Spatiotemporal patterns of evapotranspiration in response to multiple environmental factors simulated by the Community Land Model

Spatiotemporal patterns of evapotranspiration (ET) over the period from 1982 to 2008 are investigated and attributed to multiple environmental factors using the Community Land Model version 4 (CLM4). Our results show that CLM4 captures the spatial distribution and interannual variability of ET well when compared to observation-based estimates. We find that climate dominates the predicted variability in ET. Elevated atmospheric CO2 concentration also plays an important role in modulating the trend of predicted ET over most land areas, and replaces climate to function as the dominant factor controlling ET changes over the North America, South America and Asia regions. Compared to the effect of climate and CO2 concentration, the roles of other factors such as nitrogen deposition, land use change and aerosol deposition are less pronounced and regionally dependent. The aerosol deposition contribution is the third most important factor for trends of ET over Europe, while it has the smallest impact over other regions. As ET is a dominant component of the terrestrial water cycle, our results suggest that environmental factors like elevated CO2, nitrogen and aerosol depositions, and land use change, in addition to climate, could have significant impact on future projections of water resources and water cycle dynamics at global and regional scales.


Introduction
Evapotranspiration (ET), or water transferred from the land surface to the atmosphere, is an essential process in the climate system that links water, energy and carbon cycles (Nachabe et al 2005, Alton et al 2009, Jung et al 2010, Wang and Dickinson 2012. ET over land is the second largest component (after precipitation) of the terrestrial water cycle at the global scale and returns about 60% of precipitation falling on land back to the atmosphere on an annual basis (L'vovich and White 1990, Oki and Kanae 2006, Lettenmaier and Famiglietti 2006. It is expected that the hydrological cycle will be intensified under climate change (Huntington 2006) with varied impacts on ET. Physically and physiologically, ET is driven not only by climatic factors, such as precipitation, temperature, wind speed, surface humidity, and solar radiation, but also is modulated by changes in environmental factors such as the atmospheric CO 2 concentration, nitrogen and aerosol deposition, and land use and land cover change (Gedney et al 2006, Piao et al 2007, Felzer et al 2009, Shi et al 2011.
The atmospheric CO 2 concentration affects the hydrological cycle mainly in two ways: (1) the physiological effect where plants regulate the opening and closing of their stomata in response to changes in CO 2 concentration and (2) the structural effect where the increase in CO 2 concentration leads to enhanced vegetation growth, thus changing plant structure and increasing the leaf area index (Piao et al 2007, Felzer et al 2009, Cao et al 2010. Gedney et al (2006) found that increased atmospheric CO 2 concentrations during the period 1960-1994 has reduced ET with compensating increases in runoff, and attributed these changes to CO 2 physiological forcing. A recent study (Gopalakrishnan et al 2011) applied the Community Land Model version 3.5 (CLM3.5) in an offline mode to explore changes in canopy transpiration scale with CO 2 concentrations and whether the change in canopy transpiration tends to saturate at higher levels of CO 2 concentrations. Their result has shown that canopy transpiration declines at about 5.1%/100 ppmv increase in CO 2 levels. Using CLM3.5 coupled with the Community Atmosphere Model 3.5, Cao et al (2010) have investigated not only the CO 2 -physiological effect, but also the CO 2 -radiative effect on runoff. They reported that decreased ET from the continents caused by reduced stomatal conductance would increase runoff by 8.4 ± 0.6%, or by 5.2 ± 0.6% as a result of CO 2 -radiative forcing under a doubled atmospheric CO 2 concentration scenario. However, in our study, we considered the overall effect of both the physiological effect and the structural effect of atmospheric CO 2 concentration on ET. Some modeling studies have also suggested that both physiological and structural forcing as result of rising CO 2 concentration would alter ET (Cramer et al 2001, Betts et al 2007, Felzer et al 2009, Alkama et al 2010. Other environmental factors could also affect ET in different ways. For example, land cover changes could alter the depth of the soil from which plants can extract water, thus modifying energy available to drive ET by changing the land-surface albedo (Gedney et al 2006). Changes in atmospheric nitrogen deposition can alter regional patterns of nitrogen limitation (Shi et al 2011), which would in turn modify the water balance of ecosystems and the associated ET (Felzer et al 2009). One observed study showed higher levels of atmospheric aerosol concentration reduced the amount of solar radiation reaching the land surface leading to less open-pan evaporation over the past half century (Roderick and Farquhar 2002). Deposition of aerosol particles to the snow surface could also become an important factor in modulating the water cycle and climate through the snow albedo feedback (Flanner et al 2006).
During the past decades, considerable efforts have been made to improve our understanding of the factors that control variability of ET at different time scales. However, the mechanisms and factors that govern long-term trends of ET are still poorly quantified (Wang and Dickinson 2012). Hence, the purpose of this study is to quantify the possible contributions of different factors, including climate change, atmospheric CO 2 , nitrogen deposition, land use and land cover change, and aerosol deposition, to trends and variability in ET over the period of 1982-2008.

