Evaluating the responses of forest ecosystems to climate change and CO2 using dynamic global vegetation models

Abstract The climate has important influences on the distribution and structure of forest ecosystems, which may lead to vital feedback to climate change. However, much of the existing work focuses on the changes in carbon fluxes or water cycles due to climate change and/or atmospheric CO 2, and few studies have considered how and to what extent climate change and CO 2 influence the ecosystem structure (e.g., fractional coverage change) and the changes in the responses of ecosystems with different characteristics. In this work, two dynamic global vegetation models (DGVMs): IAP‐DGVM coupled with CLM3 and CLM4‐CNDV, were used to investigate the response of the forest ecosystem structure to changes in climate (temperature and precipitation) and CO 2 concentration. In the temperature sensitivity tests, warming reduced the global area‐averaged ecosystem gross primary production in the two models, which decreased global forest area. Furthermore, the changes in tree fractional coverage (ΔF tree; %) from the two models were sensitive to the regional temperature and ecosystem structure, i.e., the mean annual temperature (MAT; °C) largely determined whether ΔF tree was positive or negative, while the tree fractional coverage (F tree; %) played a decisive role in the amplitude of ΔF tree around the globe, and the dependence was more remarkable in IAP‐DGVM. In cases with precipitation change, F tree had a uniformly positive relationship with precipitation, especially in the transition zones of forests (30% < F tree < 60%) for IAP‐DGVM and in semiarid and arid regions for CLM4‐CNDV. Moreover, ΔF tree had a stronger dependence on F tree than on the mean annual precipitation (MAP; mm/year). It was also demonstrated that both models captured the fertilization effects of the CO 2 concentration.

, species composition (Renwick & Rocca, 2015), and regional diversity (Beaumont et al., 2011;Garcia, Cabeza, Rahbek, & Araújo, 2014;Ohlemüller et al., 2008) and can even result in shifts from one ecological state to another (Bush, Hanselman, & Gosling, 2010). Such changes may lead to vital feedback in the water and carbon cycles (Gonzalez-Meler, Rucks, & Aubanell, 2014); therefore, it is important to explore how climate change influences the structure and functions of forest ecosystems.
Most projections of future climate change refer to temperature and precipitation changes, as well as increasing concentrations of greenhouse gases in the atmosphere. Temperature is the main influencing factor of many ecosystem processes (Badeck et al., 2004) and the carbon balance. For example, Rustad et al. (2001) used metaanalysis to find that experimental warming of soil temperature in the range 0.3-6.0°C significantly increased soil respiration rate by 20%, net N mineralization rate by 46%, and plant productivity by 19%. Lin, Zhu, Wang, Gong, and Zou (2016) analyzed gross primary production (GPP) and net primary production products during 2000-2010 and leaf area index (LAI; m 2 /m 2 ) products during 1981-2011 and found that the air temperature had a significant positive correlation with LAI (R 2 = .311) and GPP (R 2 = .189). Meanwhile, it has been discovered that the responses of ecosystems to temperature change are spatially heterogeneous and partly uncertain (Mekonnen, Grant, & Schwalm, 2016;Williams et al., 2010;Willis, Bennett, Burrough, Macias-Fauria, & Tovar, 2012). Plenty of work has shown that because of temperature limitations, warming favors boreal forests in the form of increases in vegetation cover (Berner, Beck, Bunn, & Goetz, 2013) and northward movement of tree lines. However, for some tropical forests, temperature has a strong negative effect on stem growth by increasing respiration and decreasing photosynthesis due to reduced stomatal conductance (Schippers, Sterck, Vlam, & Zuidema, 2015). Willis et al. (2012) concluded that when regional conditions become warmer and wetter, the biomass and range distribution of trees are likely to increase, while if a transition to warmer and drier conditions occurs, grass or savanna replaces woody vegetation in many regions.
