3.2.1 Changes in mean near-surface temperature

Six of the nine models simulate a statistically significant decrease in global near-surface air temperature in response to large-scale deforestation. Results of statistically significant changes (Table 1) range from −0.11 to −0.55 ^∘C (multi-model mean −0.33 ^∘C) globally as simulated by BCC, UKESM, CanESM, CESM2, CNRM and EC-Earth, while MPI, IPSL and MIROC show no significant changes on the global scale. Over areas of deforestation (ΔTas over ΔF) the cooling is stronger (−0.33 to −1.03 ^∘C, multi-model mean −0.40 ± 0.42 ^∘C). Globally averaged, BCC simulates the weakest response as a consequence of balancing regional patterns (Fig. 2). Globally and regionally, MIROC shows almost no response of ΔTas to the deforestation forcing since the rapidly regrowing, secondary vegetation is very similar to the original land cover.

Figure 2Spatial patterns of near-surface air temperature (ΔTas) responses averaged over year 50 to year 79. Only statistically significant changes at the 5 % significance level are shown (modified t test, Zwiers and von Storch, 1995). Contours depict the areas of deforestation (Fig. 1).

The global net decrease in air temperature is dominated by the changes over the oceans and in the Arctic (Fig. 2). Using the surface energy balance (SEB) decomposition approach, we can analyse the contribution of varying energy fluxes to the change in surface temperature (ΔT_surf, Fig. 3f). ΔT_surf is directly related to the balance of surface energy fluxes. However, it might deviate from ΔTas at 2 m height (see also Winckler et al., 2019a) and is therefore shown in Fig. S3 (dashed black lines), which provides a model-wise SEB decomposition. The contributing fluxes are displayed in Fig. S4 (including cloud cover and full and clear-sky longwave radiation).

Figure 3Zonally averaged surface energy balance (SEB) decomposition component-wise for every model after deforestation (averaged over year 50 to year 79) including only areas of deforestation (ΔF). Changes in the surface temperature (T_surf) are expressed as the contribution of changes in available energy (incoming and reflected shortwave and incoming longwave, a to c) and turbulent heat fluxes (latent and sensible, d and e). T_surf is derived from the SEB decomposition method; T_{surf_model} is simulated by each model; The difference of T_{surf_model} and T_surf represents the residual flux accounting for the ground heat flux and subsurface heat storage. An approximated running mean over 10^∘ latitude was applied to smooth lines. The multi-model mean is only applied to latitudes at which all models simulate changes. Note that MIROC was not included due to only minor responses to the deforestation signal. A model-wise SEB decomposition including the simulated near-surface temperature (ΔTas) and results from MIROC can be found in Fig. S3.

Download

In the mid- to high northern latitudes all models simulate an increase in albedo in response to deforestation, which induces a cooling (Fig. 3a). This increase in albedo mainly originates from the reduction in snow-masking effect of forests allowing for a denser and longer lasting snow cover towards summer over grasslands that replace forests. Some models even simulate non-local effects: in CESM2, BCC and EC-Earth this effect is carried beyond the geographical regions of deforestation, and in CanESM and UKESM the geographical extent of the cooling is amplified due to a positive sea-ice–albedo feedback over the Arctic Ocean (see Fig. S5). Longwave radiation is reduced across all models northward of 35^∘ N mainly as a result of reduced surface temperatures leading to less atmospheric trapping and re-emission of longwave radiation (Zeppetello et al., 2019). This effect dominates over the impact of increasing cloud cover over these latitudes in UKESM, EC-Earth, BCC and CNRM, which contributes with a longwave warming (see Fig. S4 for zonal fluxes, Fig. S6 for total cloud cover and Fig. S7 for downward longwave radiation). UKESM, IPSL and CESM2 produce a “warming blob” in the North Atlantic which in turn enhances sea surface evaporation (Fig. S8) and latent heat fluxes (not shown), possibly due to the increased moisture demand of the atmosphere. This result is in line with the reversed finding by Rahmstorf et al. (2015), who found a “cooling blob” due to the freshwater input from the Greenland ice shield caused by global warming slowing down the meridional overturning circulation.

All models simulate reduction in available energy (due to reduced net shortwave and incoming longwave radiation) over areas of temperate (50 to 23^∘ S and 23 to 50^∘ N) and boreal (50 to 90^∘ N) ΔF (Fig. S4c), which dominate the reduction in temperature response, leading to cooling. The effects of increasing albedo (Figs. 3a and S5) are stronger than the reduction in longwave radiation (Fig. 3c). While net shortwave radiation reduces (Fig. 4c), the incoming shortwave radiation increases north of 40^∘ N (Fig. S3b) because the reduced evapotranspiration lowers the atmospheric water vapour content, and this increases the transmissivity of solar radiation through the atmosphere. In the MPI and IPSL models, reductions in cloud cover (Fig. S4f) contribute to enhancement of transmissivity. With less net radiative energy entering the system, less energy is available for the generation of turbulent heat fluxes (latent plus sensible heat, Fig. S4a). At these higher latitudes all models except CNRM simulate decreased latent heat fluxes as not only forests are replaced by less evapotranspirative grassland, but also the atmospheric moisture demand due to the surface cooling and less moisture supply by precipitation reduce this flux. Similarly, most models simulate reduced sensible heat fluxes as a consequence of reduced surface roughness and weaker vertical mixing. Only MPI and MIROC increase the sensible heat flux to balance the greater temperature gradient between the surface and atmosphere following the roughness reduction which is possible as net shortwave radiation is not as much reduced as in other models.

