Combined biogeophysical and biogeochemical e ﬀ ects of large-scale forest cover changes in the MPI earth system model

A ﬀ orestation and reforestation have become popular instruments of climate mitigation policy, as forests are known to store large quantities of carbon. However, they also modify the ﬂuxes of energy, water and momentum at the land surface. Previous studies have shown that these biogeophysical e ﬀ ects can counteract the carbon drawdown 5 and, in boreal latitudes, even overcompensate it due to large albedo di ﬀ erences between forest canopy and snow. This study investigates the role forest cover plays for global climate by conducting deforestation and a ﬀ orestation experiments with the earth system model of the Max Planck Institute for Meteorology (MPI-ESM). Complete deforestation of the tropics (18.75 ◦ S–15 ◦ N) exerts a global warming of 0.4 ◦ C due to an 10 increase in CO 2 concentration by initially 60 ppm and a decrease in evapotranspiration in the deforested areas. In the northern latitudes (45 ◦ N–90 ◦ N), complete deforestation exerts a global cooling of 0.25 ◦ C after 100 years, while a ﬀ orestation leads to an equally large warming, despite the counteracting changes in CO 2 concentration. Earlier model studies are qualitatively conﬁrmed by these ﬁndings. As the response of temperature 15 as well as terrestrial carbon pools is not of equal sign at every land cell, considering forests as cooling in the tropics and warming in high latitudes seems to be true only for the spatial mean, but not on a local scale.

and, in boreal latitudes, even overcompensate it due to large albedo differences between forest canopy and snow. This study investigates the role forest cover plays for global climate by conducting deforestation and afforestation experiments with the earth system model of the Max Planck Institute for Meteorology (MPI-ESM). Complete deforestation of the tropics (18.75  10 increase in CO 2 concentration by initially 60 ppm and a decrease in evapotranspiration in the deforested areas. In the northern latitudes (45 • N-90 • N), complete deforestation exerts a global cooling of 0.25 • C after 100 years, while afforestation leads to an equally large warming, despite the counteracting changes in CO 2 concentration. Earlier model studies are qualitatively confirmed by these findings. As the response of temperature

Introduction
As greenhouse gas concentrations are increasing rapidly, it is often discussed how car- 20 bon sinks can be generated in addition to emission reductions. In this regard, the terrestrial biosphere plays an important role. It is estimated to have stored about 166 Gt C (about 34% of total anthropogenic carbon) during the last two centuries, while total emissions of 200 Gt C are attributed to deforestation in this period (House, 2002). The Kyoto Protocol takes afforestation into account by considering such changes in car- 25 bon pools. As Pielke et al. (2002) point out, carbon has thus become the currency to assess the human influence on global climate. However, the vegetation cover also affects important parameters of the land surface such as albedo, roughness length and hydrological properties (Nobre et al., 2004;Pielke et al., 1998). With few exceptions, the albedo of forest canopies is lower than the albedo of other vegetation or bare soil (Alton, 2009). Therefore, the net radiation at the surface tends to be larger which acts 5 to increase near ground temperatures. In boreal latitudes, albedo differences are particularly large when snow is present, as the snow cover is partly masked by trees but not by herbaceous vegetation. In the tropics, the influence of forests on the water cycle is also important: Tropical forests are characterised by large evapotranspiration (ET) which acts to cool the surface. Due to deep roots, soil moisture can be returned to 10 the atmosphere more efficiently (Nobre et al., 2004). In addition, trees increase the surface roughness, which leads to larger diffusive fluxes. Without further feedbacks, this would also lead to a cooling because the loss of energy has to be compensated by the surface net radiation. On the other hand, changes in wind speed and direction can lead to circulation changes (Sud et al., 1996), whose impact on temperature 15 is less definite. Which mechanism prevails is a result of many nonlinear interactions and thus critically depends on the imposed changes and the original climate (Pitman et al., 2004). Moreover, these biogeophysical effects are linked to changes in the carbon cycle (biogeochemical effects) by several processes such as the dependence of transpiration on productivity and the dependence of plant physiology and structure on 20 atmospheric CO 2 concentration .
