Simulated climatology and evolution of aridity in the 21st century

Transcription

1 PUBLICATIONS Journal of Geophysical Research: Atmospheres RESEARCH ARTICLE Key Points: Future changes in aridity are analyzed using CESM-LE and CMIP5 Internal variability is smaller than that of model structural uncertainty On the interannual time scale, AI sensitivity is dominated by precipitation variability Correspondence to: L. Lin, Citation: Lin, L., A. Gettelman, S. Feng, and Q. Fu (2015), Simulated climatology and evolution of aridity in the 21st century, J. Geophys. Res. Atmos., 120, , doi:. Received 2 DEC 2014 Accepted 24 MAY 2015 Accepted article online 29 MAY 2015 Published online 25 JUN 2015 Simulated climatology and evolution of aridity in the 21st century Lei Lin 1, Andrew Gettelman 2, Song Feng 3, and Qiang Fu 1,4 1 College of Atmospheric Sciences, Lanzhou University, Lanzhou, China, 2 National Center for Atmospheric Research, Boulder, Colorado, USA, 3 Department of Geosciences, University of Arkansas, Fayetteville, Arkansas, USA, 4 Department of Atmospheric Sciences, University of Washington, Seattle, Washington, USA Abstract Future changes in aridity, defined as the ratio of annual precipitation to potential evapotranspiration (PET), are analyzed using simulations from the Community Earth System Model (CESM) Large Ensemble (LE) and the phase 5 of the Coupled Model Intercomparison Project (CMIP5) during the period Both CESM and CMIP5 ensembles can reproduce the observed temporal and spatial variability of aridity. On the interannual time scale, annual average PET is sensitive to the variability of relative humidity, net surface energy flux, and surface air temperature, while the precipitation variability is the dominant component of annual average aridity sensitivity. For the long-term trends, differences between the two ensembles illustrate that the impact of the internal variability is smaller than that of the model structural uncertainty with the trends from the CMIP5 ensemble of models having a much larger spread than those from the single-model CESM-LE. The annual mean aridity averaged over global land increases (becomes drier) by 6.4% in relative to Aridity trends differ by region in the ensemble mean. In the future, increasing precipitation leads to decreasing aridity over northwest China and central (or tropical) Africa, while decreasing precipitation leads to drying (increasing aridity) in the subtropics, northern and southern Africa, and the Amazon. Increases in PET can lead to increasing aridity even in regions with increasing precipitation American Geophysical Union. All Rights Reserved. 1. Introduction Aridity is a quantitative indicator of the degree of water deficiency at a given location [Hulme, 1995; Middleton and Thomas, 1992; Mortimore, 2009], which can have far-reaching impacts on agriculture, ecosystems, and water availability [Confalonieri et al., 2007]. A more arid climate is usually accompanied by an increase in the frequency and severity of droughts [Heathcote, 1983], in a similar way to how a warmer climate may cause more heat waves. Droughts cause tens of billions of dollars in damage and affect millions of people in the world each year [Heathcote, 1983; Wilhite, 2000]. Understanding how the aridity may change in the future is critical for better future water management in a warmer climate. Aridity index (AI) is defined by the ratio of annual precipitation (P) to annual potential evapotranspiration (PET), or AI = P/PET. The PET, a basic land climate variable, is defined as the amount of water transpired in unit time by a short green crop, completely covering the ground, of uniform height and never short of water [Penman, 1956; Hartmann, 1994; Allen et al., 1998]. Various methods were developed to estimate the PET, but the Penman-Monteith (PM) equation is regarded as the most reliable PET algorithm under all climatic conditions [Jensen et al., 1990; Donohue et al., 2010; Dai, 2011; Sheffield et al., 2012]. The PM is based on the surface energy balance and accounts for the impact of available energy, wind speed, atmospheric humidity, and air temperature on PET. Future drying over land is projected in many recent studies. For example, using output from 27 fully coupled climate models participating in the phase 5 Coupled Model Intercomparison Project (CMIP5), Feng and Fu [2013] found increasing aridity in most tropical and midlatitude land regions in the 21st century. Sherwood and Fu [2014] pointed out that, even though many regions will get more precipitation in the future, precipitation cannot keep pace with the increasing evaporative demand, collectively leading to a more arid future. Fu and Feng [2014] examined how the terrestrial mean aridity (P/PET) responds to global warming due to the increase of CO 2 concentrations using the CMIP5 transient CO 2 increase to 2 CO 2 simulations. They showed a decrease in P/PET (i.e., a drier terrestrial climate) by ~3.4%/ C ocean mean surface temperature increase. It is shown that a drier terrestrial climate is largely caused by enhanced land warming relative to the LIN ET AL. LIN: 21ST CENTURY ARIDITY 5795