Figure 4Relationship between near-surface temperature changes (ΔTas, averaged over year 50 to year 79) to the final deforestation fraction averaged over all pixels in the (a) tropical (23^∘ S to 23^∘ N), (b) temperate (50^∘ S to 23^∘ S and 23 to 50^∘ N) and (c) boreal (50 to 90^∘ N) region.

Download

The global-scale deforestation-induced cooling is only offset over tropical forests. Here, most models (with the exception of EC-Earth and UKESM) show a warming over ΔF (Fig. 2), since the reduction in evapotranspiration and the decreases in latent heat fluxes dominate the increase in the albedo due to replacement of forests by grasslands (Figs. 3a and S5). However, the geographical patterns differ across models. All models simulate an increase in albedo in the tropics as brighter grasses replaced the darker forests. However, more incoming shortwave radiation (Fig. S3b) due to reduced cloud cover (Figs. S4f and S6) more than compensates for the reduction in incoming shortwave radiation associated with an increase in albedo (Fig. S3a) in IPSL, CanESM, CNRM and BCC. UKESM is the only model that simulates tropical cooling at the surface and at 2 m height, with reduction in incoming shortwave radiation due to increasing albedo more than compensating for the reduction in latent heat and roughness decreases leading to overall cooling. This dominant effect of albedo changes is also observed in HadGEM2-ES, which shares similar model components (Robertson, 2019).

In CESM2, in the equatorial tropics, evaporation increases (Fig. S8). This unintuitive response may be due to the fact that C₄ grasses, which were parameterized for dry regions, are overproductive when they replace forests in the moist deep tropics. This cooling effect is balanced by reduced sensible heat fluxes and increased net shortwave radiation due to less cloud cover resulting in a net warming (Figs. 2c and 3c). Similarly, in EC-Earth evaporative cooling (Figs. 3e and S8i) prevails from 30^∘ N to 50^∘ S since unmanaged grasses show strong increases in leaf area that in turn increase the transfer of soil moisture to the atmosphere. However, this cooling is overcompensated for by the strongest decrease across models and latitudes in sensible heat fluxes (Figs. 3d and S4e) as a consequence of a very low surface roughness of grasses. BCC simulates the strongest temperature increases over the tropical region across all models (Figs. 2f and S3f), leading to a net increase in temperature averaged across all areas of ΔF. In the Amazon region this is mainly caused by an initial surface drying due to reduced evapotranspiration and increased sensible heat flux. This strengthens the circulation over the northern Amazon, supporting increased vertical convection of hot air that in turn causes horizontal advection of moist air from the tropical Atlantic (note that evaporation from the land decreases over the northern Amazon, Fig. S8f) – this behaviour is similar to deforestation responses found in CESM1 (Chen et al., 2019). This leads to increased cloud formation, which increases incoming longwave radiation (with all-sky surface longwave radiation being larger than clear-sky surface longwave radiation, Fig. S7f).

Comparing ΔTas and ΔT_surf (Fig. S3) reveals that there can be large differences among both variables. In CNRM, the surface warming of ΔT_surf, which is dominated by reduced evapotranspiration, is not seen in ΔTas at 2 m height (Fig. S3e). In EC-Earth the effect of reduced sensible heat fluxes causes warming at the surface (Fig. 3i) which is not mixed upwards to the 2 m level (Fig. 2i) where a cooling is observed. MPI and CanESM show the smallest deviations in ΔTas and ΔT_surf.

The SEB approach applied here neglects the ground heat flux on longer averaging periods, subsurface heat storage or changing emissivity. However, inferring the difference between modelled and analytically determined T_surf we see remaining negative differences in energy fluxes at higher latitudes in IPSL, CNRM and EC-Earth (difference of T_{surf_model} and T_surf, Fig. S3). This deviation from the simplifying assumption of our SEB approach assuming zero changes in the above-mentioned properties and fluxes needs further investigation, which is beyond the scope of this paper. Although the impact of the simulation height of T_{surf_model} and calculation height of Tas is done differently across the models (see Sect. 2.3), we cannot coherently attribute the observed gaps between ΔT_{surf_model} and ΔT_surf or ΔTas and ΔT_surf to this. For example, EC-Earth and CanESM defined T_{surf_model} to be at the surface and Tas to be 2 m above the surface, but while EC-Earth simulates clear deviations between these variables and also T_surf, differences are almost negligible in CanESM (Fig. S3).

In four out of nine models that simulate tropical warming and temperate/boreal cooling in response to deforestation the switch in sign of ΔTas from warming to cooling ranges from 11.4^∘ N in IPSL to 34.2^∘ N in BCC (multi-model mean 22.6^∘ N) if changes in ΔTas over ΔF (Fig. S3, black dashed lines) are zonally averaged (Table 1). This change in sign of the temperature response due to the biogeophysical effect is an important metric which indicates that the biogeophysical effects of re/afforestation would result in cooling south of this latitude, in addition to a global cooling effect due CO₂ removal from the atmosphere. Because the other five models show the cooling effect of deforestation at all latitudes due to non-local effects, this estimate is highly uncertain with a standard deviation of at least 10^∘ (Table 1).