Studies of the net effect of historical land cover change on global temperature have shown that biogeophysical and biogeochemical mechanisms are of the same order of magnitude (Matthews et al., 2004;Brovkin et al., 2006). Therefore, in order to quantifiy the impacts of large scale land cover changes appropriately, both effects should also be 25 included. Claussen et al. (2001) used the intermediate complexity model CLIMBER-2 to implement a complete afforestation and deforestation in different latitude bands. A factor separation yielded a cooling biogeochemical, but a warming biogeophysical contribution of increased forest cover in each latitude. With combined effects they 389 found a temperature decrease (increase) resulting from afforestation (deforestation) in the tropics but the opposite effect in high northern latitudes. This result was confirmed by Bala et al. (2007), who applied a GCM with a coupled carbon cycle (INCCA). Betts (2000) used the radiative transfer model of HadAM3 to estimate the radiative forcing due to afforestation with conifer plantations in boreal latitudes. His geograph- 5 ically explicit calculation indicate that the masking of snow may not be the dominant mechanism everywhere, although on average a mean positive forcing was obtained. Other studies even challenge the warming influence of boreal forests for larger scales: Schaeffer et al. (2006) analysed the possibilities of extratropical afforestation based on socio-economically realistic scenarios for the 21st century. Their study demonstrates that the different time scales of biogeophysical and biogeochemical effects have to be considered. While the decrease in albedo dominates the temperature response in the first half of the century, global mean surface air temperature is reduced in 2100. Bird et al. (2008) developed a conceptual stand-based model and obtain a net cooling from year 40 on when the model is applied to several sites in Canada. As the main 15 reason they identify the high cloud cover in spring which diminishes the surface forcing despite large albedo differences. Further objection is presented in Spracklen et al. (2008), who argue that the emission of cloud condensation nuclei from trees can cause a negative radiative forcing of several W/m 2 due to direct and indirect aerosol effects. Montenegro et al. (2009) used satellite observations to infer the potential net 20 effect of small scale afforestation projects. They came to the conclusion that in all latitudes CO 2 -sequestration is the dominating mechanism with a mean efficiency of 50%. Furthermore, no clear dependency on latitude was found. Considering these results, it seems definite that albedo differences counteract the carbon drawdown of afforestation in boreal latitudes and that biogeophysical effects should also not be neglected in other 25 regions. However, the sign and amplitude of the global mean temperature response remains subject to many uncertainties. In the following, the earth system model of the Max Planck Institute for Meteorology, MPI-ESM, is used to study the sensitivity of the coupled system to large scale changes in forest cover. Apart from Bala et al. (2007), such an analysis has not yet been performed with a fully coupled AOGCM. In addition to the model differences, the longer integration time and the comparison to an anthropogenically undisturbed climate, this study differs from Bala et al. (2007) by considering afforestation as well as deforestation experiments, whereas biogeophysical and biogeochemical contributions 5 are not calculated separately. Section 2 gives a short description of the model and the implementation of the experiments. Results are presented in Sect. 3 for global (Sect. 3.1) and regional changes of physical properties (Sect. 3.2) and carbon cycle effects (Sect. 3.3). In Sect. 4 these results are discussed with regard to previous studies of large-scale land cover changes, while final conclusions are drawn in Sect. 5. 10 2 Model and experiment setup MPI-ESM consists of the atmosphere general circulation model ECHAM5 (Roeckner et al., 2003), the land surface model JSBACH (Raddatz et al., 2007), the ocean model MPIOM (Jungclaus et al., 2006) and the ocean biogeochemistry model HAMOCC5 (Maier-Reimer et al., 2005). ECHAM5 was run in T31 resolution (3.75 • ) with 19 vertical 15 levels, MPIOM with approx. 3 • and 40 vertical levels. JSBACH includes a dynamic vegetation module (Brovkin et al., 2009) which is based on a tiling approach. Within the vegetated fraction of each grid box 8 different plant functional types (PFTs) are considered: tropical and extratropical trees (both deciduous and evergreen), raingreen and cold shrubs, and C3-and C4-grasses. 7 pools of land carbon are distinguished in 20 the model: a green pool, a reserve pool and a woody pool (the sum of these is referred to as living biomass), two litter pools and two soil pools. The photosynthesis scheme is based on Farquhar (1980) and for C4-grass on Collatz (1992). Soil respiration is calculated according to a Q 10 -model and is linearly dependent on soil moisture, which is represented by a "bucket" approach. Physical land surface parameters such as 25 albedo and roughness length are calculated from the individual properties of the PFTs and bare ground, weighted with their cover fractions for each land grid cell.

391
The equilibrium CO 2 concentration in the control run (CTL) amounts to 275 ppm. Orbital parameters were kept fixed at present day values and no anthropogenic land use was prescribed. Compared to observations, forest cover ( Fig. 1) is distributed reasonably well in most parts of the globe (Brovkin, 2009). However, the equilibrium carbon storage ( PFTs but forest types were removed and the total vegetation cover was set to 1 in the areas under consideration except on ice shields. In all cases, the cover fractions of the remaining PFTs were then increased while keeping their relative composition fixed. In the case of boreal afforestation this procedure fails for grid cells, where no forest had previously existed. At these cells, deciduous and evergreen extratropical forest were 20 taken with cover fractions of 50% each. As a result of the applied method, tree cover was expanded towards unproductive regions in the afforestation experiments. Since changes in biogeophysical parameters such as albedo and transpiration were calculated on a basis of changes in the carbon cycle, relative effects of afforestation are expected to be less than the deforestation effects. In all experiments, the new distribution of PFTs was kept fixed in the affected areas. For the other land cells, dynamic vegetation was still active. The land cover change module of JSBACH was used to calculate the respective initial values of the carbon pools. Half of the vegetation carbon of the removed PFTs was relocated to the atmosphere within the first year. The other half was immediately put into the soil pools. Carbon from litter and soil pools was transferred from removed tiles to expanded tiles. At the same time, the carbon densities of expanding tiles were reduced, so that there was no immediate carbon flux between these tiles and the atmosphere. As expected, the time scale of emissions is thus much faster than the time 5 scale of sequestration, because plants have to accumulate carbon according to their productivity.