2 ocean and a decrease in relative humidity over land but an increase over ocean [Sherwood and Fu, 2014; Fu and Feng, 2014]. Furthermore, part of the increase in net surface radiation goes into the deep ocean, and there is a different sensitivity of PET over land and evaporation over ocean to given changes in atmospheric conditions [Fu and Feng, 2014]. Cook et al. [2014]alsoshowedrobustcross-modeldryinginthewesternNorthAmerica, Central America, the Mediterranean, southern Africa, and the Amazon by the end of this century. However, previous studies have not examined and compared the climatology, spatial distribution, and seasonal and interannual model-simulated aridity with observations. For example, Feng and Fu [2013] adjusted the model-simulated aridity values to have the same climatology of as observations to focus on the long-term trends. Furthermore, these studies are based on CMIP5 data where individual CMIP5 ensemble members can have differing physics, dynamical cores, resolution, and initial conditions. In sum, the model uncertainty and internal climate variability, as defined by Hawkins and Sutton [2009], are difficult to disentangle. This study uses the Community Earth System Model (CESM) Large Ensemble (LE) to evaluate internal climate variability and contrast it with model structural uncertainty from the CMIP5 archive. The CESM-LE samples only internal variability, while the CMIP5 archive samples both internal variability and model uncertainty. We will first examine the climatology of the CESM simulated aridity and compare it with estimates from reanalysis, and then evaluates the changes in aridity during Simulations from 19 global climate models participating in the phase 5 of the Coupled Model Intercomparison Project (CMIP5) for the same period will be used to compare with CESM-LE results. This paper explores the climatology, spatial distribution, and seasonal and interannual variability of global climate model (GCM)-simulated aridity, which has not been done before. Furthermore, we will examine the impact of the internal variability on the trends by analyzing the CESM-LE simulations, and by contrasting results from CESM-LE with CMIP5 archive, we can obtain the insight on the impact of the CMIP5 model structural uncertainty. Section 2 describes the reanalysis data set and model simulations as well as the way in which we process the data to get PET. Section 3 compares AI from the model with reanalysis data from the period 1980 to 2005 and analyzes the sensitivity of PET and AI to the different components in the model and reanalysis data. In section 3, we then use the simulation to understand how AI might change in the future, and what drives of that change. Summary and conclusions are in section Data and Methods 2.1. Model Data The Community Earth System Model (CESM) is an Earth system model consisting of atmosphere, land, ocean, and sea ice components that are linked through a coupler for exchanging state information and fluxes among these components. [Hurrell et al., 2013]. The CESM Large Ensemble (CESM-LE) Project [Kay et al., 2014] is a community resource for studying climate change in the presence of natural climate variability. CESM-LE has performed a 30 ensemble members of perturbation simulations from 1920 to 2080 using a coupled ocean-atmosphere-land and sea ice version of CESM-CAM5. All 30 CESM-LE members use the same model and the same external forcing. Ensemble simulations have begun with random perturbations of the atmospheric air temperature in 1920 as described by Kay et al. [2014]. The horizontal resolution is (latitude and longitude) for the atmosphere and land, and 1 for the ocean. CESM-LE uses an improved land surface model and atmosphere and terrestrial biosphere processes [Hurrell et al., 2013]. Simulations use historical forcing through 2005 for greenhouse gases, ozone, solar fluxes, and aerosol emissions and then the RCP8.5 forcing scenario with ozone recovery from 2005 to See Kay et al. [2014] for further details. We use output from 19 global climate models in the CMIP5 ensemble (Table 1) with specified historical anthropogenic and natural external forcings for and with 21st century changes in greenhouse gases and anthropogenic aerosols following the RCP8.5 for [Taylor et al., 2012]. The first ensemble run was used if a model has multiple ensemble simulations Reanalysis Data Many users regard reanalysis products as equivalent to observations [Zhang and Cook, 2014; Davini et al., 2014; You et al., 2013; Zhong et al., 2013]. This is really not the case. Reanalyses are a model product constrained to a consistent set of observations, providing a coherent and spatially and temporally LIN ET AL. LIN: 21ST CENTURY ARIDITY 5796

3 Table 1. A List of CMIP5 GCMs Used in This Study a Model Name Origin 1 BCC-CSM1.1 Beijing Climate Center, China 2 BNU-ESM College of Global Change and Earth System Science, Beijing Normal University, China 3 CanESM2 Canadian Centre for Climate, Canada 4 CESM1-CAM5 National Center for Atmospheric Research, USA 5 CNRM-CM5 Centre National de Recherches Meteorologiques, France 6 CSIRO-MK3.6 Commonwealth Scientific and Industrial Research, Australia 7 GFDL-CM3 Geophysical Fluid Dynamics Laboratory, USA 8 GFDL-ESM2G Geophysical Fluid Dynamics Laboratory, USA 9 GFDL-ESM2M Geophysical Fluid Dynamics Laboratory, USA 10 GISS-E2-R NASA Goddard Institute for Space Studies, USA 11 HadGEM2-CC Met Office Hadley Centre, UK 12 HadGEM2-ES Met Office Hadley Centre, UK 13 INM-CM4 Institute for Numerical Mathematics, Russia 14 IPSL-CM5A-MR Institut Pierre-Simon Laplace, France 15 IPSL-CM5B-LR Institut Pierre-Simon Laplace, France 16 MIROC5 The Model for Interdisciplinary Research on Climate, Atmosphere and Ocean Research Institute, Japan 17 MIROC-ESM Japan Agency for Marine-Earth Science and Technology, Japan 18 MIROC-ESM-CHEM Japan Agency for Marine-Earth Science and Technology, Japan 19 MRI-CGCM3 Meteorological Research Institute, Japan a The historical run ( ) and a future scenario (RCP8.5) run ( ) form each model are used. The first ensemble run is used if a model has multiple ensemble runs. consistent representation of the state of the climate system, but subject to changes in the observations used as input. The monthly reanalysis data used in this study are obtained from the European Reanalysis (ERA-Interim) [Dee et al., 2011] and the National Centers for Environmental Prediction (NCEP)/Nation Center for Atmospheric Research (NCAR) reanalysis [Kalnay et al., 1996]. ERA-Interim is the latest global atmospheric reanalysis produced by the European Centre for Medium-Range Weather Forecasts covering the period from 1979 onward, with a spatial resolution of 1.5 (longitude) 1.5 (latitude) [Dee et al., 2011]. The NCEP/NCAR reanalysis project is using an older analysis/forecast system to perform data assimilation from 1948 to the present with a spatial resolution of (longitude) (latitude). A large subset of this data is available in daily averages [Kalnay et al., 1996] PET PET cannot be measured in the field, rather it is estimated from the meteorological variables. The Food and Agriculture Organization (FAO) Penman-Monteith equation [Monteith, 1965; Maidment, 1993; Allen et al., 1998] is an accepted method and calculates the PET in the form: PET ¼ 0:408Δ ð R n GÞþγ 900 Tþ273 u 2ðe s e a Þ (1) Δ þ γð1 þ 0:34u 2 Þ where R n is the net downward radiation; G is the heat flux into the ground; T is the mean daily air temperature at 2 m height; u 2 is the wind speed at 2 m height; e s is the saturation vapor pressure; e a is the actual vapor pressure; e s e a is the vapor pressure deficit; Δ is the slope of the saturation vapor pressure curve; and Γ is the psychrometric constant. Many studies use the mean temperature ( T mean in below) which is defined as the average of hourly temperature measurements because many reanalysis data just provide this kind of data, to calculate the saturation vapor pressure e s 17:27T mean e s ¼ 0:618 exp (2) T mean þ 237:3 According to FAO [Allen et al., 1998], it is better to compute the e s using the daily maximum and minimum air temperature (T max and T min ), as defined below h i h i 17:27T max 17:27T min 0:618 exp T maxþ237:3 þ 0:618 exp T minþ237:3 e s ¼ (3) 2 LIN ET AL. LIN: 21ST CENTURY ARIDITY 5797