Overall, the local response over deforested areas is a reduction of available energy (net shortwave plus downwelling longwave radiation, Fig. S4b) across all models at higher latitudes (by −5 to −20 and −0.5 W m⁻² in MIROC) due to snow-related albedo feedbacks of brighter grasses versus darker trees. At lower latitudes, the models' response is more diverse mainly due to differences in cloud formation (−5 to +10 W m⁻², Fig. S4f). As a result of less energy being available, turbulent heat fluxes reduce across all models and latitudes (by −1 to −12 W m⁻², Fig. S4a). Most models simulate decreased latent heat fluxes over less evapotranspirative grasses (Figs. S4d and S8). Only two exceptions were found where grass parameterizations lead to higher latent heat fluxes (CESM2 and EC-Earth). Most models simulate a reduction in sensible heat fluxes over temperate and boreal grasslands which replace forests as roughness over grasses is lower, which weakens vertical mixing. These findings are in line with those of Winckler et al. (2019b), who find a dominating role of surface roughness for local effects of deforestation. MPI and MIROC simulate an increase in sensible heat as net radiation reductions in these models are small and energy is partitioned preferentially towards the sensible heat flux. In the tropical region, stronger sensible heat fluxes are seen everywhere in response to deforestation, where ΔT_surf increases due to a stronger temperature gradient between the surface and the atmosphere. In CanESM, EC-Earth, and locally in CESM2 and BCC, however, the effect of a strongly reduced roughness outweighs the impact of the temperature gradient.

Previous studies on the temperature effects of large-scale or historical deforestation have shown that the locally induced changes in albedo after boreal deforestation are almost balanced by concurrent changing turbulent heat fluxes. However, the increased boreal albedo can also induce a non-local cooling over land and oceans via advection of cooler and dryer air (Chen and Dirmeyer, 2020; Davin and de Noblet-Ducoudré, 2010; Winckler et al., 2019a). Like in other multi-model studies on the biogeophysical effects of deforestation, it is difficult to separate local and non-local effects without further separation experiments. However, we also find a mean cooling across all models globally and locally over the areas of deforestation. Only MPI, IPSL and BCC simulate weaker non-local cooling effects, thus almost balancing global mean temperature effects of tropical warming and boreal cooling.

Still a key question is how models simulate the impact of deforestation on the turbulent heat fluxes (de Noblet-Ducoudré et al., 2012; Pitman et al., 2009), depending not only on the plant-physiological behaviour (e.g. stomatal conductance, growing seasons, leaf area index) but also on parameterizations of surface roughness and the soil hydrology schemes.

3.2.2 Forest sensitivity (FS) of ΔTas

The sensitivity of the models to the imposed deforestation signal by the end of the simulation period can be quantified in terms of the temperature change (ΔTas) per unit fraction of grid cell deforested (Fig. S9) or per unit area of deforestation (Fig. S10). We therefore call it the “forest sensitivity” (FS) to ΔF.

In the temperate and boreal regions, UKESM, CanESM and EC-Earth show temperature changes of more than −20 ^∘C frac⁻¹; CESM2, CNRM and BCC of −10 ^∘C frac⁻¹; and MPI and IPSL of up to −4 ^∘C frac⁻¹ (Note that the colour bar range is limited and extreme values are not shown). Per 10³ km² of deforestation within one grid cell, UKESM and EC-Earth simulate temperature changes of more than −2.5 ^∘C; CESM2, CNRM and BCC of more than −1.3 ^∘C; and CanESM, MPI and IPSL of less than −0.6 ^∘C 10³ km⁻². In the tropics, BCC and CESM2 show temperature increases of over 4 ^∘C frac⁻¹, MPI and IPSL show temperature increases of less than 2 ^∘C frac⁻¹, and EC-Earth and UKESM show decreases of up to −2 ^∘ frac⁻¹. Per change in 10³ km⁻² forest area, only BCC and EC-Earth show detectable changes of more than 0.5 ^∘C.

However, FS not only reflects local but also the superimposed non-local effects caused by feedback mechanisms. In the tropical region where non-local effects are smaller, we still see some differences in intensity. In particular, CESM2 and BCC and to a smaller degree MPI reveal areas of stronger sensitivity to deforestation in the tropics than elsewhere (up to 8, 6 and 4 ^∘C frac⁻¹, respectively). IPSL, CanESM and CNRM show smaller sensitivities and patterns of coupling also due to the superimposed non-local effects in the latter two models.

To draw more broad conclusions, FS is averaged for every 10 % increase in ΔF per climate zone (Fig. 4). Over tropical regions (23^∘ S to 23^∘ N, Fig. S11), MPI, CanESM, CNRM and BCC reveal increasing warming to increasing ΔF, which weakens and even stagnates in CESM2 and IPSL, respectively. On average these models show a tropical warming response of 0.27 ^∘C frac⁻¹(derived from Fig. 4a). UKESM and EC-Earth simulate increasing cooling with a larger deforestation extent of −0.34 ^∘C frac⁻¹. At higher latitudes (>50^∘ N, Fig. 4c), five models show an increasing cooling with increasing ΔF (mean −1.31 ^∘C frac⁻¹), which is increased by polar amplification. However, MPI, MIROC, BCC and EC-Earth show reverse tendencies at higher ΔF. Over temperate regions (Fig. 4b), there is a more widespread cooling (mean −0.78 ^∘C frac⁻¹) due to mingling effects of different biomes, climate zones and generally smaller forest areas.