Model description
Version 4 of the CLM, used in this study, succeeds CLM 3.5 (Oleson et al 2008) with revised runoff generation and snow parameterizations, organic soil, a 50 m deep ground column for energy budget calculations, a 3.82 m deep soil column overlaid with a groundwater aquifer for hydrologic calculations, and an updated distribution of plant functional types (Oleson et al 2010, Lawrence et al 2011. The fully prognostic carbon and nitrogen dynamics of the terrestrial biogeochemistry model Biome-BGC (version 4.1.2) Rosenbloom 2005, Thornton et al 2002) have also been merged to CLM4. The resulting model, CLM4, includes prognostic carbon and nitrogen pools and fluxes in vegetation, litter, and soil organic matter Zimmermann 2007, Thornton et al 2009). Thus, CLM4 can modulate transpiration and canopy evaporation through prognostic leaf and stem area, which influences ET. The hydrology scheme includes a revised ground evaporation parameterization that accounts for the stability of near-surface air as modified by the vegetation canopy and a litter layer. The snow parameterization incorporates the snow and ice aerosol radiation model, which includes aerosol deposition, grain-size dependent snow aging, and vertically resolved snowpack heating (Lawrence et al 2011).

Simulation setup
Global simulations were conducted using CLM4 driven by the 111-year (1901-2010) observation-constrained half-degree CRUNCEP dataset (http://dods.extra.cea.fr/data/p529viov/ cruncep/readme.htm), including temperature, precipitation, specific humidity, solar radiation, wind speed, pressure and long wave radiation at a 6 h time step. The CRUNCEP is a combination of the CRU TS.3.2 0.5 • monthly climatology covering the period 1901-2010 (http://badc.nerc.ac.uk/view/ badc.nerc.ac.uk ATOM dataent 1256223773328276) and the 2.5 • NCEP2 reanalysis data beginning in 1948 and available in near real time (Kanamitsu et al 2002, Mao et al 2012a, 2012b. Consistent with the climate forcing, the spatial resolution of the CLM4 simulations is 0.5 • , with a temporal resolution of 30 min. The simulations were spun up to equilibrium under environmental conditions (i.e., atmospheric CO 2 , nitrogen deposition, land cover, and aerosol deposition) in the year 1850, and based on a repeated 20-year subset of the transient climate dataset (1901)(1902)(1903)(1904)(1905)(1906)(1907)(1908)(1909)(1910)(1911)(1912)(1913)(1914)(1915)(1916)(1917)(1918)(1919)(1920). The simulations for 1850-1900 used the same 20-year climate forcing, but incorporated historical transient datasets for CO 2 concentration, nitrogen deposition, land use and land cover change, and aerosol deposition. Transient climate forcing was introduced at the start of the CRUNCEP record, and was used together with the other transient environmental driving dataset for the period 1901-2009. Annual land use change and harvest area were derived from the University of New Hampshire version 1 Land-Use History A (LUHa.v1) historical dataset based on that of Hurtt et al (2006) for 1850-2005, and from the RCP4.5 scenario of AR5 for the period 2006-2009, respectively. Effects of rotational wood harvest, conversion of natural vegetation to agriculture or pasture, and abandonment of managed lands are included in the land use change term (Shi et al 2011). The details of the atmospheric CO 2 concentration and nitrogen deposition are similar to Shi et al (2011). The aerosol deposition data include eight particle species: hydrophilic black carbon, hydrophobic black carbon, hydrophilic organic carbon, hydrophobic organic carbon, and four species of mineral dust. In our simulations, the aerosol deposition rates are prescribed according to rates obtained from a transient 1850-2009 CAM-chem simulation with interactive chemistry (Lawrence et al 2011).
In order to assess the relative contributions of climate, increasing atmospheric CO 2 concentration, nitrogen deposition, land use and land cover change, and aerosol deposition, we performed seven simulations in this study. In simulation S1 (i.e., the control simulation), we repeated the 20-year subset of climate drivers (1901)(1902)(1903)(1904)(1905)(1906)(1907)(1908)(1909)(1910)(1911)(1912)(1913)(1914)(1915)(1916)(1917)(1918)(1919)(1920) for the entire period 1850-2009, and kept atmospheric CO 2 concentration, nitrogen deposition, land use change, and aerosol deposition constant at their 1850 values. In simulations S2-S5, we used the same subset of transient climate, and varied one of the four remaining factors while holding the other three constant at their 1850 values (i.e., CO 2 , nitrogen deposition, land use change and aerosol deposition are varied in S2, S3, S4 and S5 simulations, respectively). In the S6 simulation (hereafter CLIM simulation), historical transient climate from CRUNCEP was applied after 1900 and the other factors were held at their 1850 values. Finally, in simulation S7 (hereafter ALL simulation), we allowed all factors (climate, CO 2 , nitrogen deposition, land use change and aerosol deposition) to vary throughout the fully transient simulation. The effect of each individual non-climate factor is calculated by subtracting S2, S3, S4 and S5 from simulation S1 (hereafter referred to CO 2 , NDEP, LUC and AERO, respectively).
To evaluate our simulations, we used the global land ET data derived from the FLUXNET network of eddy covariance towers using the model tree ensembles (MTE) approach (Jung et al 2010). The FLUXNET-MTE up-scaling provides monthly ET at 0.5 • spatial resolution over the period 1982-2008, allowing comparison with CLM4 results without need for spatial interpolation or regridding. As with observation-based ET data, only the model simulations from 1982 to 2008 were selected and analyzed.

