of the century. This is because high deforestation before 2040 CE compensates for the initial rise in NPP 170 and litterfall flux (F SC ; Fig. 3D). In restore, ∆Biomass moves from a rapid gain in the first decade towards net-zero change by the end of the century, driven by NPP initially exceeding F SC and no tropical deforestation resulting in a small flux from wood products (F WP ; Fig. 3D). Again, these differences are driven by changes in the tropics. Here biomass continuously declines in 8 https://doi.org/10.5194/bg-2020-432 Preprint. Discussion started: 30 November 2020 c Author(s) 2020. CC BY 4.0 License.
control but peaks and then declines in restore (Fig. A2F). In the extratropics, ∆Biomass turns from a net gain in the first half of the century into a net loss in the second half of both simulations, although more so in restore. The cumulative global biomass difference between both simulations is driven by the increase in biomass in the tropics in restore, which is slightly offset by the late decline in extratropical biomass (Table 1). In total the increase in biomass resulting from tropical restoration alone is 134.4 Pg C by 2100 CE, with a total global difference from control of 130 Pg C (Table 1). Averaged over the last decade of the simulation, the grid cell by grid cell difference in mean biomass carbon density in the tropics is up to 150 Mg C ha −1 between the two simulations ( Fig. 3G). Biomass growth mirrors the increase in woody PFTs and is highest around existing forest edges 180 in Amazonia, northern Mexico and the Congo Basin.
In terms of timescales, biomass carbon increases are highest in the first 20-40 years after land use ceases. On a per hectare basis the median carbon accumulation rates over first 20 years after land use cessation is 1.8 Mg C ha −1 yr −1 (range 0−5.5 Mg C ha −1 yr −1 ), using the grid cells where the share of broadleaf trees increased by at least 30% of the grid box area (n = 224) in restore (dark green in Fig. 3G). Thereafter uptake rates decrease to a median of 0. The changes in soil carbon (∆SC) are of the same magnitude as the changes in biomass carbon. In control, ∆SC increased by 141.9 Pg C by 2100 CE, mostly in the extratropics (114.1 Pg C), because of lower heterotrophic soil respiration relative to the litter influx (Table 1), likely due to cooler temperatures. In restore, global soil carbon uptake is much lower at 45.4 Pg C, with a similar strong uptake in the extratropics (100.1 Pg C), but a loss of soil carbon in the tropics (−54.8 Pg C). This is due 195 to land cover change from grass and crops to woody vegetation with lower NPP and litter input ( Table 1). The net change of the soil carbon pool is determined by the difference between litter inputs (F SC ) and heterotrophic soil respiration (R h ), the flux from the soil carbon pool into the atmosphere. R h itself is a function of temperature, soil carbon and F SC (Essery et al., 2003), so it increases as F SC and NPP increase. Therefore, overall R h increases to a much higher level in control by about 2030 CE, and stays high with a slight decline. By contrast R h declines rapidly in restore within a decade, and only partially recovers, 200 staying at a lower level until 2100 CE (Fig. 3E). As with the NPP and F SC , the global differences in R h between control and restore are driven by the decline in tropical R h in restore and amplified by marginal differences in the extratropics (Fig. A2E), ultimately linked to lower atmospheric CO 2 concentrations leading to less CO 2 fertilisation, lower NPP and less litterfall ( Table   1). Global ∆SC in both control and restore is net positive but saturates towards the end of the century, with, on average, lower values in restore (Fig. 3F). The difference comes from the tropics, where ∆SC in control varies around zero but is on average 205 ∼1 Pg C year −1 lower in control (Fig. A2G). Extratropical ∆SC slowly declines in both simulations but remains net positive.
The driver for the cumulative difference in global soil carbon content is a stronger decline in R h over F SC (Table 1).
Both simulations show that the land-to-atmosphere flux (F LA ) is a strong carbon sink at the beginning of the simulation (average 3.4 Pg C yr −1 in the 2020s in both simulations), but that this sink is diminishing (Fig. 2C). This is driven by a greater decline in NPP relative to R h . In the tropics, terrestrial carbon uptake reduces throughout both simulations and the land turns 210 into a carbon source in the final 50 years of both simulations. In the extratropics, the land is a carbon sink for most of both simulations, only turning into an occasional net source towards the end of the century. Overall, cumulative carbon uptake in restore is 16% higher than in control ( Table 1). The overall increased carbon uptake in restore relative to control becomes apparent in the difference in cumulative F LA (Fig. 4C). The difference between the two simulations first grows until 2036 CE, but then declines again after 2066 CE. This pattern is due to a greater change in NPP relative to R h in restore. First the tropical 215 land cover change increases NPP relative to R h in restore while in control this ratio remains unchanged. In the last part of both simulations NPP then declines relative to R h , but more so in restore due to lower CO 2 fertilisation from lower atmospheric CO 2 compared to control, causing the difference between both simulations to shrink. The global cumulative difference in land to atmosphere flux between restore and control is −23.4 Pg C (Table 1), with the largest difference at 2031 CE (−26.5 Pg C).
A large multi-year variability in F LA is associated with fluctuations between the components of the land-atmosphere carbon 220 flux (Fig. 2C). Global NPP in the simulations exhibits large multi-year variability of 4 Pg C year −1 throughout both simulations, driven by climate variability (Fig. 3A). This variability of ∼4 Pg C year −1 is also seen in F SC and translates into a ∼3 Pg C year −1 variability in ∆SC and R h (Fig. 3B, E, and F).