Previous studies have argued that the local temperature response to complete deforestation is stronger the smaller the initial forest cover was, and thus non-linear (Li et al., 2016; Pitman and Lorenz, 2016; Winckler et al., 2017).

Only CESM2 and IPSL seem to produce the suggested non-linear, saturating behaviour over tropical regions where non-local effects are smaller (see Fig. 2), and a clear linear behaviour cannot be found with any of the models.

However, drawing conclusions on the (non-) linearity is difficult. In our setup, ΔF reflects the top 30 % of forested grid cells and thus links to the initial forest cover but without capturing the potential effects of completely cleared grid cells, smaller forest fractions, distinct ecozones or isolation from non-local feedback effects.

At a higher level of spatial precision including climate and ecozones, results like the ones presented here could be used to generate lookup tables for climate responses of each model to a given level of deforestation. These would provide computationally inexpensive tools to draw fast conclusions on the climate effects of deforestation in, for example, future land use scenarios. However, in some models the responses show a non-linear behaviour not only to local coupling mechanisms but also due to climate feedbacks acting at the global scale. This superimposed, non-local signal should be isolated for models with strong Arctic amplifications (here CanESM, CNRM, UKESM, CESM2 and EC-Earth) to derive local climate responses. In addition, it would be preferable to use results from longer simulation periods once the models have equilibrated for such lookup tables.

3.2.3 Temporal analysis of ΔTas

The results presented so far do not take into account whether the models have reached equilibrium by the end of the simulation period. Globally, UKESM, CNRM, CanESM and EC-Earth simulate a linear response to the deforestation signal with CanESM, EC-Earth and CNRM showing a continuing downward trend after the end of deforestation, while UKESM stabilizes over ΔF and globally (Fig. 5a). BCC simulates a more or less constant temperature increase over ΔF dominated by tropical warming, though the global signal is a slight cooling. MIROC drives hardly any change in ΔTas for any level of deforestation. MPI, IPSL and CESM2 show only small responses on the global scale due to balancing signals, but regionally ΔTas scales with the intensity of deforestation. Over South America, BCC, CESM2, MPI and IPSL simulate a linear increase while UKESM, CNRM and EC-Earth simulate decreases in Δtas with ΔF (not shown). In the boreal region, all models but MIROC simulate a linear decrease with ΔF over time, which clearly continues after 50 years in CanESM, EC-Earth and CNRM over North America and CanESM over Eurasia.

Figure 5Time series of temperature (a) and precipitation changes (b). Solid lines depict changes over areas ΔF, while dotted lines depict global changes. A 30-year moving average is applied. The black line shows the multi-model mean with 1 standard deviation in shaded grey.

Download

The temporal evolution of ΔTas reveals not only the sensitivity of models to large-scale deforestation but also the strength of non-local high-latitude feedbacks. For most models it would have been beneficial to extend the simulation period to allow the climate variables to reach equilibrium. The two models providing 150 years of data, MPI and MIROC, are less sensitive models without strong feedbacks and thus equilibrate quickly. For UKESM, recovering forests in the remaining parts of the deforested grid cells shape the evolution of the signal.

For models that provided several ensemble members of the deforestation experiment (MPI, IPSL and CESM2) we calculated the time of emergence (ToE). In the tropics, near-surface temperature changes emerge over the regions of strongest deforestation before the end of the first 50 years of the simulation (Fig. 6). Interestingly, the signal propagates from the centre of deforestation to the edges in the tropical zone. In the central tropics, the signal becomes robust (that is, exceeds the signal-to-noise ratio (SNR) of 2) with up to 20 % to 35 % of deforestation still left (Fig. S12). This hints at the advection of temperature changes towards the centre of deforested area due to non-local effects. In boreal zones, CESM2, and to a lesser degree in MPI, demonstrates signals propagating westwards starting from the boreal east coasts with about 30 % and 10 % of deforestation, respectively, still left (Fig. S12). The main attributors here are the westerly winds that carry the modified air by deforestation from the west to the east coast of the continent where the signal is therefore strongest and emerges earlier. In CESM2, Arctic amplification further amplifies this process (see Sect. 3.2.1), leading also to responses outside ΔF. In MPI and IPSL, the advection of temperature changes from neighbouring grid cells is limited to areas of ΔF. In the majority of areas, the signal takes more than 50 years to emerge despite the strong imposed deforestation forcing (Fig. 6 green and blue colours).

Figure 6(a–c) Time of emergence (ToE) of ΔTas and (d–f) equivalent fraction of emergence (FoE) of ΔTas. Only statistically significant areas as found in Fig. 2 are shown; oceans are masked out.