Spatial patterns of ET
To provide some credibility for the model predicted ET changes, we first compare the CLM4 simulated globally averaged and spatial patterns of ET with observation-based FLUXNET-MTE product. Global mean ET as predicted by CLM4 is 639 mm yr −1 for the period 1982-2008, compared to FLUXNET-MTE estimation of 574 mm yr −1 . Figure 1 shows the spatial distributions of land annual mean ET over the study period for ALL simulation, FLUXNET-MTE and the difference between them. It can be seen that CLM4 captures the global distribution of ET well (figures 1(a) and (b)), but CLM4 predicted ET is higher than FLUXNET-MTE over the tropics (figure 1(c)), which is the major contribution to the higher value of CLM4 simulated globally averaged ET when compared to MET product.

Interannual variation in ET
Both the CLM4 modeled ET in simulation ALL (including all forcing factors) and the observation-based FLUXNET-MTE product show significant interannual variability between 1982 and 2008 (figure 2(a)). The interannual variability of ET due to the combined effects of all the forcing factors is consistent with the observation-based ET data (R = 0.65, P < 0.005). On average, the simulated global land area mean ET from simulation ALL demonstrates a significant positive trend with the rate of 0.60 ± 0.14 mm yr −2 , slightly higher than the trend of 0.47 ± 0.12 mm yr −2 for the observation-based ET over the study period 1982-2008, but both at high confidence levels with p-values less than 5%. Jung et al (2010) also have reported that ET shows a declining trend for the subset time period of 1998-2008. However, our predicted ET trend does not follow that decreasing trend as the MTE product does. We have separated the effect of each individual factor for climate, increasing atmospheric CO 2 concentration, nitrogen deposition, land use and land cover change, and aerosol deposition. The results show that ET predictions from the individual factor simulations also demonstrate substantial interannual fluctuation over the study period (figure 2(b)). We will attribute the effect of each individual factor to the ET trends in section 3.3.
Simulation ALL captures regional-scale interannual variability in ET over major continents when compared to the FLUXNET-MTE product, with the exception of South America (figure 3). CLM4 can reproduce more than 50% variance over three continents (North America, Africa and Australia), and 25% variance over the other continents (Europe and Asia). While, CLM4 fails to captures the variance of South America.