4 Using mean air temperature instead of daily maximum and minimum temperatures results in a lower estimate for the mean saturation vapor pressure e s, used to get the vapor pressure deficit e s e a. What is more, for standardization, the mean air temperature for 24 h periods (T mean )isdefined as the mean of the daily maximum (T max ) and minimum temperatures (T min ). T max and T min for longer periods such as months are obtained by dividing the sum of the respective daily values by the number of days in the period [Allen et al., 1998]. The mean air temperature is employed in the FAO Penman-Monteith equation to calculate the slope of the saturation vapor pressure curves Δ. Using T mean instead T mean results in a different Δ. Although it results in slightly lower PET, we choose to use equation (3), the average temperature T mean to calculate e s and PET because it facilitates comparisons between the reanalysis data and model data. But we will explore the impact of using equation (3) over equation (2) quantitatively in section 3.3. For R n G (surface heat flux), we use the sum of the Latent and Sensible heat (LH + SH) because G is not available [Scheff and Frierson, 2014; Fu and Feng, 2014] Sensitivity of PET and AI We use stepwise linear regression, similar to Zhang et al. [2009], to analyze the relationship between PET and the climate variables used to calculate it [Fan and Thomas, 2013; Liu et al., 2013]. Stepwise linear regression uses the magnitude of the correlation coefficient between a given factor and PET to estimate the role of this factor on the variations of PET [Gao et al., 2006]. We estimate the sensitivity coefficients (S(x)) of annual average PET in response to a given climate variables x, e.g., surface temperature (T), relative humidity (RH), wind speed (WS), and net surface flux (FLX). The method is used for each individual ensemble member of the CESM-LE or reanalysis system and is detailed as follows: 1. Based on monthly data from 1980 to 2005, we first calculate the annual mean T i,rh i,ws i, and FLX i for each year (i = 1, 26) and their average values in the 26 years, represented as T 0,RH 0,WS 0, and FLX 0. The standard deviation (σ T, σ RH, σ WS, and σ FLX ) is calculated from the annual mean values in each year (X i ) minus the annual mean climatology (X o ) to represent the interannual variability. 2. The standard annual average PET 0 is calculated by equation (1) using T 0,RH 0,WS 0,andFLX 0. Thus, PET 0 = f(t 0,RH 0,WS 0, and FLX 0 ). 3. PET(x + σ x ) (PET(x σ x )) is calculated by adding (subtracting) the standard deviation of a given climate factor σ x to the long-term average x while keeping others factors unchanged. For example, on the temperature, PET(T 0 + σ T )=f(t 0 + σ T,RH 0,WS 0, FLX 0 ) (PET(T 0 σ T )=f(t 0 σ T,RH 0,WS 0, FLX 0 )). 4. The sensitivity coefficients, S(x), for climate variables x are calculated as below Sx ðþ¼50% * absðpetðx þ σ x Þ PET 0 Þ=PET 0 þ 50% * absðpetðx σ x Þ PET 0 Þ=PET 0 5. The above procedure is repeated until the S all input climate variables are calculated to estimate the sensitivity of AI to variations of T, RH, WS, FLX, and P (precipitation). 3. Results 3.1. Model Comparisons With Reanalysis Data Figure 1 shows the global distribution of AI averaged during derived from CESM-LE (Figure 1a), CMIP5 (Figure 1b), ERA (Figure 1c), and NCEP/NCAR (Figure 1d) reanalysis. For a comparison purpose, we also show the AI distribution (Figure 1e) from Feng and Fu [2013] who used observed temperatures and precipitation from the Climate Prediction Center (CPC), and the solar radiation, specific humidity and wind speed from the Global Land Data Assimilation System (GLDAS). Simulated AI broadly agrees with the reanalyses and previous work [IIASA/FAO, 2003; Mortimore, 2009; Feng and Fu, 2013]. The ensemble mean of CESM-LE overestimates AI (relative to reanalysis data) over northeastern China, Australia, southern Africa, and to a lesser extent tropical Africa. CESM-LE AI may be overestimated because the model overestimates precipitation over those regions (not shown). Despite these biases between model and reanalysis, the model is broadly able to simulate the climatology of annual mean AI. The spatial rootmean-square (RMS, defined as the square root of the arithmetic of the squares of the values) difference of the AI components P and PET is shown in Table 2 to give pattern correlations of simulations and reanalysis LIN ET AL. LIN: 21ST CENTURY ARIDITY 5798