The results of the ToE analysis have to be treated with caution since only a few ensemble members were available, and thus uncertainty remains high. However, following up on the earlier analysis (Sect. 3.2.1), the observed patterns make sense from a causal perspective. After 30 years, all three models demonstrate a propagation of signals from the centre to the edges of ΔF, and two show a westward propagation across the boreal zone. This emphasizes the importance of non-local biogeophysical effects during and after large-scale deforestation (Chen and Dirmeyer, 2020; Pitman and Lorenz, 2016; Winckler et al., 2019b). FoE is more universally applicable across models as the same amount of deforestation can happen at a different time in each model. Notably, while ToE patterns of ΔTas are very diverse across models, the FoE patterns are more alike, especially in the tropics. These results lead to the conclusion that even after large-scale deforestation of one-third to half of the grid cell's forests, the signal only becomes robustly detectable after a few decades as climate variability and mediating effects from the ocean have to be overcome (Davin and de Noblet-Ducoudré, 2010). The applied method based on ensemble trends gives an optimistic estimate of ToE compared to alternative approaches based on multi-ensemble or temporal means relative to the variability of a reference period (Fig. S13, note that SNR was lowered to 1).

Our results have important implications for ongoing land cover changes and climate policies. Between 2001 to 2018, 3.61 Mkm² of forests was cleared (Hansen et al., 2013; and http://earthenginepartners.appspot.com/science-2013-global-forest, last access: 18 May 2020); thus the deforestation rate was about 20 % of that applied in this study. Our results suggest that the detection of climate effects of this recent deforestation would possibly take decades. Likewise, climate response times to the reversal of deforestation as a mitigation measure would be long compared to climate policy timescales.

3.2.4 Changes in precipitation

The global net effect of precipitation changes over ΔF is negative across all models ranging from −10 to −108 mm yr⁻¹ (mean −37 ± 33 mm yr⁻¹, Table 1). All models show shifts of atmospheric patterns over the oceans and distinct changes over ΔF (Fig. 7). Again, MIROC is the least sensitive model with only minor increases over South Africa and Alaska.

Figure 7Spatial patterns of precipitation responses averaged over year 50 to year 79. Only statistically significant changes shown. Contours depict the areas of deforestation (Fig. 1).

Over time, MPI, UKESM and IPSL simulate linear responses to the deforestation signal with only MPI showing global stabilization after ending forest removal (Fig. 5b). Other models exhibit longer time periods (CanESM and MIROC) of continuing positive or negative changes depending on the region. For example, over North America (not shown), CanESM simulates a downward and MIROC an upward trend in precipitation.

The strongest global mean reduction of moisture transfer to the atmosphere via evapotranspiration (Fig. S8) and resulting precipitation over ΔF is found in MPI, followed by UKESM, CanESM and IPSL. Generally, these decreases result from the replacement of forest by less evapotranspirative grasses. Precipitation increases occur mostly outside deforested regions but also on small scales over areas of ΔF in CESM2, CNRM, BCC and EC-Earth for different reasons: in CESM2, C₄ grasses replace forest in tropical regions which are parameterized to be productive under unfavourable climate conditions (e.g. too try or hot) and, hence, are overly productive in tropical zones, leading to transpiration increases. EC-Earth simulates increases in the tropics following increased evapotranspiration (ET) due to a strong increase in leaf area. CNRM is the only model simulating precipitation increases over ΔF at northern latitudes providing moisture for enhanced evapotranspiration during snow-free months. In BCC, local vertical convection in the Amazon region causes horizontal advection of moist air from the Atlantic and west Amazon, which locally increases precipitation there (see Sect. 3.2.1).

Over tropical ΔF, most models show a linear relationship with relative temperature increases correlating with relative precipitation decreases and vice versa (Fig. S14). UKESM shows a linear decrease in ΔPr with a decrease in ΔTas. Due to the above-mentioned model specifications, CESM2 and BCC show a very weak relationship between ΔTas and ΔPr over tropical ΔF, with also positive ΔPr paired with positive ΔTas. Over boreal ΔF, most models simulate decreases in ΔPr correlated with decreases in ΔTas, while in CNRM ΔPr increases despite the cooling air.

Global deforestation affects precipitation by altering circulation patterns and by changing the moisture inputs from the surface to the atmosphere. The SEB analysis (Sect. 3.2.1) demonstrated how new plant types govern the land–atmosphere interaction via turbulent heat fluxes. In most cases we could infer a causal link between changes in turbulent heat fluxes, longwave radiation linked to cloud cover and precipitation, which is in line with previous studies (e.g. Akkermans et al., 2014; Lejeune et al., 2015; Spracklen et al., 2012). While most models simulate moisture decreases as less productive and evapotranspirative grasses replace trees, some models simulate local increases due to advected moisture (e.g. BCC) or favourable parameterizations of grasses (e.g. CESM2 and EC-Earth).

3.3.1 Changes in mean and temporal development of carbon pools and fluxes

Land carbon (cLand, the sum of vegetation, soil and litter carbon) losses range from −169 to −338 GtC until the end of the deforest-glob simulation. Note that CanESM and BCC were excluded in this calculation due to a major divergence from the protocol. By the end of the experimental period the spread across models increases to −144 to −350 GtC, with UKESM and MPI simulating continuing declines and MIROC simulating increases. The multi-model mean decreases from −191 GtC after 50 years (mean over year 45 to 54) to −203 GtC after 75 years (mean over year 70 to 79, Table 1) after the start of the simulation. If only models were included that followed the protocol closely (MPI, CNRM and CESM2), the multi-model mean would be −282 GtC after 50 years and −300 GtC after 75 years. The spatial patterns of ΔcLand are displayed in Fig. S15. The spread across models is the result of several factors that make the model behave differently.