Changes in global temperature and the carbon cycle
As a result of tropical deforestation, global mean temperature is increased by approx. 10 0.4 • C; the warming in high latitudes is particularly pronounced due to greenhouse forcing (Fig. 3). Because living carbon pools and forest cover are large in the tropics, the CO 2 increase of initially 60 ppm is much higher than in DB (Fig. 4). Changes in the ocean's carbon content are primarily a result of the CO 2 -anomalies and not shown here; Fig. 5 depicts the changes in global terrestrial carbon. It is evident that DT is not only characterised by large primary emissions of approx. 123 Gt C due to biomass reduction, but also by net secondary emissions due to soil decomposition of almost the same amount. In fact, in the end of the experiment, tropical land cells contain 390 Gt C less than in CTL, of which the extratropical regions compensate 190 Gt C. During the first decades, the emissions from tropical soils are almost exactly balanced by the 20 extratropical land and ocean sinks. This explains the constant CO 2 concentration in Fig. 4. As these sinks change on a larger time scale than the tropical land areas, they prevail after some 60 to 80 years and CO 2 concentration decreases. Around the year 90, extratropical land regions alone overcompensate the tropical sink, so that terrestrial carbon increases again. 25 In comparison to DT, terrestrial carbon pool anomalies are small in AT. Tropical land 393 areas take up approx. 0.33 Gt C/yr, decreasing atmospheric CO 2 by only 5 ppm within some decades. Thereafter, net emissions by the ocean and the extratropical biosphere are large enough to balance the tropical anomaly which approaches an equilibrium in the final century. Global mean temperature decreases by only 0.06 • C. Changes in AT are much smaller than in DT, because the converted area is smaller. Besides, climatic 5 limitations play a role, as will be discussed in Sect. 3.3.
Things look quite different in the boreal experiments. In DB only approx. 20 Gt C are emitted instantaneously due to the low carbon storage of living biomass in boreal latitudes. The trend in global terrestrial carbon is close to zero because soil respiration in the cold regions is slow enough to be compensated by an enhanced productivity in 10 the tropics. Therefore, the ocean uptake alone is responsible for the reduction of the CO 2 -anomaly from 10 ppm to 3.7 ppm within the first 60 years. An additional reason for the slow response in the carbon pools of the deforested region is the large proportion of litter. Due to the experiment setup, litter is redistributed to expanding tiles and not put into the soil pools immediately. Because of the permanently increased litter flux during 15 the first decades, soil carbon is first increased before it returns its original anomaly of +20 Gt C at the end of the experiment.
In the case of AB, extratropical as well as tropical regions exchange much more carbon than in AT. Large areas are available especially in the northern regions, so that 20 Gt C are stored in the first 15 years alone. CO 2 -concentration is reduced by 20 5 ppm in this period, later the tropical and oceanic sources almost compensate the sequestration. While pools of living biomass and litter become saturated at +35 Gt C and +30 Gt C in the last decades, soil carbon increases almost linearly in the final 200 years and already contains 40 Gt C more than in CTL in the year 300. As the time scale of tropical soils is shorter, they compensate a considerable part of the boreal 25 sink (more than 60 Gt C by the year 300), about twice as much as the ocean. Despite the CO 2 forcing, global mean temperature increases in AB (and decreases in DB) by approx. 0.25 • C. It is thus evident that biogeophysical mechanisms dominate the global mean temperature response in the boreal experiments.