Precipitation is another vital factor, influencing tree growth (Subedi & Sharma, 2013;Voelker, Meinzer, Lachenbruch, Brooks, & Guyette, 2014) and affecting forest population dynamics (Booth et al., 2012;De Steven, 1991). More precipitation during the wettest quarter increases tree diameter growth (Subedi & Sharma, 2013), whereas reductions in photosynthesis occur during droughts, which decrease GPP (Schwalm et al., 2010;Van der Molen et al., 2011). Wu, Dijkstra, Koch, Peñuelas, and Hungate (2011) demonstrated that decreased precipitation suppressed aboveground biomass, whereas increased precipitation stimulated aboveground and belowground biomass. Moreover, the CO 2 concentration is the third factor related to climate change because it is expected to have a direct fertilization effect Zak, 2011) andlead to warming. Kimball (1983) had estimated that a doubling of the CO 2 concentration, all else constant, will increase growth and yield approximately 34 ± 6% in C3 plants and 14 ± 11% in C4 plants. However, elevated CO 2 does not always have a positive relationship with biomass and growth, and its fertilization effects partly depend on forest age (Körner et al., 2005) and individual tree size (Kim, Oren, & Qian, 2016).
In recent two decades, dynamic global vegetation models (DGVMs) have become important tools to investigate and predict the rate and direction of changes in global vegetation biomes in response to climate change and rising atmospheric CO 2 (Cramer et al., 2001;Notaro, 2008;Shafer, Bartlein, Gray, & Pelltier, 2015;Woodward & Lomas, 2004). Some are coupled with climate models to predict climate-vegetation interactions (Sitch et al., 2003;Notaro, Chen, & Liu, 2011), while others are run offline with different scenarios to explore the effects of changes in climate or CO 2 on vegetation (Ni, Harrison, Prentice, Kutzbach, & Sitch, 2006;Peng et al., 2009;Plattner et al., 2008;Ruosch et al., 2016;Shafer et al., 2015;Sitch et al., 2008;Woodward & Lomas, 2004;Zhang et al., 2015). For example, Cramer et al. (2001) used six DGVMs to investigate the responses of ecosystem carbon to changes in climate and CO 2 concentration. Woodward and Lomas (2004) used SDGVM (the Sheffield DGVM) to find that a scenario of future global warming resulted in a gradual decline in the terrestrial carbon sink. Galbraith et al. (2010) used three DGVMs to explore the mechanisms of Amazonian forest biomass changes under climate change; and it was found that high temperature directly increased plant respiration and declined photosynthesis and then led to reduction in forest biomass losses (Galbraith et al., 2010). Furthermore, large uncertainties may exist among different DGVMs. Sitch et al. (2008) used five DGVMs to explore that significant discrepancies were associated with the response of tropical vegetation to drought and boreal ecosystems to elevated temperatures and changing soil moisture status.
Attention has been given to the relationship between terrestrial ecosystems and climate change and atmospheric CO 2 . However, much of the research has focused on the influences of climate change and/or atmospheric CO 2 on carbon fluxes or water cycles, and few work considered how climate change and CO 2 influence the ecosystem structure (e.g., fractional coverage change) and which ecosystem types are susceptible to varying climate and CO 2 . Such issues are very important because they have a direct impact on global biogeography, carbon and water cycles, vegetation succession, and the time scale of vegetation ecosystem recovery.
In this work, two DGVMs (a revised version of IAP-DGVM1.0 and CLM4-CNDV) were used to investigate the responses of forest ecosystems to climate change with respect to changes in temperature, precipitation, and CO 2 concentration. The following questions are addressed: (1) Which regions are sensitive to climate change? (2) When the temperature, precipitation, and CO 2 concentration vary, how do the forest area and fractional coverage change? (3) Which factor has larger influences on the change in F tree (ΔF tree ; %), climate or forest ecosystem structure, and how? (4) Which climate conditions favor forest ecosystems in different regions?

| A revised IAP-DGVM1.0
IAP-DGVM1.0  was developed by the Institute of Atmospheric Physics, the Chinese Academy of Sciences, to investigate ecological processes and to study land-atmospheric interactions.