Figure 8Changes in carbon cycle pools over time smoothed by a 10-year moving average. Note that for CanESM and BCC only ΔcVeg could be analysed. ΔcLand refers to the sum of ΔcSoil, ΔcVeg and ΔcLitter. The black line shows the multi-model mean with 1 standard deviation in shaded grey.

Download

For all models but MIROC, it is mainly the changes in vegetation carbon (ΔcVeg) dynamics that dominate changes in cLand followed by changes in litter carbon (ΔcLitter) and soil carbon (ΔcSoil, Fig. 8).

In UKESM dynamic vegetation adjustments in the remaining natural parts of the deforested grid cells drive the continuing decline and weak recovery of ΔcVeg as a consequence of strong cooling followed by stabilizing climate, respectively (Fig. 8a and b). Below-ground carbon is transferred to the fast soil carbon pool (ΔcSoilFast, residence time of a year), which reduces with ongoing deforestation (Fig. S17c). Fast soil carbon decays and thereafter accumulates in the medium soil carbon pool (ΔcSoilMedium, residence time of several decades; Fig. 17b). The development of cLitter and heterotrophic respiration (rh) reductions and the subsequent development of all soil carbon pools correlate with the progression of deforestation and vegetation recovery afterwards. Note that for UKESM below-ground carbon from coarse roots is removed from the system and not transferred to the soil carbon.

In MIROC, because the vegetation type was fixed as woody types in the deforestation, forest recovery started soon after the deforestation process. As a result, the magnitude of ΔGPP is moderately decreasing among the models (Fig. 10a) and ΔcVeg recovers as secondary woody vegetation, leading to the positive large net ecosystem productivity (NEP; Fig. 10d).

CESM2 shows a steep decline in cLand, which is mainly caused by the initial vegetation loss and enhanced fire activity (carbon emissions by fire, ΔfFire) during deforestation of tropical forests due to degradation fires from deforestation (Li and Lawrence, 2017). The sudden initiation of deforestation fires in the Tropics contributes to emissions of 12 GtC yr⁻¹ at the beginning of the deforestation period and levelling off at around 1.7 GtC yr⁻¹ after 50 years (Fig. 10e), with the highest values in the tropics (Fig. S16). These initial deforestation fires cause GPP to drop for the first two decades before the overly productive C₄ grasses in CLM5 (Lawrence et al., 2019), especially in the deep tropics (Fig. 9), start to compensate for the carbon losses. ΔcLand saturates around −342 GtC, with a minor negative drift after deforestation stops. The path is slightly non-linear as the tropical grasses lead to recovery in ΔcLand after deforestation stops. In combination with a constant decline of heterotrophic respiration (Fig. 10c), net ecosystem productivity (ΔNEP, Fig. 10d) turns slightly positive by the end of the simulation period.

Figure 9Spatial patterns of ΔGPP responses averaged over year 70 to year 79. Only statistically significant changes at the 5 % significance level are shown. Contours depict the areas of deforestation (Fig. 1).

Figure 10Changes in carbon fluxes over time smoothed by a 10-year moving average. GPP and NPP are the gross and net primary productivity, respectively; rh is the heterotrophic respiration; NEP is the net ecosystem productivity; fFire denotes the emissions by fire; and NBP denotes the net biome productivity. NBP is based on year-to-year variations in ΔcLand and thus not provided for CanESM and BCC. The black line shows the multi-model mean with 1 standard deviation in shaded grey.

Download

The slight continuing downward trend of ΔcLand in MPI is dominated by changes in the tropics (not shown). As grasses replace trees, ΔNPP is reduced strongly (Fig. 10b), and consequently litter pools are reduced as well (in fact, MPI has the strongest NPP reduction across models with up to −2.2 GtC yr⁻¹ found in the tropics). Grass litter flux is not only smaller in amount, but also of changed quality, leading to faster decomposition. Like in UKESM, the gain of fast soil carbon pool due to root decomposition during the first 50 years has only a minor effect in the long term (Fig. S17). Fire activity is fostered globally as grasses are more fire-prone than trees. In MPI, fire activity is enhanced because of a warmer tropical and globally drier climate (Fig. 10e), slowly diminishing land carbon pools at a similar rate as in CESM2 (∼1.3 GtC yr⁻¹). Regional precipitation reductions cause heterotrophic respiration to decrease even more strongly than ΔNPP, resulting in positive ΔNEP by the end of the simulation period.

EC-Earth and CNRM seem to behave similarly in terms of ΔcVeg, ΔcLitter and ΔcSoil at the global scale. However, regionally the models show fundamentally different responses. EC-Earth simulates higher deforestation rates in the tropics with subsequent higher cVeg loss than in CNRM and vice versa for higher latitudes. Interestingly, ΔcLitter and ΔcSoil increase in EC-Earth globally and CNRM in higher latitudes. In CNRM, grasses produce more below-ground litter fall than trees (due to a higher root-to-shoot ratio), which accumulate, accompanied by lower overall Δrh fluxes, in all soil carbon pools.

In EC-Earth, increases in cLitter and cSoil are partly caused by the deforestation itself since portions of root and, against protocol, leaves and wood biomass are left on-site for decay. In addition, reductions in autotrophic respiration of grasses more than compensate for GPP losses due to deforestation, leading to a positive ΔNPP and thus more litter. This litter is further contained as fire emissions in this model are reduced compared to the previous forest landscape ($\sim -\mathrm{2}$ GtC yr⁻¹ globally). Even the substantial heterotrophic respiration increases due to local moisture input combined with mild cooling can therefore not deplete cSoil.