The tropical energy balance
Local temperature changes cannot be understood without analysis of the surface energy balance (Table 1). In DT a warming occurs despite of the increase in surface albedo. Without any feedbacks, this albedo increase would cause a reduction in net 5 short-wave radiation by approx. 10 W/m 2 . Instead, anomalies lie between −8 and +8 W/m 2 because cloud cover is reduced by up to 0.06, with a mean of 0.028 over tropical land cells. Due to these compensating effects, the increased long-wave emission is the most important contribution to the decrease in net radiation at the surface. Because of warmer and drier conditions, the sensible heat flux is increased in DT, 10 despite the reduced net radiation. The lack of energy is thus balanced by a strong reduction in ET. Changes in AT are smaller than in DT and generally of opposite sign. However, two exceptions are apparent. Firstly, incoming long-wave radiation increases in both experiments. In AT, this is caused by increased cloud cover which overcompensates 15 the reduced greenhouse effect. In DT, CO 2 forcing dominates and despite its decrease towards the end of the experiment, no qualitative changes in the energy balance occur over time (Table 1). The second difference between the experiments concerns the pattern of temperature changes. As Fig. 3 shows, cooling in AT does not take place at every individual land grid cell. In Northern Africa, the albedo difference amounts 20 to up to 0.14 and net short-wave radiation is increased by 30 W/m 2 . This is possible because of the large desert fraction in CTL at the northern boundary of the afforested region. Although ET and cloud cover are increased in Northern Africa, the albedo changes have the dominating influence on 2m-temperature. In Southern Africa cloud cover is not changed as uniformly as in other areas, so that in places with decreasing 25 cloud cover and sufficient albedo changes, temperature does also increase. It is due to these warming regions, that the mean change in sensible heat flux is also positive. 395

Changes in the tropical water cycle
Apart from the exceptions mentioned in Sect. 3.2.1 it seems clear that the anomaly in ET is the main driver of temperature changes in the tropical land areas. This is corroborated by the changes in the annual cycle. Because of the seasonal shift of the ITCZ, precipitation mainly occurs in the summer of each hemisphere. Evaporation closely 5 follows precipitation, while transpiration increases during the wet season when buckets are filled with water and stays high during the dry season until soil moisture becomes too small. In the dry period, transpiration contributes to total ET with more than 90%. In DT, temperature is particularly increased during the transition from the rainy to the dry season, when transpiration is reduced the most (Fig. 6 shows results for South 10 America as an example). In addition, the reduction in ET leads to an increase in moisture convergence during March and August, which also dominates the annual mean. As a result, the positive soil moisture anomaly increases and shows the largest value during the moisture minimum after the dry season. In this time of the year, increased soil moisture can outweigh the reduced productivity of grass so that transpiration is 15 equal or in some regions even higher than in CTL. The temperature increase is then probably only due to the elevated CO 2 concentration. Table 2 summarises the annual mean changes in precipitation, ET and moisture convergence for different regions. In DT, the decrease in relative humidity and increase in surface albedo act to supress convection which explains the reduced precipitation.

20
A large-scale sinking motion is induced in the mid-troposphere over most tropical land cells, while a rising anomaly occurs over the surrounding tropical oceans with exception of the South Atlantic. In AT, opposite changes are obtained. Table 2 also documents that changes in the centre of Amazonia are different from those in other regions. In central Amazonia, the reduction in precipitation is large enough to exceed that of ET, 25 so that mean moisture convergence and soil moisture decrease. The precipitation decrease during the dry season (Fig. 6) is mostly attributable to this area. Also, ET is decreased in the second half of the year, so that the warming is even stronger during this time, with up to 4 • C in September. The strong annual mean warming in central Amazonia which is evident in Fig. 3 is the result of these differences in the seasonal cycle. In contrast, the surrounding areas show an increase in soil moisture and moisture convergence as is shown in Fig. 7. While they dominate the spatial mean in DT, this is not the case in AT. Therefore, temperature in AT is mainly decreased (and soil 5 moisture increased) in August and September. Because of the smaller size and the dryer climate of the afforested areas, changes in AT are much smaller than in DT.

The boreal energy balance
In the boreal experiments, temperature is mainly affected by albedo changes due to snow masking. The effect is strongest in spring when solar insolation is already large 10 while snow cover is still high. As a result, land surface temperature anomalies are at a maximum during this season and reach 3 • C on zonal average. In DB, snow melt is delayed because of a lower spring temperature, while in AB it occurs earlier than in CTL (Fig. 8). The mean albedo anomaly is larger in AB than in DB, as snow cover and converted area in the northern parts are larger. It is due to the lower insolation 15 that temperature does not change more than in the deforestation case. However, the temperature anomaly in AB stays large until summer, because snow melt occurs later in the north. Due to the albedo changes the upwelling short-wave radiation is the largest contribution to the changes in net radiation (Table 1). In contrast to the tropics, the long-wave 20 fluxes almost cancel each other. Annual mean changes in cloud cover are also low: DB shows a decrease by 0.003, AB an increase by only 0.001, which is both about one order of magnitude below the changes in the tropical experiments. However, annual mean short-wave insolation does change because of seasonal and spatial differences in cloud cover. In winter and spring, cloud cover is decreased by deforestation (in win- 25 ter particularly at the southern edge of the affected area, so that it still exerts a radiative effect), in summer increased.