Factor contributing to the global and continental ET trends
For the period 1982-2008, globally averaged increasing trends of CRUNCEP climate forcings over land in precipitation, temperature, specific humidity, shortwave radiation, and long wave radiation are significant. Wind speed also shows an increasing rate but the trend is not significant. In contrast, surface pressure shows an insignificant decreasing trend over the study period. Other forcing factors, such as atmospheric CO 2 and aerosol deposition show significant increasing trends, while nitrogen deposition shows insignificant increasing trend   (table 1). It should be noted that all these statistics are global land values over the period of 1982-2008, and that regional and different temporal features may vary.
We investigated the relative contributions of major environmental driving factors to global and continental-level ET. For global land, historical climate variation generates a significant increasing ET trend, with a rate 0.78 mm yr −2 , while the CO 2 -only simulation produces a decreasing trend (−0.20 mm yr −2 ). Nitrogen deposition generates an increase in ET of 0.02 mm yr −2 , and land use change results in a decrease of 0.001 mm yr −2 . Aerosol deposition does not exert any significant effects for global land ET. Figure 4 shows that climate is the strongest driving factor, and rising CO 2 concentration is the second most important for trends in ET over global land, or regionally over Europe, Africa and Australia, while nitrogen deposition, land use change There is a clear disparity in global-scale ET trends between our climate-only simulation (CLIM) and the observation-based MTE product (figure 4), which suggests that ET change driven by climate alone is not sufficient to account for the total trend in MTE ET. Because the MTE product is based on flux observations, it implicitly includes the influence of rising CO 2 , anthropogenic nitrogen deposition, and other non-climate forcing factors. Our all factor simulation (ALL) has a global-scale trend closer to the MTE product, with rising CO 2 indicated as the most important factor modifying the influence of climate. We see that this pattern also holds for several sub-regions (South America, Africa, and Australia), but that in several other regions our ALL simulation is worse than CLIM, when compared to MTE (North America, Europe, and Asia, figure 4).

Spatial patterns of trends in ET
The general agreement of the spatial distribution between the modeled ET from the simulation driven by all factors (ALL) and observation-based estimates by Jung et al (2010) is evident as seen in figure 1, suggesting that it is useful to explore the spatial patterns of changes in ET. Figure 5 displays the spatial distribution of trends in modeled and observation-based ET, and CRUNCEP precipitation over the 27-year study period. The ET trend from all factor simulation ALL (figure 5(a)) follows the spatial distribution of precipitation in general (figure 5(h)), and has similar spatial patterns compared to observation-based MTE data ( figure 5(b)). However there are some notable differences between the modeled and MTE ET over Europe, China, southeastern North America and southeastern Africa. As discussed earlier, the MTE ET product relies heavily on FLUXNET observations and is subject to uncertainty rooted from the limited spatial and temporal coverage of flux towers, which made it an imperfect metric for measuring model performance. In addition, MTE ET dataset does not explicitly take into account the other environmental factors, such as CO 2 concentration, nitrogen deposition, land use and land cover change, and aerosol deposition.
The spatial pattern of ET trends in climate-only simulation CLIM follows the pattern from all factors simulation ALL very well, suggesting that the increasing trend of ET over the past 27 years is induced mainly by the changes in mean climate, as well as by its variability ( figure 5(c)). Nitrogen deposition enhances the upward trends in ET over most parts of the global land (figure 5(e)), which is consistent with a previous study examining the influence of single forcing factors on river flow (Shi et al 2011). In contrast, elevated atmospheric CO 2 concentration exerts significant decreasing trends in ET over almost global land surface ( figure 5(d)), a result in agreement with previous studies (Cramer et al 2001, Betts et al 2007, Felzer et al 2009, Alkama et al 2010, Shi et al 2011. The land use change generates inhomogeneous trends over most portions of global land, but induces upward trends over high latitude regions of the Northern Hemisphere, southeastern China and western parts of Australia (figure 5(f)). Piao et al (2007) also have reported that land use change plays an important role in controlling regional runoff values. Aerosol deposition enhances ET over low latitude areas and Southern Hemisphere, while decreases ET over middle and high latitude regions of Northern Hemisphere ( figure 5(g)).