IPSL shows the smallest response in land carbon, which is dominated by cVeg changes and hardly by any changes in soil or litter carbon pools (Fig. 8). The exception is in central Africa, where the higher NPP of grasses increases the litter flux affecting mainly the long-term soil carbon pool (not shown).

In BCC, tropical changes dominate the global average with the highest observed ΔGPP across models, which is, however, diminished by a similarly high soil respiration flux, resulting in a negative global ΔNEP throughout the simulation period (Fig. 10d). Only in the northern Amazon region, GPP is reduced under high temperatures and despite the observed precipitation increase (Fig. 7). BCC is the only model that simulates cVeg increases outside ΔF in the temperate regions where cooling and precipitation increases overlap, leading to a higher ΔGPP.

From CanESM, we only investigate ΔcVeg and carbon fluxes since carbon was not transferred to the atmosphere as requested by the protocol but to a great proportion left on-site for decay. Still, CanESM shows a very interesting behaviour that diverges from the other models: CanESM simulates a uniform global increase in NPP (Fig. 10b) associated with the highly productive C₄ grasses, especially in the tropics (Fig. 9d). The increase in NPP is accompanied by almost as strong heterotrophic respiration increases (as a consequence of increased litter and soil carbon pools), resulting in net ecosystem carbon gains. GPP changes are, however, not obviously different with mainly less productive grasses everywhere but in Africa and south-east Asia (Fig. 9d), meaning that autotrophic respiration of grasses decreases more than in all other models after deforestation.

The net biome productivity (ΔNBP = Δ_tcLand with Δ_t referring to the year-to-year changes in cLand, Fig. 10f) summarizes the effect of land carbon fluxes. In that, all models show similar carbon fluxes of −4.45 ± 1.07 GtC yr⁻¹ during deforestation except for CESM2, with dominating influences from ΔfFire. After the end of deforestation, the ΔNBP reduction declines on average to −0.45 ± 1.06 GtC yr⁻¹. The outliers MPI (continuous reductions of rh) and UKESM (recovering ΔNPP) show a positive trend, thus initiating a slowdown of the loss in ΔcLand opposed to MIROC in which Δrh increases with the growth of secondary vegetation.

Land carbon changes emerge as a signal within the first 10 years in most places (Fig. S18). MPI and IPSL show more distinct patterns at the outermost edges of deforestation, where 30 years pass before ToE occurs. In IPSL and CESM2, many patches outside ΔF show ToE of up to 50 years; however, changes are smaller than ± 0.5 kg m⁻² in these areas. ToE of ΔGPP (Fig. S19) is more interesting. ΔGPP is uniformly affected by the replacement of trees by grasses but influenced also by the changes in local climate. In MPI and IPSL, the earliest ToE values appear where strong GPP reductions are observed (Fig. 9), while in CESM2 these locations experience strong GPP increases in this time (10–40 years). Changes in climate impose their signal, and thus similar patterns of propagation can be observed as for ToE of ΔTas.

Although forest removal was implemented in a similar way across models, the trajectory and spatial patterns of carbon changes differ strongly. The major part of the land carbon model spread stems from the removal of vegetation carbon based on the differences in initial forest distribution and carbon densities. A similar divergence across multiple models but of lower magnitude was already found for a previous study investigating the effect of future land use and land cover changes on the carbon cycle in CMIP5 models (Brovkin et al., 2013). The changes in ΔcLand (−260 ± 74 GtC) in the deforest-glob simulation are ∼25 % higher than the estimated historical emissions of 205 ± 60 GtC from land use and land cover changes and wood harvest and wood products between 1850 and 2018 (Friedlingstein et al., 2019).

The loss of land carbon follows a similar trajectory at the global scale, with only vegetation recovery (MIROC, UKESM) and grass parameterization (CESM2) causing non-linearities. We find that not only whether fire is represented (MPI, CESM2, EC-Earth) or not can have substantial effects on the NBP and thus overall carbon losses but especially how it is implemented. In CESM2 fire is used as a deforestation tool, while it only depends on litter fluxes in MPI and EC-Earth. In the latter model, fire activity decreases with the expansion of grassland opposed to the other two models. To narrow down the sign and magnitude of fire emissions thus needs further consolidation by incorporating observational data. The protocol allowed models to simulate dynamic vegetation processes outside the deforestation area based on the assumption that the time horizon of the experiment was too short for climate change effects to affect the remaining woody vegetation (Lawrence et al., 2016). UKESM disproves this assumption since forest cover continuously declined in the remaining part of the grid cell. The separation of land carbon pools by land cover type would have therefore been advantageous. Across all models that witness a declining or constant fast soil carbon pool with the onset of deforestation (CESM2, IPSL, MIROC), the fate of below-ground plant materials (roots) remains unclear considering that root biomass is about one-fifth of the above-ground biomass (Lewis et al., 2019).