The cloud cover changes are in line with changes in the latent heat flux. In DB, the 397 growing season is delayed so that transpiration is reduced by approx. 0.15 mm/day in May. Evaporation is reduced by more than 0.1 mm/day in early spring because of lower temperatures. In summer and autumn, an increase in moisture convergence leads to higher soil moisture, so that ET hardly deviates from the control climate. This is in analogy to the situation at the end of the dry period in DT. Similar changes apply for AB 5 but with opposite sign. Annually averaged sensible heat flux anomalies are smaller but of the same sign as the changes in latent heat flux. Outside the latitudes where land cover change was imposed, zonal mean changes in radiative and diffusive fluxes are small. 10 The locally induced temperature changes are subject to feedbacks on a larger scale which concern sea-ice cover as well as the oceanic and atmospheric circulation. Although the changes in sea ice cover found here show large fluctuations, differences between the experiments are apparent: In DB, even in the Arctic Ocean the relative increase of sea ice cover is only 0.6% and not significant in most places in the years 15 101-300, although ice volume increases by 10.7% in this period. In the afforestation experiment, annual mean sea ice cover in the Arctic Ocean is decreased by 2.4% in the years 101-300 (1.65% in 151-300). The slightly stronger response in AB is most probably due to the fact that the largest albedo changes over land are located in more northern areas than in DB. The difference between DB and AB is most discernable in 20 autumn ( Fig. 9): Although the cooling over the Arctic Ocean in DB is stronger than over the adjacent land in most areas, the geographical pattern of the largest temperature anomalies resembles the short-wave forcing. This is not the case in AB, where the strongest warming occurs in proximity to the Arctic Ocean. Changes in meridional overturning circulation (MOC) are more pronounced than 25 those in sea ice cover. As Fig. 3 illustrates, the temperature response is weak and in some places even reversed in the north-western Atlantic in both experiments. In DB, the mass flux below 1000 m is increased by 1.5 Sv from 15.7 to 17.2 Sv at 30 • N on average over the final 200 years; in AB it is reduced by 1.5 Sv. At 60 • N however, overturning is strongly enhanced in DB (from 5.7 Sv to 6.5 Sv), but hardly affected in AB (−0.1 Sv). This may be related to the decrease in ice cover in AB, so that larger buoyancy fluxes than in CTL are obtained. The boreal land cover changes also influence atmospheric circulation. In DB the ver-5 tically extended cooling leads to an increase in baroclinity and thus in wind speeds in temperate northern latitudes, especially in spring. In 200 hPa, a zonal mean increase of up to 1 m/s is obtained. In contrast, high level wind speed in the subtropics is decreased in this season over the Atlantic and the Mediterranean. In AB, corresponding zonal mean anomalies are found. However, characteristic deviations from these zonal averages occur. The spatial heterogenity of the surface flux anomalies leads to barotropic Rossby wave patterns which show seasonal variations. In particular, an anticyclonic (cyclonic) anomaly occurs over Southeastern Europe in DB (AB) in spring which shifts to the east during the year. This shift indicates an interaction with a natural pattern: In CTL an anticyclone is simulated over the North Atlantic in spring (Azores 15 high) which shows a similar shift and therefore seems to be extended to the east in DB and contracted in AB.

Regional changes in carbon pools
The climatic changes discussed above affect the local carbon pools. In addition, plant productivity is altered by the exchange of PFTs and changes in CO 2 concentration. 20 Therefore, anomalies in total land carbon are not of the same sign everywhere in the tropical or boreal regions (Fig. 10). By rescaling the carbon pools of the individual tiles in CTL according to their altered cover fractions, the contributions of the PFT exchange alone can be obtained for each experiment. The difference to the actual changes is then due to climate and CO 2 feedbacks (neglecting all synergies). These feedbacks 25 can be further separated by inferring changes in NPP and soil respiration with the same method. On the basis of this analysis, Fig. 10 can be interpreted in a more profound way: In the dry regions of Northern Australia and Northern Africa, as well as in the 399 north-east of South America, grass is simulated to be more productive than forest, so that deforestation leads to an increase and afforestation to a loss in soil carbon. In addition, the decreased moisture convergence in AT is accompanied by reduced soil moisture. As a result, total tropical soil carbon in AT decreases, despite the fact that more wood carbon tends to increase the turnover time. In the moist tropics, the increase (de-5 crease) in soil moisture in DT (AT) acts to increase NPP with the exception of central Amazonia. Changes in soil moisture also have the largest impact on soil respiration in both experiments, so that respiration changes generally counteract the climatic effects on NPP. For example, in DT central Amazonia shows an increased turnover time even though the strongest warming occurs in this region. The extratropical sink in DT is 10 due to CO 2 -fertilisation, higher temperatures and, in some regions, a northerly shift of boreal forest. In the boreal areas simulated differences between grass and forest NPP are small in CTL. The very small amount of secondary emissions in DB is in line with this expectation; soil carbon has not at all contributed to secondary emissions by the year 300. In contrast to the tropics, the impact of temperature on NPP outweighs the 15 effect of increased soil moisture except for the southern edge of the affected area in Central Asia. In AT, the reduced carbon storage of some land cells in Eastern Asia and Western North America is due to moisture induced soil respiration.
The spatial averages of carbon pools in and outside the areas of land cover change are presented in Table 3. For soil carbon, the feedback calculation explained above 20 was applied. The separated impact of the redistribution of cover fractions is refered to as uncoupled. In DT, all contributions act to decrease the tropical carbon pools which explains the large secondary emissions in this experiment. The increased soil respiration overcompensates the enhanced productivity of grass in the tropics. However, in AT feedbacks also lead to a loss in soil carbon, because of increased soil respiration in 25 Amazonia and larger water stress in other tropical areas. In addition to the low productivity differences, feedbacks are also responsible for the small emissions in DB. It has to be considered though, that soil carbon pools have not reached an equilibrium by the year 300 so that the absolute feedback contribution would further decrease in DB as well as in AB.