Subsequently, a revised IAP-DGVM1.0 introduced the effects of soil moisture during the growing season on the establishment rate of woody plant functional types (PFTs) in the establishment scheme (Song, Zeng, Zhu, & Shao, 2016). When coupled with CLM3, compared with the default IAP-DGVM1.0, the revised version reduced biases in forest fractional coverage in approximately 78.8% of the global grid cells, especially in arid and semiarid regions and the transition zones of forests (Song et al., 2016). In this work, the revised IAP-DGVM1.0 coupled with CLM3 is used and abbreviated as IAP-DGVM in the following sections.

| EXPERIMENTAL DESIGN
Two types of global offline simulations were conducted: one using IAP-DGVM coupled with CLM3 and the other using CLM4-CNDV. All simulations were forced circularly with 50 years of reanalysis surface atmospheric fields   from Qian, Dai, Trenberth, and Oleson (2006). IAP-DGVM ran for 800 years with T62 resolution (79 × 192 grid cells covering 60°S-90°N) to equilibrium and then restarted for another 50 years with the default atmospheric fields (control case) and climate change (i.e., with changes in temperature, precipitation, or CO 2 concentration) in several separate cases: (1) temperature ±1°C, ±2°C, and ±3°C at each time step (abbreviated as mean annual temperature [MAT] ± 1°C, MAT ± 2°C, and MAT ± 3°C, respectively); (2) precipitation increased or decreased by 15% (abbreviated as MAP115 and MAP085); and (3) doubling the CO 2 concentration (2CO 2 ). For CLM4-CNDV, the 20th-century control simulation documented by Bonan and Levis (2010) (initial conditions supplied with the CCSM4 release) was used as the initial data to run CLM4-CNDV for 600 years to equilibrium with 96 × 144 grid cells. The simulation was then restarted for the same ten cases with climate change and one control case, as IAP-DGVM. The last 50 years of simulation results were analyzed. In each simulation, only one climate factor was changed, and the others remained at the default settings. For simplicity, variables from the control cases of the two models were marked "ctrl" in the subscript.
In IAP-DGVM and CLM4-CNDV, natural plants are classified into 12 PFTs according to their physical, phylogenetic, and phenological characteristics, including seven trees (Table 1), two shrubs, and three grasses, in which trees have the highest hierarchy for the competition of establishment. Therefore, the simulation performance of tree PFTs has a direct influence on other PFT simulations, and this work mainly focused on how climate change influences forest coverage and its relevant variables using IAP-DGVM and CLM4-CNDV. The definition of fractional coverage and related parameterizations is shown in Appendice S1.

| Comparison among different sensitivity tests of temperature change
Global distribution of regions sensitive to temperature change First, to investigate which areas are sensitive to temperature change, the global distribution of differences between the maximum tree fractional coverage (F tree,max ; %) and the minimum tree fractional coverage (F tree,min ; %) from seven temperature sensitivity tests is shown in Figure 1. In IAP-DGVM, most forest regions were influenced by temperature change, and the most sensitive areas were distributed in the core areas of forests, especially in boreal forests, where the amplitude of the F tree change was approximately 10%-20%, exceeding 35% in some grid cells. In CLM4-CNDV, boreal regions also had significant sensitivity to temperature change; however, the most influenced areas were distributed in the transitional areas of boreal forests, the peripheral zones of tropical forests, and some semiarid or arid regions (e.g., western America).
The influence of temperature change on gross primary production Temperature change influences terrestrial ecosystems in various ways, and one of the most direct ways is affecting the GPP of the ecosystem

| Tree fractional coverage change and its influencing factors
Comparison of different temperature sensitivity tests showed some common points, e.g., the negative relationship between GPP eco and warming, as well as the similar response to increasing or decreasing temperature for a given tree PFT. In the following, the cases with MAT ± 1°C were used to investigate the difference in F tree (ΔF tree , %; F tree in cases of temperature change-F tree in the control simulation (F tree,ctrl ) for each model) due to temperature change in different regions and to identify the influencing factors for the two models.