The analysis of CMIP5 models revealed that substantial uncertainty in model responses was due to implementation differences (i.e. land use patterns, Boysen et al., 2014; Brovkin et al., 2013). Having a very simple experimental protocol of replacing trees with grasses, we now show that the underlying processes themselves also explain large parts of the model spread. Strong or weak model responses may originate from including or not representing certain processes explicitly, e.g. fire activity or soil biochemistry. Our analyses also highlighted the relevance of the comparative response of different vegetation types. While most evaluation is done for total land carbon stocks and fluxes, assessment of land use change requires adequate representation of individual land use/cover types at each location relative to each other. This highlights the need for improving the process understanding of soil carbon dynamics (e.g. Chen et al., 2015; Don et al., 2011; Giardina et al., 2014; Luo et al., 2017), fluxes (Atkin et al., 2015; Huntingford et al., 2017) and biomass carbon stocks (Erb et al., 2017) using observations and field experiments.

3.3.2 Forest sensitivity (FS) of land carbon

Similarly to ΔTas, the FS of ΔcLand is analysed. Across models, MIROC is the most sensitive with local land carbon losses per fraction of deforestation (Fig. S20; note that the extreme values are not shown in the colour bar) of ∼ −95 GtC in boreal North America, followed by UKESM with $\sim -\mathrm{60}$ GtC and other models with −40 to −50 GtC. On average, UKESM, CESM2 and MPI amount to −16 GtC while the other models stay around −9 GtC frac⁻¹ of deforestation. Per 10³ km² of deforestation within one grid cell (Fig. S21), EC-Earth removes on average −2.7 GtC, CESM2 and UKESM remove around −1 GtC, and the other models stay above −0.5 GtC.

The sensitivity of land carbon changes with regard to the deforestation fraction in a grid cell across latitudinal climate zones (Figs. S22 and S23) depends on the initial biomass carbon, soil carbon dynamics, the characteristics of the replacing vegetation and probably even climate.

Most models, except for MIROC and IPSL, show an almost linear decrease in FS in the boreal and temperate region, although the magnitudes vary strongly (Figs. S22 and S23). On average, models decrease cLand by 4.1 and 4.9 kg m⁻² frac⁻¹ in the boreal and temperate zone, respectively. In the tropics, IPSL and CNRM (to a lesser degree MPI and UKESM) simulate on average weaker decreases of around 30 % deforestation than with lower and especially larger forest removals (Fig. S22a). These models including EC-Earth remove very similar amounts of carbon per deforestation amount. However, in EC-Earth, cLand changes in the tropical region decrease above 60 % deforestation. CESM2 simulates a strong negative non-linear behaviour (ca. −9 kg m⁻² frac⁻¹) dominated by vegetation carbon removal in South America (Fig. S23c). MIROC reveals an increasing non-linearity the further north the forest is removed due to the interplay with forest regrowth. On average, the tropical cLand loss is quantified with 5.1 kg m⁻² frac⁻¹.

Although climatic changes affect the carbon cycle negatively via droughts or positively via favourable warming (see also Fig. 9), the main contribution comes from the deforestation itself as also the temporal analysis revealed. Therefore, the carbon response to ΔF is mainly local and almost linear. The FS approach can be used to analyse the effects on land carbon pools in future land use scenarios to derive gross CO₂ emissions. The mechanisms behind FS of ΔcLand may differ across models, and non-linear dynamics from vegetation distribution changes at the local scale can influence the results. For models like MIROC and BCC, we would not apply this approach as clearly drivers from climate or parameterization play a role. Also, the effects of changing atmospheric CO₂ concentrations on the carbon cycle are not captured in this study.

3.3.3 Estimated changes in temperature due to ΔcLand

Carbon emissions from deforestation in the real world act as a greenhouse gas with a potential warming effect. In absence of varying CO₂ concentrations in this experimental setup, we can therefore only approximate the temperature effects of deforestation. The TCRE serves as a good tool to estimate the biogeochemical (BGC) effect on climate from large-scale deforestation (ΔTas in regard to ΔcLand). While the overall biogeophysically (BGP) induced effect of deforestation was a cooling on both the global scale and also over most areas of ΔF, the BGC effect results in an estimated global warming of 0.17 to 0.87 ^∘C. For MIROC, IPSL and MPI (0.18 to 0.57 ^∘C) this is the main temperature response to ΔF (because BGP-induced effects are non-significant), assuming that the TCRE concept allows us to calculate a significant temperature change from any statistically significant change in land carbon pools. Note that the result for CanESM and BCC is only based on cVeg changes. On the global scale, the BGC warming is at least similarly strong (CNRM, UKESM) or 4 times larger (CESM2) than the BGP cooling. When considering areas of deforestation alone, the robust BGP cooling dominates in CanESM, CNRM and UKESM.

The estimate of ΔTas in regard to ΔcLand depends not only on each model's TCRE but also on the sensitivity of ΔcLand in regard to ΔF (see Sect. 3.3.2). Thus, models with strong carbon losses (e.g. MPI) may still have lower climate sensitivities (e.g. 1.6 ^∘C EgC⁻¹) and vice versa, leading to a similar range of results. Although TCRE was shown to be a useful tool regardless of non-CO₂ and aerosol forcings, we here ignore the carbon–concentration feedback in the absence of variable CO₂ concentrations which could potentially enhance the land carbon sink via CO₂ fertilization (Bathiany et al., 2010). Nevertheless, while large-scale deforestation could lead to an overall BGP cooling globally and over ΔF, BGC warming dominates on the global scale with the possibility to balance boreal BGP cooling or enhance tropical BGP warming.