As an afforestation project in reality will be most efficient in terms of carbon storage where a converted area has the strongest possible carbon uptake, the changes in carbon storage are also calculated with reference to the absolute afforested or deforested area only (Table 4). In the tropics, deforestation still has a much larger impact on car-5 bon storage than afforestation because of the climatic limitations in the dry regions. In boreal areas, living biomass and litter are also less affected in AB than in DB. However, additional carbon is stored in the soil. It is evident that in AT and DB, where changes in soil carbon counteract the changes in biomass, the soil anomalies are smaller than in the corresponding experiments DT and AB. The difference between afforestation and 10 deforestation illustrates, that for the whole area afforestation still leads to a gain and deforestation to a loss in soil carbon. Moreover, Table 4 documents that in all experiments the global mean anomalies are of the same sign as the immediate change in biomass, despite all productivity differences and feedbacks.

Tropical experiments
Many GCM studies on the biogeophysical impact of large-scale deforestation in tropical regions have been conducted; a selection is presented in Table 5. Most of these studies show an increase in temperature and a decrease in ET and precipitation, which is in line with the results presented above. In addition, the large-scale sinking in response 20 to the decrease in net energy at the surface is in accordance with expectations (Mylne and Rowntree, 1992;Dirmeyer and Shukla, 1994). The characteristic changes in surface energy balance and cloud cover also agree with CLIMBER-2 results , with the exception that downwellig long-wave radiation is increased in DT due to CO 2 emissions. Changes in moisture convergence and the annual cycle  Sud et al. (1996) find the largest decrease in Amazonian precipitation during the rainy season, whereas in DT the dry period is intensified in central Amazonia. The comparison with the studies listed in Table 5 is subject to many uncertainties which are introduced by the choice or calculation of surface parameter values (Mylne and Rowntree, 1992;Pitman et al., 1993;Dirmeyer and Shukla, 1994), the treatment of the ocean, different integration times and the definition of areas (McGuffie et al., 1995), as well as the applied model. In addition, no change in CO 2 concentration was accounted for in any of these experiments. However, these differences may not explain why the changes in central Amazonia are much larger than elsewhere. By also using MPI-ESM, but without taking carbon cycle effects into account, Brovkin et al. (2009) find a similarly pronounced warming in Amazonia after global deforestation. Thus, these regional differences might rather be related to regional characteristics such as the amount of water recycling, which is known to be large in Amazonia (Nobre et al., 2004). For example, Pitman et al. (1993), McGuffie et al. (1995 and Zhang et al. (1996) also find the largest reduction of moisture 15 convergence and the strongest (or only) temperature increase in Amazonia (Table 5). Sud et al. (1996) find a large warming and an exceptional decrease in moisture convergence in Amazonia as compared to the tropical mean. Also, Claussen et al. (2001) obtain the strongest warming in South America with CLIMBER-2.
In South East Asia, the water balance is less influenced by local recycling but rather 20 by the monsoon circulation (Zhang et al., 1996). Changes in precipitation and ET are smaller than in Amazonia and the largest relative increase in moisture convergence after deforestation is obtained. In comparison to tropical Africa, absolute ET reduction in DT is larger while the temperature change is similar. The studies listed in Table 5 even report a cooling due to deforestation. Delire et al. (2001) obtained a reduction of SSTs 25 due to increased upwelling. As CO 2 concentration is increased in DT, this cannot be expected here. However, a westward anomaly of ocean surface speed occurs during boreal spring and summer due to strengthened trade winds. The increase in SST is relatively low to the west of Indonesian land cells during this period.
In addition to these regional differences, the deforested continents may not be independent of each other. In Amazonia, the anomalies in moisture convergence (Fig. 7) hardly extend to the Pacific because of the Andes, as is also discussed by Lean and Rowntree (1997). In this regard it seems plausible that in Africa the anomalies can affect the nearby ocean more easily. In Fig. 7 a dipole structure is evident over the tropical 5 Atlantic with more moisture convergence north and less moisture convergence south of the equator. This pattern is inherited by precipitation changes between July and October, which reflect a northerly shift of the ITCZ. Anomalies at the eastern boundary of South America could then propagate inland and add to the locally induced changes. However, to determine the extent to which the Amazonian climate is influenced by land 10 cover changes in Africa further studies are needed.