Ocean and Atmosphere feedbacks
The ocean net carbon sink (i.e. ocean to atmosphere carbon flux, F OA ) declines in both control and restore over the whole simu-225 lation period (Fig. 2D). This net sink is smaller in restore compared to control by 15 Pg C, resulting from the lower atmospheric CO 2 concentrations, leading to a reduced disequilibrium between ocean pCO 2 and the atmospheric CO 2 concentrations. This negative feedback limits the impact of restoration on atmospheric CO 2 . The cumulative difference in F OA increases between both runs between 2014-2036 CE before saturating in the second part of the century ( Fig. 4C and Table 2).
The net carbon flux into the atmosphere (F A ) peaks in the mid-2040s before turning negative (Fig. 2C). This is driven firstly 230 by the specified RCP 2.6 fossil fuel emission scenario which peaks at 2020 CE (Fig. 2B). The difference between control and restore is that peak emissions are lower because of the additional forest restoration and halting anthropogenic land use emissions (Fig. 5B). Cumulatively, F A is lower by 18 Pg C in restore than control (Table 2).
This 18 Pg C difference is equivalent to a benefit of 9.5 ppm atmospheric CO 2 in restore. The maximum difference between both simulations is 17.1 ppm at 2037 CE (35.1 Pg C), with an extended period of relatively lower CO 2 concentrations until 235 2070s (Fig. 5C), and seen more clearly as decadal means (Fig. 5D). After the 2070s the difference between control and restore diminishes, caused by a combined decline in ocean and land carbon uptake in restore relative to control. Both can be explained through the carbon cycle response to lower atmospheric CO 2 : a decrease in the difference in partial pressure of CO 2 between atmosphere and oceans, and a decline in plant NPP due to lower CO 2 fertilisation. This results in a CO 2 sensitivity to forest restoration of −0.61 ppm CO 2 per 10 3 Mha restored (Table A3).  Finally, ET and other land surface properties may affect precipitation. While global precipitation increases in control and restore over the first 40 years, this broadly follows the increase in global temperature. Mean land-only precipitation in the tropics over the whole simulation is 2% higher in restore than in control, but as no persistent pattern is seen over the tropical restoration area it is unclear if this is internal variation or a substantive change.

Surface temperature
Restoration reduces global surface temperatures by nearly 0.2 • C in the first 30 years of the simulation as a result from the net impact of radiative CO 2 forcing, albedo and ET changes. However, this difference weakens to become negligible by the 2090s ( Fig. 6B and C). The early temperature differences overlap with the periods with the largest differences in atmospheric CO 2 and before changes in surface energy, due to lowered albedo, counteract the reduced radiative CO 2 forcing later in the century.

265
This suggests that tropical forest restoration has near-term climate benefits, and in combination with fossil fuel mitigation can lower peak warming.
Under both simulations global average temperature stays below 2 • C, but surpasses 1.5 • C relative to pre-industrial during nearly 50 years after 2035 CE (Fig. 6A). As the standard deviation of the decadal variability of HadGEM2-ES is 0.1 • C (Jones et al., 2011), only the short-term impacts of restoration result in a detectable temperature reduction.

270
No persistent spatial temperature patterns are observable over time in comparing control and restore. There is, however, a temperature gradient between sites with increased broadleaf trees and adjacent areas. Grid cells with increased tree cover are up to 0.6 • C warmer than adjacent cells that have no new tree growth, as seen on the tropical restoration boundaries, suggesting that albedo warming is stronger than ET cooling in these cells, causing localised surface warming, which in turn may lead to lower plant carbon uptake and increased soil respiration.