Sensitivity of ET to climate change
To clarify the relationship of ET to climate variability, correlation coefficients between ET and climatic variables are  (table 2). Modeled ET has a significant positive correlation with precipitation in all regions and globally, while observation-based MTE ET shows significant positive relationship only for Australia, Africa and globally. It is noticeable that the correlation between ET and precipitation is higher over dry regions (e.g., Australia), in which the ET processes are controlled by water availability in the root zone or shallow surface. CLM4 ET has significant positive connection with temperature globally and over Europe and Africa, while MTE ET has significant positive correlation with temperature globally and for all regions except Australia. Both CLM4 and MTE ET have significant negative correlation with temperature over Australia, where high temperature is generally associated with drought conditions. Modeled ET shows significant positive correlation with specific humidity globally and over Europe, Asia and Africa, while MTE ET shows significant positive correlation globally and over all continents except Australia. Both CLM4 and MTE ET show significant positive correlation with long wave radiation for global and continental scale except over South America. There are few significant correlations between either CLM4 or MTE ET and shortwave radiation or wind speed, although both CLM4 and MTE approaches have a significant negative correlation with shortwave radiation over Australia, which might be related to drought effect. We also calculated the correlation coefficients of CLM4 predicted net radiation and ET, but not for MTE dataset because of lacking the observed net radiation data. The result shows that there are significant positive relationship between CLM4 predicted net radiation and ET for global land, Africa, Austria, North America (the correlation coefficient is 0.50, 0.78, 0.73 and 0.78, respectively), while there are not significant relationships over Asia, Europe and South America (the correlation coefficient is 0.33, 0.44, and 0.12, respectively). The only region where either approach shows a significant correlation between ET and surface pressure is Table 2. Correlations between annual mean climatic variables and ET (Prec: precipitation, Temp: temperature, SH: specific humidity, LWR: long wave radiation, SWR: short wave radiation, Wind: wind speed, Psrf: surface pressure). These seven variables are from CLM4 climate forcing data, CRUNCEP, while the net radiation (NetR) is from CLM4 output. Bold values represent trends with significance (P < 0.05). Europe, with CLM4 predicting a negative relationship, while MTE shows a positive relationship. The spatial distribution of correlation coefficients between modeled ET from all factor simulation ALL and observation-based ET and selected climatic variables also demonstrates that the modeled ET has significant positive correlation with precipitation (figure 6(a)), while the observation-based ET tends to highly correlate with temperature and specific humidity (figures 6(d) and (f)). This could be related to the fact that the MTE ET dataset is biased for energy-limited regions (i.e., humid regions), such as limited coverage of FLUXNET sites over the tropics. Both modeled and observation-based ET show significant positive relationship with long wave radiation over most global land areas (figures 6(g) and (h)).