Boreal experiments
With regard to the albedo induced cooling in spring and early summer, other GCMstudies of large scale boreal deforestation are qualitatively corroborated. Thomas and Rowntree (1992) as well as Chalita and LeTreut (1994) analysed the impact of albedo 15 differences between closed snow cover and snow-covered forest. The temperature changes in MPI-ESM are similar to their results. Douville and Royer (1997) used the ARPEGE climate model and additionally considered the change in roughness length. They found a temperature maximum of more than 3 K in April and May between 50 • N and 60 • N, similar to MPI-ESM. However, the duration of the cooling is much less 20 than reported by Douville and Royer: While they found a cooling of more than 2 K between December and June in the same latitudes, a comparable anomaly only persists for 3 months in MPI-ESM (Fig. 8). One reason might be the delayed snow melt in ARPEGE. It must be considered though, that in all three studies there was no coupling between 25 atmosphere and ocean. Bonan et al. (1992) and  showed that if taking the ocean into account, the cooling is stronger, more evenly distributed over the year and geographically more extended. Bonan et al. (1992) found a cooling of locally 403 up to 5 K in July and 12 K in April. Snyder et al. (2004) obtained a cooling of 2.8 K (6.2 K in MAM) even with fixed SSTs because of a strong increase in low level cloudiness. As MPI-ESM includes the interaction with the ocean, the simulated temperature decrease in DB is much less than anticipated by comparison with other studies. The choice of replacement vegetation, changes in ET and the additional CO 2 of 5 ppm can not ac-5 count for this difference. In contrast, large discrepancies to other models are apparent in the short-wave fluxes. Douville and Royer (1997) find a reduction of net short-wave radiation at the surface of more than 20 W/m 2 between 50 • N and 65 • N in March and April. In MPI-ESM the maximum value amounts to only 15 W/m 2 . Bonan et al. (1995) and Thomas and Rowntree (1992) report an annual mean decrease in net radiation of As these values are generally averaged over different areas and time periods and represent different components of the surface radiative balance, the comparison cannot be perfectly consistent. However, it seems obvious that the sensitivity of the short-wave radiative balance to land cover change in high 15 latitudes is comparatively weak in MPI-ESM. The reason seems to be related to the parameterisation of surface albedo: In comparison to observations from the BOREAS study, presented in Betts and Ball (1997), the albedo of snow-covered forests is too high in MPI-ESM. This is especially the case for deciduous forest, for which Betts and Ball find an albedo of 0.21, while in the model the values lie in the range of 0.4-0.7. 20 For boreal evergreen forest, Betts and Ball state a mean of 0.13; in the model a range of 0.2-0.45 is found. Even in the boreal afforestation case, where maximal forest cover is assumed, albedo values are between 0.2 and 0.25. Also, measurements indicate that snow masking by deciduous and evergreen forest is similar (Betts and Ball, 1997;Robinson and Kukla, 1984). This is apparently not the case in MPI-ESM, although 25 stems and branches are accounted for by a stem area index.
It seems plausible that the small change in sea ice cover in the boreal experiments is at least partly due to the weak temperature response in MPI-ESM.  found a 20% increase in global sea ice cover due to boreal deforestation in CLIMBER-2. They also came to the conclusion that the thermal lag of the ocean and the ice-albedo feedback are the main reasons for a cooling in summer. This seems to be true for the autumn in AB, but not DB (Fig. 9). Results are not as distinct for the summer months. In contrast, the MOC increase of 1.5 Sv in DB agrees well with the 2 Sv obtained by Ganopolski et al. (2001). However, the diversity of model differ-5 ences makes a causal assessment difficult. This also applies for atmospheric circulation changes. It is striking, that Douville and Royer (1997) also found a disturbance of the formation of the Azores High due to boreal deforestation. However, the similarities to their results are very limited, as Douville and Royer (1997) found a delayed rather than an earlier high. Also, they report weakened westerlies over Northern Europe and 10 Russia, a southward shift of the North Atlantic westerly jet, increased surface pressure in high latitudes and reduced surface pressure over the mid-latitude Atlantic. All these features are of opposite sign in this study. A similarity can be found in the weakening of the Indian summer monsoon winds in DB, although this is not accompanied by a reduction in precipitation as in Douville and Royer (1997).