The global distribution of tree fractional coverage change characteristics? To answer this question, the relationship between ΔF tree and MAT as well as F tree,ctrl was investigated ( Figure 5). Globally, warming led to negative area-averaged ΔF tree (ΔF tree ; %) in any case of F tree,ctrl for both models (the blue lines in Figure 5a,c), while the effects of decreasing temperature on ΔF tree were different between the two models (the blue lines in Figure 5b, the area-averaged F tree (F tree ; %) increased in areas with F tree,ctrl < 32% and decreased in regions with F tree,ctrl > 50% in the IAP-DGVM simulation ( Figure 5b); however, for CLM4-CNDV, F tree increased when 0 < F tree,ctrl < 72% (ΔF tree was approximately 7% when F tree,ctrl was 55%), and then, with F tree,ctrl > 72%, F tree decreased due to decreasing temperature ( Figure 5d).
In the IAP-DGVM simulations, there were two distinct tendencies in the relationship between ΔF tree and F tree,ctrl , and these tendencies depended on MAT (Figure 5a,b). When MAT increased by 1°C, F tree increased (ΔF tree > 0) in most grid cells with MAT < 0°C and decreased (ΔF tree < 0) in most grid cells with MAT > 0°C ( Figure 5a); however, Figure 5b shows that forests in warm regions benefited from cooling, in accordance with Figure 4. Figure 5a and b illustrates that the most impacted forest ecosystems were in regions with F tree,ctrl ~ (60%, 80%; the absolute value of ΔF tree (|ΔF tree |) was almost 4%). CLM4-CNDV was similar to IAP-DGVM, although the boundaries of |ΔF tree | between regions with MAT > 0 and MAT < 0 were not obvious, but |ΔF tree | was larger in CLM4-CNDV (Figure 5c-d). Overall, MAT determined whether ΔF tree was positive or negative, and the amplitude of ΔF tree was relative to F tree,ctrl .
To quantitatively explain the dependence of ΔF tree on MAT and F tree,ctrl , the correlation coefficient (R 2 ) was calculated (see Appendices i.e., for the two models: (1) In all the cases, although ΔF tree had a signif- Compared with IAP-DGVM, ΔF tree in CLM4-CNDV varied over a wide range, especially for forest ecosystems with F tree,ctrl ~ (25%, 85%; Figure 5). To determine which types of forest ecosystems have large change in ΔF tree for the two models, the relationship between F tree,ctrl and global area-averaged standard deviation of ΔF tree (σ; %) was analyzed ( Figure 6). The results showed that forest ecosystems simulated by CLM4-CNDV usually had larger σ when the temperature varied. Except for the case with F tree,ctrl at approximately 70%, σ from IAP-DGVM was almost less than 5%, while in the CLM4-CNDV cases, the maximum σ reached approximately 15% (F tree,ctrl ~ 25%) and 20% (F tree,ctrl ~ 85%) for warming and cooling, respectively ( Figure 6). These differences may be due to the larger change in ΔGPP in CLM4-CNDV.

| The effects of precipitation change on forest ecosystems
Precipitation is another key factor that influences the vegetation distribution and ecosystem structure; therefore, its effects on ΔF tree were investigated in the following. Figure 7 shows the global distribution of tree fractional coverage change due to precipitation change from IAP-DGVM and CLM4-CNDV, and following Figure 4, only grid cells with |ΔF tree | greater than 5‰ were shown. Compared with the cases of temperature change, the responses of forest ecosystems to mean annual precipitation (MAP) change were uniform, i.e., increased MAP led to globally increased F tree , while reduced MAP led to decreased F tree . However, the sensitive regions varied slightly between IAP-DGVM and CLM4-CNDV. In IAP-DGVM, large changes in F tree occurred in eastern North America, northern Asia, and most regions in South America (Figure 7a,b). However, in the CLM4-CNDV simulations, the sensitive areas mainly covered western North America, Central Asia, and the peripheral areas of the core forests (e.g., the southeast of Central Africa; Figure 7c,d).