5 Journal of Geophysical Research: Atmospheres Figure 1. Global distribution of AI for climatology based on (a) CESM-LE simulations, (b) CMIP5 simulations, (c) ERA, (d) NCEP reanalysis data, and (e) CPC-GLDAS data. Drylands are characterized by AI < 0.65 [United Nations Convention to Combat Desertification, 1994] and further divided into hyperarid (AI < 0.05), arid (0.05 < AI < 0.2), semiarid (0.2 < AI < 0.5), and dry subhumid (0.5 < AI < 0.65) regions. Purple boxes indicate focus regions: (1) northwest USA: 30 N 40 N, 240 E 250 E; (2) North Africa: 30 N 40 N, 0 E 10 E; (3) Sahel: 10 N 20 N, 10 E 20 E; (4) northwestern China: 30 N 40 N, 100 E 110 E; (5) Australia: 30 S 20 S, 130 E 140 E; (6) East Africa: 0 N 10 N, 40 E 50 E; (7) South Africa: 20 S 10 S, 20 E 30 E; and (8) Patagonia: 50 S 40 S, 290 E 300 E. to quantify the model performance. We define the global land as terrestrial regions between 60 S and 60 N. The RMS difference between model and reanalysis data is of similar magnitude to the RMS differences between ERA and NCEP reanalysis data sets. Specifically, the RMS of CESM-LE precipitation is nearly the same with respect to ERA (1.20) as NCEP is to ERA (1.21). CESM-LE simulations have smaller RMS (spatial root-mean-square) in PET than P (Table 2). The CMIP5 ensemble mean has slightly smaller RMS relative to reanalysis data than that of CESM-LE (Table 2). To evaluate the temporal variations of aridity, average AI is calculated over eight arid/semiarid regions (marked in Figure 1) and all global land. Regional averages of PET and P are used to then calculate a regional value for AI. The arid/semiarid regions are analyzed because they are very sensitive to global changes due to fragile ecosystems [Huang et al., 2012; Rotenberg and Yakir, 2010]. Table 2. The Root-Mean-Square (RMS) Error Between CESM Model, CMIP5 Model, ERA, and NCEP Data Over Land ( 60 S to 60 N) Root-Mean-Square Error of Precipitation (mm/d)/pet (mm/d) CESM-LE CMIP5 ERA NCEP LIN ET AL. CESM-LE CMIP5 ERA NCEP 0.79/ / / / / /0.57 LIN: 21ST CENTURY ARIDITY We use the ensemble mean of CESM-LE to derive the model simulated seasonal cycle of AI across the globe. We plot the seasonal cycle variations of averaged AI of in Figure 2. One standard deviation is also calculated to evaluate the uncertainties associated with internal variability among the ensemble members. 5799

6 Figure 2. (a i) Annual cycle of monthly mean AI averaged over in the focus areas (see Figure 1 for definitions) from CESM-LE (black line and blue shading), ERA (green), and NCEP (red) reanalysis data. Black lines are the ensemble mean of 30 CESM ensembles. The grey blue shading denotes 1 standard deviation of 30 ensembles, and the thin black lines are the maximum and minimum monthly average of the ensemble members. The seasonal cycle of AI is well captured by CESM (Figure 2). AI exhibits strong seasonal fluctuations since it is affected by many meteorological variables. P and PET also exhibit strong seasonal fluctuations (not shown). AI has a large annual cycle in the northwest USA (Figure 2a) and South Africa (Figure 2g), while the AI changes in North Africa (Figure 2b) and Australia (Figure 2e) are small. Additionally, different areas have different peaks in the annual cycle. For example, the maximum monthly mean AI value occurs in June and July in Patagonia (Figure 2h) but in August in the Sahel (Figure 2c), and the minimum AI value occurs between April and September in South Africa (Figure 2g). The temporal variations of annual mean average AI anomalies of CESM-LE (from 1980 to 2080), ERA, and NCEP (from 1980 to 2005) in the study areas (relative to the baseline period of ) are illustrated in Figure 3. One standard deviation is also used to measure the internal variability between the ensemble members. Generally speaking, CESM-LE is able to reproduce the observed temporal variability in (Figure 3). The anomalies of CMIP5 (not shown) are similar to CESM-LE. The variance in the ensemble spans the range of the reanalysis results, and the variance across the ensemble members is quite different between regions in Figure 3. Figure 3 indicates that the CESM-LE is capturing the variability appropriately. Note, however, the big differences between the reanalyses in the Sahel (Figure 3c), where the interannual variability in NCEP is nearly zero, while for ERA it is larger than the spread in the CESM-LE. Both reanalysis, CESM-LE, and CMIP5 suggest an increasing aridity in semiarid regions, consistent with previous work [Fu et al., 1999; Ma and Fu, 2007; Huang et al., 2012]. On a global basis, CESM-LE projects increasing terrestrial aridity (AI decrease) in the 21st century, consistent with CMIP5 and previous studies [Feng and Fu, 2013]. However, aridity trends are not globally uniform. On a regional basis, the model predicts decreases in aridity in some regions like the Sahel (Figure 3c) and East Africa (Figure 3f). We will quantitatively analyze these changes in section 3.4. LIN ET AL. LIN: 21ST CENTURY ARIDITY 5800

7 Figure 3. (a i) Temporal variations of annual mean average AI anomalies in the focus areas and global based on CESM-LE (black), ERA (green), and NCEP (red) reanalysis. Black lines are the maximum and minimum of the 30-member ensemble. The grey blue shading denotes 1 standard deviation from 30 CESM simulations. As with AI, the modeled temporal variations of surface temperature (T), precipitation (P), and PET are similar to reanalysis data sets. Table 3 shows the standard deviation of interannual variations of annual mean T, P, PET, and AI during derived from reanalysis CESM-LE and CMIP5. Interannual standard deviations are comparable between reanalysis and CESM. This is also indicated qualitatively in Figure 3, where the model ensemble and standard deviation are shown with the reanalysis time series in each region: the reanalysis variation generally falls within the variability envelope of the CESM-LE. The similarity between CESM-LE and CMIP5 values are similar to reanalysis data, suggesting that the model ensembles estimate the interannual variability of these quantities quite well compared to the reanalysis. Figure 4 shows the linear trends of annual mean T, P, PET, and AI during derived from reanalysis and the model. Even between reanalysis, the trends sometimes do not even agree in sign. The temperature trend of NCEP shows a decrease 0.08 C/yr in North Africa and 0.09 C/yr in Australia (Figure 4a) when the other three data show positive trends. The CESM-LE AI trend in Australia is 0.15 ± 0.67%/yr, while the trend of CMIP5 is 0.15 ± 0.88%/yr. The trends from the reanalyses should be taken with extreme cautious because of the inhomogeneity of the observational data used in terms of both quality and amount. Table 3. Standard Deviation (σ) of the Annual Mean Surface Temperature ( C), Precipitation (mm/d), PET (mm/d), and AI for From CESM-LE Simulations, CMIP5 Simulations, and Reanalysis (CESM-LE/CMIP5/ERA/NCEP) a CESM-LE/CMIP5/ERA/NCEP σ_t( C) σ_p(mm/d) σ_pet(mm/d) σ_ai Northwest USA 0.69/0.60/0.53/ /0.29/0.21/ /0.12/0.11/ /0.11/0.06/0.06 North Africa 0.48/0.60/0.60/ /0.11/0.12/ /0.12//0.15/ /0.03/0.03/0.02 Sahel 0.50/0.47/0.48/ /0.16/0.41/ /0.13/0.24/ /0.03/0.08/0.01 Northwestern China 0.43/0.47/0.47/ /0.23/0.18/ /0.07/0.07/ /0.16/0.09/0.15 Australia 0.62/0.70/0.50/ /0.31/0.21/ /0.27/0.23/ /0.08/0.04/0.04 East Africa 0.28/0.35/0.25/ /0.27/0.25/ /0.14/0.12/ /0.06/0.05/0.06 South Africa 0.43/0.48/0.41/ /0.33/0.34/ /0.13/0.15/ /0.10/0.09/0.12 Patagonia 0.41/0.40/0.48/ /0.17/0.19/ /0.08/0.14/ /0.09/0.06/0.09 Global Land 0.26/0.30/0.29/ /0.05/0.03/ /0.04/0.05/ /0.02/0.02/0.02 a The CESM-LE and CMIP5 results are the mean of the standard deviation from each ensemble simulations. LIN ET AL. LIN: 21ST CENTURY ARIDITY 5801