Global sensitivities
In order to assess the impact of forest cover changes independently of the area's size, Table 6 presents the sensitivity of global mean temperature and CO 2 content to changes in forest area. These are compared to Claussen et al. (2001) who present an analogue analysis for CLIMBER-2 in Fig. 1  Although CLIMBER-2 and MPI-ESM differ in many respects, the results are qualitatively similar. Nonetheless, with the exception of tropical deforestation, CO 2 sensitivities are smaller in MPI-ESM. In AT, the unproductive climate of dry regions prevents 25 a larger CO 2 uptake, a feature that may not be as important in CLIMBER-2 because of the low resolution. In the case of boreal afforestation it must be considered that the carbon cycle had not yet reached a new equilibrium due to the shorter integration time. 405 The low CO 2 -sensitivity of MPI-ESM to boreal deforestation is due to the low carbon storages of above ground biomass. However, this deficiency seems to be cancelled by the weak snow masking so that temperature sensitivities in high latitudes are similar in both studies. In the tropics, the differences in the sensitivity of temperature seem to reflect those in CO 2 content. As CO 2 concentration in DT is still decreasing in the year 5 300, sensitivities for this experiment are probably not well comparable. In comparison to Bala et al. (2007), temperature changes are small in both deforestation experiments. The global warming of 0.7 K in the year 100 after tropical deforestation in Bala et al. (2007) may be related to larger carbon pools in INCCA. The corresponding CO 2 anomaly of 199 ppm is approx. 4 times larger than in MPI-ESM. In 10 response to boreal deforestation, CO 2 anomalies are similar in both models, but albedo changes and thus global cooling are much more pronounced in INCCA. Nonetheless, the signs of global mean temperature changes obtained here are in line with the results of Bala et al. (2007), and for boreal latitudes also with Betts (2000). However, uncertainties remain concerning the role of individual processes. Claussen et al. (2001) 15 as well as Bala et al. (2007) find a cooling biogeophysical contribution of tropical deforestation. The latter suggest that the albedo change dominates over the reduced ET in its impact on temperature because the ET of grass in INCCA is comparatively high. As the changes in tropical surface temperature and the energy balance demonstrate, this is not the case in MPI-ESM. Claussen et al. (2001) also obtain a warming at 20 tropical land cells, even though it is counteracted on a global scale by the diminished greenhouse effect, resulting from reduced ET. In agreement with this,  report a warming over tropical land and a cooling over the oceans. As only coupled experiments have been conducted here, it cannot be determined whether this holds true for MPI-ESM. 25

Summary and conclusions
The experiments presented here show that in MPI-ESM forests tend to warm the surface in high northern latitudes but act to cool the surface in the tropics. Earlier model studies are corroborated by this result. A global temperature change of +0.4 • C and a CO 2 anomaly of initially 60 ppm is obtained after tropical deforestation because of 5 large primary and secondary carbon emissions. In addition, the strong reduction in evapotranspiration leads to a pronounced warming in tropical land areas. While this is in line with the majority of biogeophysical GCM-studies, changes in the annual cycle and moisture convergence remain uncertain. For tropical afforestation, results are opposite to the deforestation experiment, but of much smaller magnitude, because 10 productivity remains low in dry areas. CO 2 as well as global mean temperature are therefore hardly affected. In high latitudes the snow masking of trees in spring dominates the temperature response, although this effect is weaker than in other models. Changes in sea ice cover, meridional overturning in the ocean as well as atmospheric circulation modify the temperature anomalies, but the contribution of these feedbacks 15 is also model dependent. Primary and secondary emissions are low in DB compared to DT because biomass pools and productivity differences between grass and trees are small. In the case of boreal afforestation, the large carbon sink is quickly cancelled by the ocean and tropical forests. CO 2 anomalies therefore do not exceed some ppm in both boreal experiments. Despite many model differences, the order of magni-20 tude of global temperature and CO 2 sensitivities is similar to CLIMBER-2 results from Claussen et al. (2001). For some land cells, a negative relation between forest cover and carbon storage is obtained, although they do not dominate the spatial mean in any experiment. Also, local temperature changes opposite to the global mean occur due to local differences in 25 surface properties or feedbacks. This is mostly the case in Africa, where afforestation resulted in a warming in places with high surface albedo. Because of the dry conditions there, an afforestation would certainly not be feasible in reality. Nonetheless, as realistic 407 afforestation or reforestation projects must always be confined to much smaller areas than considered here, the spatial mean sensitivities do not apply in such cases. The magnitude and even the sign of biogeophysical and biogeochemical effects may then depend on the location. Therefore, studies such as Bird et al. (2008) and Montenegro et al. (2009), who challenge the idea of a warming boreal forest might be valid on 5 a local scale and different from large-scale experiments.
In addition, many model limitations exist. In high latitudes, these primarily consist in the albedo of snow-covered forest. The representation of soil moisture as a single bucket neglects many important aspects such as the root depth of different plants. This may be particularly inadequate in the tropics, where changes in the water cycle are es-10 sential. As soil moisture has shown a large impact on productivity and soil respiration in the experiments, these uncertainties also affect the carbon cycle. In addition, neither the actual size of soil carbon pools nor the dependencies of NPP on temperature, soil moisture and atmospheric CO 2 as well as the dependency of soil respiration on temperature and moisture are constrained well and thus differ among models (Friedlingstein 15 et al., 2006). As these mechanisms can act in opposite directions, the net effect on carbon pools may thus also be model dependent. In order to assess the impacts of forest cover changes in a more appropriate way, a better quantification of these effects from observations is therefore essential.   Table 5. Changes in surface temperature (T ), precipitation (P ), evapotranspiration (ET) and moisture convergence (P -ET) in Amazonia (Am), tropical Africa (Af), Indonesia (In), Southeastern Asia (SEA) and the whole tropics (Trop) in previous (biogeophysical) model studies of large-scale deforestation. Zhang et al. (1996) refer to surface air temperature instead of ground surface temperature. In Pitman et al. (1993) this information is not definite.