Discussion
Stopping anthropogenic land use in the tropics has two impacts on land cover. First as most deforestation in RCP 2.6 is projected to occur in the tropics over the first half of the 21 st century, 286 Mha deforestation is avoided by the end of the century (9.6 Pg C reduction in emissions). Second, trees replace grassy vegetation during secondary succession, renaturalising 1529 Mha of land. NPP, litterfall, soil carbon, and soil respiration respond to the land cover change. Overall, as NPP declines 280 13 https://doi.org/10.5194/bg-2020-432 Preprint. Discussion started: 30 November 2020 c Author(s) 2020. CC BY 4.0 License. less than litterfall under restoration, tropical total biomass increases by 105.5 Pg C in restore compared to a loss of 28.8 Pg C in control, given a net biomass increase of 134.4 Pg C. However, while often neglected, lower CO 2 fertilisation reduces biomass growth outside the tropics by 4.4 Pg C, and changes in soil carbon dynamics store 96.5 Pg C less in soils in restore over the simulation period, as reductions in R h are lower than those in F SC following restoration. This leaves a net land carbon sink under both simulations, but an additional land carbon benefit from restoration of only 23.4 Pg C. The ocean air-sea carbon 285 exchange responds almost instantaneously to the enhanced land carbon sink through a cumulative reduction in carbon uptake of 15 Pg C over the century following restoration, relative to control. The combined effect leads to a reduction in atmospheric carbon of 18 Pg C between both simulations, equivalent to modest a 9.3 ppm CO 2 by the end of the century, given 1529 Mha were allowed to be restored and 286 Mha additional deforestation was avoided. The largest difference is found in the 2040s (30 Pg C; 14.6 ppm CO 2 ).

290
The net global temperature benefit from the combined effect of lower radiative forcing due to lower atmospheric CO 2 concentrations and the biogeophysical responses to land cover change is up to 0.2 • C in the decades after restoration, but is minimal (<0.01 • C) by the end of the century. Overall, large-scale restoration increases carbon storage on land, and can reduce warming in the decades after restoration. Critically, when combined with aggressive mitigation actions, stopping deforestation and restoring forest in the tropics can limit peak warming later this century. Yet, the carbon sequestration and implied climate 295 benefits are substantially smaller than recent, widely publicised claims of a potential tree restoration carbon uptake of 205 Pg C globally (104 Pg C of that in the tropics) (Bastin et al., 2019, but see Bastin et al., 2020). The difference originates likely from ignoring the full range of Earth System feedbacks.

Restoration timescales and carbon uptake
Generally succession is reproduced in HadGEM2-ES following a grass-shrub-tree cycle, which is likely too slow as obser-300 vations show that carbon accumulation is driven almost immediately by tree growth in most tropical locations. Meanwhile the absence of fire disturbance in HadGEM2-ES means that succession is never interrupted. Given the slow tree growth in the model it is not surprising that the modelled pantropical median biomass increase of 1.8 Mg C ha −1 yr −1 (range 0.5-5.5 Mg C ha −1 yr −1 ; see SI) for first 20 years is lower than the observed net carbon uptake rates over 20-30 years after the cessation of land use for the tropics (2.5-6.6 Mg C ha −1 yr −1 , Bonner et al., 2013;Koch et al., 2019), although values as low as 305 1.5 Mg C ha −1 yr −1 have been reported (Brown and Lugo, 1992). The median uptake rate of 0.5 Mg C ha −1 yr −1 simulated in later decades is in agreement with other studies from different parts of the tropics (Houghton and Nassikas, 2018;Lewis et al., 2009;Phillips et al., 2009;Poorter et al., 2016). Consequently, the modelled biomass after 20 years (41 Mg C ha −1 yr −1 , Fig.   A4) is also smaller than the biomass observed in recovering forest in various regions of the Neotropics (135-150 Mg C ha −1 , Orihuela-Belmonte et al., 2013;Poorter et al., 2016). The median biomass (165.2 Mg C ha −1 ) by the end of the simulation is in 310 the range of reported values for the tropics (100-200 Mg C ha −1 , Saatchi et al., 2011). Initial grid-box biomass, however, was already higher in restore than biomass found in real world post-disturbance monitoring plots, meaning less modelled biomass growth is sufficient to match observations. The low biomass increase leads to an underestimate in carbon accumulation in the model by a factor of up to 2 (adjusting for this would equal a 15 ppm CO 2 uptake in restore compared to control, see Methods). This suggests that assisted restoration, 315 i.e. reducing competition from grasses and shrubs, is preferable from a carbon uptake perspective over more natural restoration approaches.