8 Figure 4. Trends of annual mean average (a) surface temperature (T), (b) precipitation (P), (c) PET, and (d) aridity index (AI) anomalies in the study areas for based on CESM-LE (red), CMIP5 (blue), ERA (black plus), and NCEP (black star). The red (blue) error bar denotes 1 standard deviation of CESM-LE (CMIP5) ensembles. To estimate the influence of natural climate variability on simulated aridity, we also show the global distribution of the standard deviation (σ) of simulated P, PET, and AI based on the 30 CESM-LE ensemble simulations (Figure 5). The standard deviation of precipitation of global mean is 0.05 mm/d, much bigger than the standard deviation of PET (0.02 mm/d). The maximum σ of P can be over 0.6 mm/d, creating over 100% variations in AI around the Himalayan mountain range (Figure 5c). LIN ET AL. LIN: 21ST CENTURY ARIDITY 5802

9 Journal of Geophysical Research: Atmospheres Figure 5. Global distribution of 1 standard deviation of (a) precipitation (P), (b) PET, and (c) AI climatology based on the ensemble of 30 CESM-LE simulations for Sensitivity of PET and AI on Interannual Time Scale What drives the variability and changes in PET and AI? To understand the sensitivity of PET and AI to different input climate variables, we analyze sensitivity coefficients for each of the input climate variables as described in section PET Sensitivity by Meteorological Variables Figure 6 illustrates the zonal of relative changes (in percent) of annual average PET in response to changes in individual variables based on CESM-LE (Figure 6a) and reanalysis data (Figures 6b and 6c). For CESM, the maximum and minimum of every member of the 30-member ensemble are shown (Figure 6a). Note that at the zonal mean level, the variance across the ensemble is small. The NCEP/NCAR flux sensitivity does not agree well with ERA, especially over the polar regions. Generally, the roles of relative humidity (RH), net flux (FLX), and surface air temperature are comparable for the PET. CESM-LE results agree with the reanalysis that RH and FLX are the most important factor controlling the PET variations in the subtropics and higher latitudes, while FLX is the major factor for PET variations in the tropics AI Sensitivity by Meteorological Variables Figure 7 shows the zonal mean of relative changes of annual average AI for each variable based on CESM (Figure 7a) and reanalysis data (Figures 7b and 7c). The maximum and minimum of every variable of the 30-member ensemble was added to Figure 7a. Generally, sensitivity of model simulated AI agrees with reanalysis data, except NCEP/NCAR relative humidity sensitivity over the north polar regions and net flux (FLX) over the polar regions. Precipitation (P) is the most important variable for AI. Relative humidity (RH) and net flux (FLX) sensitivity for AI are similar to the sensitivity of PET to these variables. FLX is the major factor for AI variations in the tropics. LIN ET AL. LIN: 21ST CENTURY ARIDITY 5803

10 Figure 6. Zonal mean of the sensitivity coefficients of PET for different elements based on (a) CESM-LE simulations, (b) ERA, and (c) NCEP data. These elements include T: surface temperature (blue), RH: surface relative humidity (red), WS: wind speed (green), FLX: net flux in the surface (brown). In Figure 6a, thin lines (blue, red, green, and brown) are the maximum and minimum of the variable (T, RH, WS, and FLX) sensitivity for the 30-member CESM ensemble. The contributions of each climate variable to the variations of annual average AI in CESM-LE at each location are shown in Figure 8. Relative humidity (RH) has the largest impact on AI variations over high northern latitudes, North America, western Russia, and Australia, while the impacts of T are weak in the tropics. Wind speed (WS) generally has a weak impact on AI (less than 2%) over most of the land. Surface net flux (FLX) has a little influence in high northern latitudes and western China. Precipitation (P) is the most important (can be 30%) of the five variables to AI. As seen in Figure 1, P is important in regions of low aridity. Although both T and FLX have impacts in the northern temperate latitudes, AI is more sensitive to RH than T and FLX. These sensitivities do not change with the simulated time period chosen. There is no change between three periods: the reference period ( ), (near future), and (far future) (not shown) Sensitivity of T mean and T min _T max As discussed in section 2, the daily mean temperature can be calculated as the arithmetic average of the daily maximum and minimum temperatures [Allen et al., 1998] or the average of hourly temperature, as in CESM. The saturation vapor pressure (e s ) can be calculated using the daily minimum and maximum temperature or the daily mean temperature. The latter method may lead to lower saturation vapor pressure, which in turn may result in lower PET and larger AI. Figure 9 compares the PET and AI calculated using the daily mean temperature (the standard method) and the daily minimum and maximum temperature. Figure 7. Zonal mean of the sensitivity coefficients of AI for the different elements based on (a) CESM-LE simulations, (b) ERA, and (c) NCEP reanalysis data. These elements include T: surface temperature (blue), RH: surface relative humidity (red), WS: wind speed (green), FLX: net flux in the surface (brown), P: precipitation (black). In Figure 7a, thin lines (blue, red, green, brown, and black) are the maximum and minimum of the variable (T, RH,WS, FLX,and P) sensitivity for the 30-member CESM ensemble. LIN ET AL. LIN: 21ST CENTURY ARIDITY 5804