In the responses of F tree to MAP change, CLM4-CNDV also had larger ΔF tree than IAP-DGVM (Figure 8). When increasing MAP by 15%, larger ΔF tree occurred in areas with approximately 30% < F tree < 80% in both models. However, in the case of decreasing MAP, obvious ΔF tree appeared in the grid cells with approximately 20% < F tree < 45% in IAP-DGVM, while the sensitive regions were areas with approximately 60% < F tree < 85% in CLM4-CNDV.
Similarly, to further investigate the influences of MAP and F tree,ctrl on ΔF tree , the correlation coefficients between ΔF tree and F tree,ctrl , as well as MAP, were calculated (see Appendices S4 and S5). In the same way, normalization of MAP (mm/year; i.e., MAP′ = MAP/MAP max , where MAP max was the maximum value of MAP around the globe) was performed before regression. Furthermore, ΔF tree and F tree,ctrl were used in decimal form rather than percentages (%).ΔF tree had a significant relationship with F tree,ctrl (p < .0001) and MAP′ (p < .0001), especially with F tree,ctrl , for both IAP-DGVM and CLM4-CNDV. ΔF tree in the case of increasing MAP had greater dependence on F tree,ctrl than cases with decreasing MAP (R 2 = .377 vs. .191 in IAP-DGVM; R 2 = .181 vs. .154 in CLM4-CNDV). Additionally, the ΔF tree simulated by IAP-DGVM had much stronger sensitivity to F tree,ctrl and MAP′ than in the CLM4-CNDV simulations. The relationship between the standard deviation of ΔF tree and F tree,ctrl was also considered. Similar to the cases of temperature change, σ in CLM4-CNDV was generally larger than that in IAP-DGVM for most forest ecosystems when MAP changed ( Figure 9). For IAP-DGVM, σ in the case of decreasing MAP was higher than σ in the case of increasing MAP for all groups of forest ecosystems, especially regions with F tree,ctrl ~ 70%. However, for CLM4-CNDV, grid cells with F tree,ctrl < 48% had larger σ in the case of increasing MAP (except for areas with 25% < F tree,ctrl < 36%), especially with F tree,ctrl ~ 40%, whereas in the decreasing MAP sensitivity test, higher σ occurred in regions with 70% < F tree,ctrl < 82%.

| The effects of CO 2 concentration change on forest ecosystems
Changes in the carbon dioxide level have attracted attention because increasing CO 2 concentration not only results in global warming but also increases carbon fertilization. In this work, increasing CO 2 does not lead to rising temperature, i.e., only carbon fertilization effects were considered. Figure 10 shows the relationship between F tree,ctrl and area-averaged ΔF tree in the sensitivity tests with doubled concentration (2CO 2 ) simulated by IAP-DGVM and CLM4-CNDV. It was shown that (1) the simulated F I G U R E 6 Dependence of the standard deviation of ΔF tree (σ; %) on tree fractional coverage (F tree,ctrl ; %) simulated by IAP-DGVM and CLM4-CNDV in the case of MAT ± 1°C F tree in the two models had a positive response to CO 2 concentration; (2) when doubling the CO 2 concentration, ecosystems with 35% < F tree,ctrl < 40% had the strongest sensitivity to CO 2 change, ΔF tree reached approximately 12% and 14% for IAP-DGVM and CLM4-CNDV, respectively.

| CONCLUSIONS AND DISCUSSION
Forests are particularly vulnerable to changing environmental conditions due to the longevity of tree species (Kräuchi, 1993). However, climate change effects on forests may also be subtle, affecting individual tree growth and forest composition and structure from years to decades (Pederson et al., 2015).
In this study, the responses of forest ecosystems to changes in climate and CO 2 concentration were investigated by IAP-DGVM coupled with CLM3 and CLM4-CNDV. In the temperature change sensitivity tests, it was shown that (1)  Compared with the temperature change, the responses of forests to precipitation and CO 2 concentration changes were more uniform, i.e., F tree increased with precipitation and CO 2 concentration around the globe. The regions sensitive to increasing and decreasing MAP were different. Areas with 30% < F tree < 60% (in IAP-DGVM) or semiarid and arid regions (in CLM4-CNDV) had strong sensitivity to increasing MAP; however, as MAP decreased, F tree in areas with large F tree decreased remarkably in IAP-DGVM, while F tree in semiarid and arid regions in CLM4-CNDV dropped significantly. Similar to the temperature change simulations, ΔF tree was more dependent on F tree,ctrl than MAP, and the standard deviations of ΔF tree in CLM4-CNDV were higher than those from IAP-DGVM. For the CO 2 concentration simulations, both DGVMs captured the CO 2 fertilization effects.