The uncertainty of ET estimation
Comparison to other global-and continental-scale estimates of ET is helpful in assessing the uncertainty in CLM4 predictions. CLM4 all factor simulation (ALL) predicted globally averaged ET is 639 mm yr −1 for the period 1982-2008, compared to FLUXNET-MTE estimation of 574 mm yr −1 . Both numbers are comparable to Zeng et al (2012), who reported the globally averaged land ET at 604 mm yr −1 with a range of 558-650 mm yr −1 . Mueller et al (2011) compared 30 global observation-based ET datasets and the modeled ET from 11 coupled global climate models of the IPCC Fourth Assessment Report and concluded that the global mean annual ET of the 41 datasets ranged from 511 to 650 mm yr −1 , with an average value of 580 mm yr −1 . Table 3 shows the global-and continental-scale ET trends from our all factor simulation ALL, MTE estimates, and two other data sources: one from Zhang et al (2010) (hereafter ZHA10), and the other from Zeng et al (2012) (hereafter ZENG12). Both MTE and ZHA10 estimated ET by up-scaling local eddy covariance flux measurements from global FLUXNET network, through integration with gridded satellite FPAR (or NDVI) and climate data. MTE used Global Precipitation Climatology Center (GPCC) precipitation and temperature from CRU while ZHA10 used NCEP precipitation and temperature. ZENG12 was estimated by using a simple regression approach using temperature and precipitation from CRU and NDVI data from NOAA/AVHRR. Both our simulations and the three studies agree that there is a significant increasing trend in ET globally and over Africa. Over Europe, all the other three studies show significant upward trends in ET, while CLM4 shows an insignificant increasing trend. CLM4 predicts insignificant decreasing trends in ET over North America and South America, while the other three methods estimate significant increasing trends. Over Asia, CLM4 demonstrates an insignificant downward trend, while ZHA10 shows a Figure 6. The spatial distribution of correlation coefficients between modeled annual ET (left panel) from all factor simulation ALL and observation-based ET (right panel) and climatic variables (precipitation, temperature, specific humidity and long wave radiation from top to bottom of each panel, respectively). significant decreasing trend, and MTE and ZENG12 show significant upward trends in ET. Both CLM4 and the MTE and ZENG12 estimate insignificant increasing trends in ET over Australia, while ZHA10 estimates an insignificant decreasing trend.
It should be noted that our predicted ET from the all factor simulation ALL considers not only climate change, but also the rising atmospheric CO 2 concentration, nitrogen deposition, land use and land cover change and aerosol deposition. However, ET of MTE, ZHA10 and ZENG12 do not explicitly take into account these other environmental factors. It has been shown in recent years that the effect of rising atmospheric CO 2 concentration on the water cycle is very important (Gedney et al 2006, Piao et al 2007, Shi et al 2011. Figure 4 clearly shows that the ET trend should change from negative to positive if we exclude the effect of CO 2 from our all factor simulation ALL over North America, Asia and South America, confirming that the rising atmospheric CO 2 concentration is very crucial for land ET. Piao et al (2007) and Shi et al (2011) have reported that it is necessary to consider the effect of land use change on the hydrological cycle, and Felzer et al (2009) and Shi et al (2011) have demonstrated that nitrogen limitation is another factor controlling the water balance of ecosystems. Zeng et al (2012) have noted that excluding the effects of rising atmospheric CO 2 concentration, land use change and humidity on the estimation of ET, introduced uncertainty for their dataset.
The variation among these independent methods for estimation of global and regional trends in ET highlights the need for continued and expanded collection of ET observations for validation. The advantage of the modeling approach capable of quantifying ET over large areas and field experiments capable of providing accurate local-scale ET estimates with process-level understanding could serve as constraints for improving models and reducing uncertainty in ET estimate, respectively, making collaborative efforts between field scientists and modelers essential, especially in data-poor regions.

The uncertainty of our prediction
This study investigates the relative contributions of multiple environmental factors on temporal and spatial variability in ET during 1982ET during -2008. Although this relatively comprehensive analysis is intended to quantify the factorial contributions to the changing rate of ET, it is also important to recognize the uncertainties that are inherent in such a study. First, our simulations do not consider some possible disturbances (except for fire) or environmental factors that may influence the terrestrial ecosystem water cycle, for example ozone pollution (Felzer et al 2009), or irrigation (Doll 2002, Gordon et al 2005, all of which might influence regional and global trends of ET. Further studies are still needed to quantify the effects of these factors on ET. Second, uncertainties induced by model structure, parameters, and the driving data remain to be evaluated.

Conclusions
In this study, the process-based model CLM4 is used to explore the relative importance of changes in climate, atmospheric CO 2 concentration, nitrogen deposition, land use change and aerosol deposition on spatial and temporal variations in ET at global and continental scales during the period 1982-2008. Our model results suggest that the changing rates in ET at global and continental scales have been mainly a consequence of climate and rising CO 2 concentration during the study period. The relative roles of nitrogen deposition, land use change and aerosol deposition are small and variable by region. Our simulated results not only provide insights for large-scale field experiments, but also highlight the importance of biosphere feedbacks and anthropogenic influence on hydrological cycle (Gedney et al 2006, Piao et al 2007. The roles of non-climate factors, such as the rising atmospheric CO 2 concentration, nitrogen deposition (Shi et al 2011), land use change (Piao et al 2007, Shi et al 2011 and aerosol deposition should not be ignored when project future changes in water cycle and climate.