11 Journal of Geophysical Research: Atmospheres Figure 8. AI sensitivity coefficients for the different variables from CESM-LE simulations. These elements include (a) T: surface temperature, (b) RH: surface relative humidity, (c) WS: wind speed, (d) FLX: net flux in the surface, and (e) P: precipitation. The global mean difference between these methods of estimating daily temperature for calculating PET in CESM-LE results in 6% differences in PET (Figure 9a). This in turn leads to 8% higher AI (Figure 9b). Although the AIs of Northern Canada and Northern Europe are both larger than 0.65 (Figure 1), due to sensitivity to T, the errors in Northern Canada are higher (10% more) because the sensitivity of RH in these two areas is different (Figure 9b). The difference in PET can be up to 20% in some areas. Over most land regions, the AI differences ranges from 5% to 10%. That is important since AI can be as small as 0.01, and the classification of drylands are characterized by AI [Mortimore, 2009; Hulme, 1996; Middleton and Thomas, 1992]. Results indicate that Tmin and Tmax should be used cautiously to define PET because the absolute value of PET might change. Here we have used the average temperature to calculate es and PET for consistency with the reanalysis data (NCEP/NCAR and ERA provide average T only). So we will focus just on the relative change and the anomalies of PET and AI The Future Projected Changes The temporal variations of annual mean average aridity index anomalies of ERA, NCEP, and CESM-LE averaged over the global and the eight regions during are shown in Figure 3. The long-term area average trends of surface temperature, precipitation, PET, and aridity index (AI) of CESM-LE and CMIP5 for in these regions are shown in Figure 10. For surface temperature and PET, the simulations agree well with each other (Figures 10a and 10c). There is a larger spread to the CMIP5 ensemble than the CESM-LE. There are larger discrepancies in regional precipitation trends (Figure 10b). For example, the CESM-LE precipitation trend in Australia is 0.11 ± 0.14%/yr when the trend of CMIP5 is 0.09 ± 0.33%/yr. Note that the values after ± are one sigma, and for the rest of the study. Based on CESM-LE, the deviation of P trend is bigger than P trend in northwest USA, Australia, South Africa, and Patagonia (Figure 10b) which means that the natural (internal) variability can impact the sign of P trends on region scale. Most of these areas indicate projected decrease AI (Figure 10d). Note that the internal variability should have little impact on LIN ET AL. LIN: 21ST CENTURY ARIDITY 5805

12 Journal of Geophysical Research: Atmospheres Figure 9. (a) PET and (b) AI sensitivity of T mean and Tmin_Tmax. the ensemble mean trend. Based on CESM-LE, a significant negative trend in the annual aridity index (drying) is projected over North Africa at the rate of 0.36 ± 0.10%/yr while a significant positive trend (moistening) is projected in East Africa at the rate of 0.36 ± 0.13%/yr. A very weak trend ( 0.05 ± 0.16%/yr) is expected over Australia. Only two of eight study areas (Sahel and East Africa) show positive (moistening) trends in AI. Overall, the aridity index average over global land is projected to decrease at the rate of 0.10 ± 0.02%/yr from the CESM-LE. Similar values are found globally for the CMIP5 ensemble of 0.12 ± 0.04%/yr. Figure 11 shows the projected changes in ensemble zonal mean surface air temperature ( C) (Figure 11a), precipitation (%) (Figure 11b), PET (%) (Figure 11c), and AI (%) (Figure 11d) between 2055 and 2080 (the far future ) and the reference period 1980 to 2005 based on CESM-LE (red) and CMIP5 (blue). One standard deviation between the ensemble members is shown as shading. As seen throughout this work, the CMIP5 multimodel ensemble has more variability than the single-model CESM-LE. CESM-LE and CMIP5 zonal mean T is expected to increase from 1.5 C to 5.0 C (depending on latitude) in the far future (Figure 11a). But CESM-LE zonal mean precipitation increases in the far future over all latitudes except from 30 S to 35 S and 45 S to 55 S when CMIP5 decreases in the far future over most of the Southern Hemisphere (Figure 11b). However, the CMIP5 and CESM ensembles are not generally statistically different (the ensemble means are generally within 1 standard deviation of the CMIP5 ensemble). The spread in the CMIP5 results indicates that model error could impact the sign of precipitation (shaded region includes zero). Increases in zonal mean precipitation (CESM-LE and CMIP5) are larger in the Northern Hemisphere and consistent between ensembles. The CESM-LE zonal mean PET increases from 8% to 28% in the far future (Figure 11c), CESM-LE and CMIP5 ensembles project decreases AI in the far future (Figure 11d) except from 15 N to 30 N. It should be noted that in terms of the ensemble mean trend, the result from CMIP5 is expected to be more reliable than CESM-LE. This is because the ensemble mean trend from CMIP5 not only minimizes the impact of internal variability but also the impact of model structural uncertainties, while the CESM-LE ensemble mean trend only minimizes the impact of internal variability and does not sample structural uncertainty. LIN ET AL. LIN: 21ST CENTURY ARIDITY 5806

13 Figure 10. Trends of annual mean average (a) surface temperature (T), (b) precipitation (P), (c) PET, and (d) AI anomalies in the study areas for based on CESM-LE (red) and CMIP5 (blue). The red (blue) error bar denotes 1 standard deviation of CESM-LE (CMIP5) ensembles. We show area average changes of surface temperature ( C), precipitation (%), PET (%), and AI (%) for relative to from CESM-LE and CMIP5 in Figure 12. Globally, both simulations project a decrease AI ( 6.4 ± 0.7% for CESM-LE and 7.3 ± 3.2% for CMIP5). However, the projected precipitation change in CESM-LE (6.9 ± 0.6%) is nearly double that from CMIP5 (3.8 ± 2.2%). On regional scales, two of the eight areas show the opposite signs in the projected change of precipitation between ensembles: LIN ET AL. LIN: 21ST CENTURY ARIDITY 5807