Tropical restoration and carbon cycle feedbacks
Stopping tropical deforestation has a greater carbon benefit on a per unit area basis (0.034 Pg C Mha −1 ) compared to forest restoration (0.027 Pg C Mha −1 ). Taken together, preventing tropical deforestation of 286 Mha (9.6 Pg C) and the subsequent 320 tropical forest restoration (41.8 Pg C) of 1529 Mha has a carbon benefit of 51.4 Pg C over control by 2100 CE, excluding any feedbacks. This is at the lower end of published uptake estimates when normalised over the renaturalised area (Table 3).
The difference is primarily influenced by the choice of model (e.g. interactive calculation of biomass vs fixed biomass scaling, processes included), CO 2 scenario choice, reference point (time zero, before restoration vs an evolving control simulation) and the type of renaturalisation. This shows that estimates that do not include carbon cycle and climate feedback processes 325 overestimate the carbon uptake potential of natural climate solutions. Our simulations are closest to unguided forest restoration, while other studies calculate uptake rates from either simulated reforestation or assisted natural regeneration (without successional cycle). Assisted natural regeneration obtains higher carbon uptake rates faster but needs to be actively managed (i.e. more expensive) and, if implemented incorrectly, prone to detrimental impacts on biodiversity and other ecosystem functions (Lewis and Maslin, 2018). The soil carbon response is also important, but is often ignored (Bastin et al., 2019) or uncertain 330 (Friedlingstein et al., 2014).
The carbon benefit from tropical forest restoration (51.4 Pg C) is partially offset by the Earth System response to lower atmospheric CO 2 . In particular by a lower extratropical uptake in restore compared to control (18.5 Pg C) and a lower ocean carbon uptake (15 Pg C). This leads to 65% of the carbon benefit being overwhelmed by negative feedbacks. This is larger the 20-50% range found in previous studies employing coupled carbon cycle models (Arora and Montenegro, 2011;Bathiany 335 et al., 2010;Jones et al., 2016;Pongratz et al., 2009;Stocker et al., 2011) and the 53% of anthropogenic emissions taken up by land and ocean carbon sinks over 1990-2018 (Friedlingstein et al., 2019b). The difference is likely to be down the use of dynamically regrowing vegetation vs prescribed land cover, uncertainties in the CO 2 fertilisation effect on plants, the sensitivity of the land carbon to temperature changes, the sensitivity of ocean carbon to changing temperatures and atmospheric CO 2 concentrations, and different time scales (decades to multiple centuries). The actual carbon benefit when considering the 340 Earth System response (18 Pg C) is smaller than the 51.4 Pg C from emission reduction and tropical vegetation uptake alone due to soil carbon, CO 2 fertilisation, and ocean carbon cycle feedbacks. This is important, as other approaches often do not take into account these negative CO 2 feedbacks or the response of the carbon cycle to climate change. Indeed, the most high-profile restoration potential estimate Bastin et al. (2019) includes none of these feedbacks, suggesting a CO 2 benefit normalised by area far higher than all other estimates (Table 3).

345
All carbon fluxes and carbon cycle responses combined add up to a CO 2 reduction of 9.3 ppm at the end of the century. A prior study utilising a fully-coupled ESM and reforesting 270 Mha cropland in the tropics (compared to 1529 Mha renaturalised Table 3. Carbon uptake estimates for tropical afforestation (affor), reforestation (refor), and forest restoration (restor), normalised by area, cumulative uptake by 2100 CE, renaturalised area, type of renaturalisation strategy, and method for calculation, whether carbon cycle feedbacks (CC), biogeophysical feedbacks (BP), CO2 fertilisation (β), or plant temperature responses (γ) were included, successional dynamics, CO2 concentrations in control experiments. ESM = Earth System Model. All values are for tropics only (tropics as specified by each study), except Lewis et al. (2019b) includes some extratropical regions under the Bonn Challenge. Note that Arora and Montenegro (2011) uses the term afforestation for turning present-day cropland that would be forest back into forest, here we label this reforestation. here) under the SRES A2 scenario finds a CO 2 benefit of more than double: 20 ppm CO 2 (Arora and Montenegro, 2011).
Their CO 2 sensitivity to tropical vegetation regrowth (CO 2 benefit normalised by the area returned to natural vegetation, Arora and Montenegro, 2011) was substantially higher than both other studies (Claussen et al., 2001;Bathiany et al., 2010) 350 and the sensitivity found by this experiment (Table A3). The higher normalised CO 2 benefit is due to prescribed vegetation, meaning cropland instantaneously becomes forest rather than is converted through secondary succession, the choice of a higher CO 2 scenario resulting in a greater CO 2 fertilisation of vegetation, and a 30% lower ocean carbon sensitivity to changes in atmospheric carbon in CanESM2 (Friedlingstein et al., 2014), resulting in a lower reduction in ocean carbon uptake following an increased land carbon sink. This demonstrates that ESM estimates of the benefit of forest restoration vary with model 355 formulation and baseline scenario. The sensitivity found here is within the range of earlier studies (Table A3), and employs the most sophisticated ESM and scenario choice so far used for such an experiment.