As shown in Figure 3, tropical PFTs had opposite responses to temperature change between two models. Our other research showed that such distinctions were likely to result from the differences in seedling establishment scheme and photosynthesis parameterization (see Appendice S1). IAP-DGVM explicitly considers the impact of soil moisture on the establishment rates of woody PFTs. When temperature decreased, lower evapotranspiration increased soil moisture, not only benefiting seedling establishment rates which increased tree population densities, but also improving the maximum rate of carboxylation (V max ) and GPP eco (Figure 2 In IAP-DGVM, the widest range of ΔF tree appeared in the grid cells with 60% < F tree,ctrl < 80%. However, in CLM4-CNDV, ΔF tree varied over a large range, as shown by the smaller number of grid cells with 25% < F tree,ctrl < 85% ( Figure 5). In addition to the differences in GPP variance, due to the significant dependence of ΔF tree on F tree,ctrl , differences in the simulated F tree,ctrl accounted for the discrepancies in ΔF tree between the two models. The results showed that excluding grid cells with F tree,ctrl < 5%, approximately 19.0% and 31.2% of the grid cells fell in the intervals F tree,ctrl < 20% and F tree,ctrl > 85% in IAP-DGVM, whereas in CLM4-CNDV, the percentages reached approximately 16.0% and 64.5%, respectively ( Figure S2; to concentrate on F I G U R E 9 Dependence of the standard deviation of ΔF tree (σ; %) on tree fractional coverage (F tree,ctrl ; %) simulated by IAP-DGVM and CLM4-CNDV in the cases of MAP increasing or decreasing by 15% F I G U R E 1 0 The relationship between tree fractional coverage (F tree,ctrl ; %) and the change in tree fractional coverage (ΔF tree ; %) in the cases of doubled concentration (2CO 2 ) simulated by IAP-DGVM and CLM4-CNDV the core areas of forests, only grid cells with F tree,ctrl > 5% were considered, and in the two models, F tree,ctrl is assumed not to exceed 95% in each grid cell; therefore, there were no results when F tree,ctrl < 5% or F tree,ctrl > 95%). The combination of the differences in the simulated F tree,ctrl and the dependence of ΔF tree on F tree,ctrl largely accounted for the differences in ΔF tree and its standard deviation in these two models.
As discussed in previous research, the responses of forest ecosystems are spatially heterogeneous and partly uncertain (Mekonnen et al., 2016;Williams et al., 2010;Willis et al., 2012). To further investigate the differences in the response of forest ecosystems to temperature change in different regions, the optimal temperature change (relative to the current temperature) was defined as temperature condition in the seven temperature sensitivity tests under which F tree was the largest. Only grid cells with F tree,ctrl greater than 1% were considered. Large discrepancies existed in the global distribution of the optimal temperature between IAP-DGVM and CLM4-CNDV MAT − 1°C or MAT in many boreal grid cells. In the arid and semiarid regions (e.g., the western USA) and transitional zones of forests (e.g., the peripheral areas of the tropical rainforests in Central Africa), decreased temperature was good for forest coverage, and F tree was largest at MAT − 3°C because cooling relieved drought or reduced respiration, decreasing tree mortality in these regions, which was somewhat in accordance with Williams et al. (2010). The tropical forests in CLM4-CNDV mostly had the largest F tree in the case of MAT + 3°C; however, due to F tree,ctrl being close to the upper 95% limit provided in the models, the increment of F tree was small in these areas.
This work provided valuable ideas to investigate the responses of forest ecosystems to climate change and several vital clues to explore the uncertainties in the current vegetation dynamic models. In the following work, the combined effects of changes in temperature and precipitation on vegetation will be considered.