14 Figure 11. Projected zonal mean changes in (a) surface temperature ( C), (b) precipitation (%), (c) PET (%), and (d) AI (%) changes for relative to from CESM-LE (red solid lines) and CMIP5 (blue dash lines) simulations. The grey red (blue) shading denotes 1 standard deviation of CESM-LE (CMIP5) for Australia (8.5 ± 6.6% for CESM-LE/ 5.2 ± 16.5% for CMIP5) and Patagonia (4.4 ± 5.0% for CESM-LE/ 6.6 ± 7.2% for CMIP5). Random internal variations have large effect on the regional AI change, based on CESM-LE, 0.6 ± 7.6% in Australia and 5.3 ± 5.6% in Patagonia. To examine future components in more detail, we focus on CESM-LE from which we can isolate the impact of the internal variability, although the impact of the model structural uncertainty can be larger. The air temperature over land increases everywhere in both the near future (Figure 13a) and the far future (Figure 13b). A large increase appears over most of North America, northwestern Eurasia, and Eastern equatorial Africa in the near future (Figure 13a), and the difference becomes larger in the far future (Figure 13b). Precipitation in most of Eurasia, most of Australia, tropical Africa, and the middle- and high-latitude North America is expected to increase up to 30% (relative to mean) in the future (Figures 13c and 13d) but decrease as much as 30% (relative to mean) in the areas near the Mediterranean Sea, southwestern North America, Southern Africa, and South America. These projected precipitation changes are consistent with previous studies [Feng and Fu, 2013;Dai, 2011; Solomon, 2007]. The potential evapotranspiration (PET) over global land is expected to increase up to 20% in the near future (Figure 14a) with large increases in Europe and North America. However, the PET in India is expected to decrease slightly ( 2%) in the near future. PET increases of 10% to 40% (relative to present) are projected in the far future (Figure 14b), with larger values in Northern Hemisphere middle and high latitudes. The projected changes in precipitation and increase in PET contribute to projected changes in the aridity index (Figures 14c and 14d). AI decreases (drying) up to 20% over land between 60 N and 60 S. There are significant changes ( 10% to 20%) over most of Europe, North, and South America. AI increases (moistening) 5% to 30% over India, northern China, Eastern equatorial Africa, and the southern Saharan LIN ET AL. LIN: 21ST CENTURY ARIDITY 5808

15 Figure 12. Area average of (a) surface temperature (T), (b) precipitation (P), (c) PET, and (d) AI to projected changes for relative to from CESM-LE (red) and CMIP5 (blue). The red (blue) error bar denotes 1 standard deviation of CESM-LE (CMIP5) ensembles. regions in the near future (Figure 14c). Larger decreases (up to 50%) and larger increases (up to 50%) in AI are expected in the far future (Figure 14d), with the same spatial pattern seen in the near future. Generally, areas expected to dry (moisten) in the near future are expected to see more drying (moistening) in the far future. LIN ET AL. LIN: 21ST CENTURY ARIDITY 5809

16 Journal of Geophysical Research: Atmospheres Figure 13. Projected changes in (a, b) surface temperature ( C) and (c, d) precipitation (%) for (a, c) and (b, d) relative to under the RCP8.5 scenario from 30 CESM ensemble members. Stippled grid points indicate that at least 80% of the 30 simulations (24) agree on the sign of the change. It is clear that AI is expected to decrease (drying) in most of the study areas in the future (Figure 12d), but the decrease is not uniform across the annual cycle. Figure 15 shows annual cycle of mean average AI for , (near future), and (far future). The AI in Northern America during June to October is expected to change little; while the other months are expected to decrease (Figure 15a). The AI in Patagonia during April to October is expected to decrease when the other months do not (Figure 15h). Figure 14. Projected changes in (a, b) PET (%) and (c, d) aridity index (%) for (a, c) and (b, d) relative to from 30 CESM ensemble simulations. Grid points are stippled where more than 80% of the ensemble members (24) agree on the sign of the changes. LIN ET AL. LIN: 21ST CENTURY ARIDITY 5810

17 Journal of Geophysical Research: Atmospheres Figure 15. (a i) Annual cycle of mean average aridity index (AI) of (black), (green), (red) in the study areas based on the ensemble mean of 30 CESM simulations. Figure 16. Contribution of (a, b) precipitation and (c, d) PET to projected changes in AI (in percent) for (a, c) and (b, d) relative to under scenario RCP8.5 from CESM-LE simulations. Grid points are stippled where more than 80% of the 30 ensembles (24) agree on the sign of the changes. LIN ET AL. LIN: 21ST CENTURY ARIDITY 5811