Temperature benefit to low emissions
The permissible carbon budget to stay within 2 • C (Rogelj et al., 2018) is 445 Pg C between 2006 CE and2100 CE, with 139 Pg C already used (Friedlingstein et al., 2019b). is another factor influencing the carbon benefit in our experimental set up. For example, including processes such as tillage leads to increase historic land use change emissions by 70% (Pugh et al., 2015). The response of vegetation to changes in atmospheric CO 2 depends on the magnitude of CO 2 fertilisation which is limited by nutrient (nitrate and phosphate) availability.

380
Nutrient limitation is not included in HadGEM2-ES, models that include nutrient limitation generally show a weaker CO 2 fertilisation impact on plant productivity (50-100%, Huntzinger et al., 2017;Fleischer et al., 2019) but the magnitude of the effect in the real world is still uncertain (Haverd et al., 2020). The decline of the ocean carbon sink when lowering atmospheric CO 2 is a robust feature in ESMs (Schwinger and Tjiputra, 2018), its sensitivity to changes in atmospheric CO 2 (and temperature), however, is between 39% lower and 30% higher than the carbon sink in HadGEM2-ES over the historical period (Friedlingstein 385 et al., 2014). Comparing these processes with HadGEM2-ES, the greater carbon benefit from cutting higher land use emissions would be balanced by the smaller negative impact of lower atmospheric CO 2 on extratropical carbon uptake, with an uncertain magnitude in ocean carbon uptake change. A coordinated effort exploring the intermodel spread of earth system feedbacks to nature-based solutions (e.g. in the Carbon Dioxide Removal Model Intercomparison Project-CDRMIP, Keller et al., 2018) would be beneficial given the importance of these feedbacks and the policy relevance of nature-based solutions.

Conclusions
We find small temperature benefits from large-scale tropical forest restoration over a few decades, and no impact in the longer term. Tropical restoration reduces peak atmospheric CO 2 concentrations, but the Earth System response to tropical restoration itself offsets nearly two-thirds of the initial carbon benefit from restoration. This work provides further insight into the Earth System response to negative emissions, particularly under a policy-relevant low CO 2 trajectory. Some of these findings may 395 be model-dependent, due to low modelled NPP of broadleaf trees, large modelled changes in albedo, and lower modelled changes in ET. Given the idealised extent of the forest restoration (1529 Mha benefit in the long term and is no alternative to reducing fossil fuel emissions, it can, however, contribute to reducing peak CO 2 concentrations and peak temperatures, which may be critical for societal and ecosystem adaptation. Furthermore negative emissions and lower atmospheric CO 2 concentrations represent a system of diminishing returns. Land and ocean carbon sinks decline as atmospheric CO 2 decreases and thereby erasing up to two thirds of the additional carbon sink from LULCC. While well known in the modelling community Schwinger and Tjiputra, 2018), it is also important to consider 405 these negative feedbacks in estimates on the carbon impact of forest restoration to avoid making misleading statements (e.g. Bastin et al., 2019) on the potential of carbon sinks from tropical restoration (Bastin et al., 2020;Friedlingstein et al., 2019a;Lewis et al., 2019a). In short, the more processes are considered in mitigation estimates for negative emission technologies, dynamic vegetation, climate, and carbon cycle response, the smaller their mitigation potential becomes.
However, this idealised experiment shows that in the short to medium term (∼30 years) carbon uptake from tropical forest 410 restoration, alongside radical reduction in fossil fuel use can provide a valuable additional time until other negative emission technologies become more widely available to remove countries' remaining residual emissions to meet the societal goal of stabilising the climate by reaching net zero greenhouse gas emissions.     Table A3. CO2 and temperature sensitivities to forest regrowth (incl. land use change) in Earth System Models. Area weighted CO2 and temperature for the last 20 simulation years (this simulation and Arora and Montenegro, 2011), the last 30 simulation years (Bathiany et al., 2010), and an average of the last 150 simulation years (Claussen et al., 2001). 26 https://doi.org/10.5194/bg-2020-432 Preprint. Discussion started: 30 November 2020 c Author(s) 2020. CC BY 4.0 License.