18 Figure 17. Zonal mean percent contribution of precipitation (P) and PET to projected changes in aridity index (AI) for (a) and (b) relative to CESM ensemble simulations Reasons for the Aridity Index Change We quantify the roles of precipitation and PET in AI using ΔAI ¼ gðδpþþf ðδpetþ; ΔAI mean is the change of aridity index (AI in the future minus AI in the reference period, e.g., minus ), the index change caused by P is g(δp)and the AI change caused by PET is f(δpet). These functions can be written in the following form [Feng and Fu, 2013]: gðδpþ ¼ ΔP=PET f ΔPETÞ ΔPET*P=PET 2 þ P* ðδpetþ 2 =PET 3 or f ðδpetþ ¼ ΔðP=PETÞ ΔP=PETÞ Figure 16 shows the changes in AI for the near future and the far future due to changes in precipitation (Figures 16a and 16b) and PET (Figures 16c and 16d) by CESM-LE. Figure 17 shows the zonal mean (CESM-LE). The projected increases in precipitation lead to small increases in AI (moistening) over most global land areas. AI decreases (drying) near the Mediterranean Sea, southwestern North America, South America, most of Africa, and the west coast of Australia (where precipitation decreases). Largest decreases are projected in northern Africa in the far future (Figure 16b). It is clear that PET leads to a decrease in AI in everywhere (since PET increases everywhere) and the amplitude is expected to be large in the far future (Figures 16c and 16d). Basically, the contribution of P and PET to AI changes is similar between the near future and the far future. The change of PET is the major reason for the change of AI except for the region between 10 and 30 N (Figure 17), where the large projected increases in precipitation over Africa drive increases in AI. Table 4 quantifies the changes in AI and the contributions from P and PET, averaged over the eight regions based on CESM-LE. Over global land average, drying is expected (lower AI) in the near future and the far future. Some regions are projected to become moister (higher AI) in the near future. But most regions are expected to dry (lower AI) in the far future. The projected change of P leads to increasing AI in most regions except North Africa and South Africa. The change of projected PET leads to an AI decrease (drying) in all regions. The uncertainties in the projected PET change are small (±1), while the uncertainties in the projected change of P are large, leading to the relatively large uncertainly in the projected change of AI, specifically the relative change of AI in Australia (2.0 ± 7.2%/ 0.6 ± 7.6%), East Africa (4.0 ± 7.1%/18.5 ± 7.4%), and Patagonia ( 0.7 ± 5.4%/ 5.3 ± 5.6%). 4. Summary and Conclusions This study assessed the spatial and temporal variations of aridity index (AI) using various observational data sets for along with 30 model simulations from CESM and the multimodel ensemble from CMIP5 for Two future periods, defined as near future ( ) and far future ( ), were compared with current climate ( ). Multiple ensemble simulations from a model offers more information than a single-model simulation because internal variability of the model can be quantified by LIN ET AL. LIN: 21ST CENTURY ARIDITY 5812

19 Table 4. Area Average Contribution of Precipitation (P) and PET to Projected Changes in AI for / Relative to From CESM-LE Near Future/Far Future/(%) ΔAI g(δp) f(δpet) Northwest USA 5.7 ± 7.3/ 8.2 ± ± 6.8/4.5 ± ± 1.0/ 12.7 ± 0.9 North Africa 5.6 ± 5.0/ 21.2 ± ± 4.8/ 9.9 ± ± 0.5/ 11.3 ± 0.6 Sahel 6.6 ± 3.0/9.7 ± ± 2.7/19.2 ± ± 0.5/ 9.5 ± 0.5 Northwestern China 1.7 ± 2.6/ 4.6 ± ± 2.2/15.8 ± ± 0.8/ 20.4 ± 1.1 Australia 2.0 ± 7.2/ 0.6 ± ± 5.8/8.5 ± ± 1.3/ 9.1 ± 1.3 East Africa 4.0 ± 7.1/18.5 ± ± 6.1/23.2 ± ± 0.9/ 4.7 ± 0.9 South Africa 4.4 ± 2.7/ 12.8 ± ± 1.9/ 0.8 ± ± 1.1/ 12.0 ± 1.0 Patagonia 0.7 ± 5.4/ 5.3 ± ± 4.5/4.4 ± ± 1.3/ 9.7 ± 1.2 Global Land ( 60 S 60 N) 2.1 ± 0.6/ 6.4 ± ± 0.5/6.9 ± ± 0.2/ 13.3 ± 0.2 using the standard deviation across the ensemble members. This leads to better confidence in the model results and allows an assessment of sensitivity for different factors. The use of a multimodel (CMIP5) and single-model (CESM-LE) ensembles clearly shows the role of internal variability versus model structural uncertainty. CESM has been demonstrated to represent major modes of climate variability [Hurrell et al., 2013]. So if we assume that CESM is representative of the typical internal variability is in any individual CMIP5 model, then CESM is representative of typical internal variability in the CMIP5 ensemble. Other models could be used, but the CESM-LE is the largest such ensemble available. The much larger spread of the CMIP5 ensemble indicates that on these long time scales, structural uncertainty is larger than internal variability at global scales, consistent with the framework of Hawkins and Sutton [2009]. Thus, the ensemble mean trend from the CESM-LE is subject to a potential large impact of the model structural uncertainty. Results from reanalysis and historical simulations ( ) indicate significant discrepancies in annual aridity index trends. Even between with two reanalyses, there are large differences at regional scales. Overall, global land mean aridity has been decreasing at the rate of 0.18%/yr for Trends continue in the future. In the near future ( ), AI decreases up to 20% over most of global land regions between 60 N and 60 S. There are significant changes ( 10% to 20%) over most of Europe and northern South America. AI increases 5% to 30% over India, northern China, Eastern equatorial Africa, and the southern Saharan regions in the near future. In the far future ( ) CESM-LE projects larger decreases (up to 50%) and larger increases (up to 50%) in AI, with the same spatial pattern as in the earlier period. Generally, areas expected to dry (moisten) in the near future are expected to see more drying (moistening) in the far future. However, western North America is projected to dry in the near future but not the in the far future. Different areas have a different annual cycle of AI. Trends in AI come preferentially from months with higher AI. Increases in projected future precipitation cause moistening in the high northern latitudes, China, and central Africa; and precipitation declines lead to drying in the subtropics, northern and southern Africa, and the Amazon. The relationship between global precipitation and global temperature is approximately linear. Increased PET not only amplifies the precipitation-induced drying but also increase aridity (decreased AI) even when precipitation increases. We computed the sensitivity coefficients for annual average PET from four climate variables and for annual average AI from five climate variables in CESM-LE in interannual time scale. The sensitivity coefficients are different in different areas. For example, relative humidity is the most important factor affecting PET in the subtropics and higher latitudes, the net surface flux is the major factor for sensitivity in PET change in the tropics. AI is more sensitive to precipitation than to other variables over most of land. Furthermore, the AI sensitivity does not change with the simulated time period chosen. Finally, the average of hourly temperature measurements results in 6% lower PET and 8% higher AI than using daily maximum and minimum temperature. Natural variability can strongly impact the regional-scale precipitation and AI change in the future. For a given region, internal variability can change the sign of a signal. Trends are dominated by trends in precipitation. Variance in precipitation changes is smaller in the CESM-LE than in the CMIP5 multimodel ensemble. This is not surprising as precipitation is a complex process that is parameterized differently in different models. LIN ET AL. LIN: 21ST CENTURY ARIDITY 5813