Diagnostic analysis of interannual variation of global land evapotranspiration over 1982–2011: Assessing the impact of ENSO



[1] Global land evapotranspiration (E) between 1982 and 2011 was estimated by using a canopy conductance-based process E model (Air Relative Humidity-Based Two-Source model) [Yan et al., 2012]. To analyze the impact of precipitation forcing on E, an ensemble of six E data sets was derived from a driving ensemble of six precipitation data sets (i.e., Global Historical Climatology Network, Global Precipitation Climatology Centre, Climate Research Unit, Global Dataset of Meteorological Forcings, Global Precipitation Climatology Project, and Delaware). The result shows that ensemble average E over global land had an annual mean of 64.8 ± 0.8 × 103 km3 yr−1 and a significant linear trend of 4.6 mm per decade (p < 0.01). Significant partial correlations were found between the ensemble average E and its three controlling variables (i.e., precipitation (Pr), vegetation leaf area index (Lai), and potential evaporation (Ep)). These correlations explained 95% of the interannual variation of global land E with Pr as the dominant forcing contributing 37% variation of E; i.e., global land E was slightly sensitive to Pr than Lai and Ep. Pr, Lai, and Ep all showed increases of 8.8 mm (p < 0.01), 0.4 m2 m−2 (p < 0.01), and 2.0 mm (p < 0.1) per decade, respectively, which characterized a favorable environment for the increase of E over past 30 years. Both negative Multivariate El Niño–Southern Oscillation (ENSO) Index (MEI) and Southern Oscillation Index (SOI) displayed an increasing trend. The La Niña phase tended to be dominant from 1982 to 2011 and caused a significant increase of land Pr and further enhanced land E. Impacts of ENSO and corresponding Pr variation require attention to increase the understanding of the interannual variation of global land E.

1 Introduction

[2] Through biophysical processes including vegetation transpiration and soil evaporation, global land actual evapotranspiration (E) connects land surface water, energy, and carbon cycles and links the atmosphere to vegetation and soils in terrestrial ecosystems. As a key component in global land water cycle, E returns about 60% of annual land precipitation to the atmosphere [Oki and Kanae, 2006] and consumes more than half of absorbed solar energy [Trenberth and Fasullo, 2009].

[3] Knowledge of E is crucial to understanding how the water cycle has been impacted by climate change. Due to the lack of direct observations of land E at the global scale, many E models of varied complexity have been formulated based on different physical principles. These include surface conductance-based E models [Leuning et al., 2008; Mu et al., 2011; Zhang et al., 2010; Yan et al., 2012], energy-balance E models [Su, 2002; Kustas and Norman, 1999], coupled-stomata models for transpiration and photosynthesis [Ryu et al., 2011; Priestley and Taylor, 1972], equation-based E model (GLEAM) [Miralles et al., 2011b], and empirical E models [Wang and Liang, 2008; Zeng et al., 2012; Jung et al., 2010], such as the model tree ensemble approach (MTE) [Jung et al., 2010] based on a set of explanatory variables (i.e., remote sensing data and surface meteorological data). These models have been evaluated with flux E data from subsets of the 400 available flux stations worldwide. Current estimation of global E from this suite of remote sensing-based models, as well as reanalysis (e.g., Modern Era Retrospective Analysis for Research and Applications (MERRA)) and off-line land surface models (e.g., Second Global Soil Wetness Project (GSWP-2)), ranges widely from 58 × 103 to 85 × 103 km3 yr−1 by GSWP-2 [Dirmeyer et al., 2006] and from 68 × 103 to 80 × 103 km3 yr−1 by the LandFlux-EVAL project [Mueller et al., 2011]. Along with these considerable differences, opposite trends in the change of the pattern of variation have been found in the period from 1984 to 2007 in different studies [e.g., Jung et al., 2010; Vinukollu et al., 2011]. Nevertheless, short-term analyses of global land E consistently show a decline from 1998 to 2008 [Jung et al., 2010; Vinukollu et al., 2011; Zhang et al., 2010; Zeng et al., 2012]. The first objective of the present study is to further examine E from 1998 up to 2011 to determine whether this decline continues.

[4] Intensification of global hydrologic cycle with warming temperatures has been confirmed based on evidence from precipitation and runoff data sets [Huntington, 2006; Alkama et al., 2011]. Both climate models and satellite precipitation observations indicate that global (land and ocean) total atmospheric water, precipitation, and ocean evaporation have had similar, increasing trends (1.2 to 1.4% decade−1) due to global temperature increase from 1987 to 2006 [Wentz et al., 2007] and from the Pacific decadal variability as well [Gu and Adler, 2012]. However, intense and opposite precipitation anomalies can be seen over land and ocean due to El Niño–Southern Oscillation (ENSO) events [Dai and Wigley, 2000; Trenberth et al., 2007].

[5] ENSO is regarded as the most important coupled ocean-atmosphere phenomenon causing climate (rainfall, temperature, vegetation, drought, flood, etc.) variability throughout the world on interannual time scales [Dai et al., 1997; Wolter and Timlin, 1998]. Warm ENSO events (El Niño) tend to decrease global land precipitation [Gu et al., 2007], but drought occurs during the warm El Niño and cold La Niña events of the ENSO phenomenon in different areas of the world [Vicente-Serrano et al., 2011]. Reference evapotranspiration has a higher value up to 17%–30% in La Niña years than that in El Niño years in warm climates of Iran [Sabziparvar et al., 2011] and in the Maipo River basin of Chile [Meza, 2005]. In addition, Poveda et al. [2001] pointed out that satellite normalized difference vegetation index (NDVI) exhibits strong negative anomalies during El Niño years in Colombia.

[6] Our second objective is to investigate whether ENSO events affect land E and contribute to the decline of E from 1998 to 2008 in addition to the soil moisture stress addressed by Jung et al. [2010]. Precipitation representing water supply Ep indicating atmosphere evaporation demand and vegetation Lai showing canopy status are key drivers of E from the Penman-Monteith E model [Monteith, 1965]. Thus, ENSO-related precipitation anomalies should affect land E. However, the impact of ENSO on land E remains unknown except at small-river-basin scale [Twine et al., 2005].

[7] The third objective in this study is to answer whether different precipitation data set affect the estimated trend of global land E. As global land E is more sensitive to Pr than net radiation perturbations [Schlosser and Gao, 2010] and uncertainty in Pr mostly translates to uncertainty in E [Nasonova et al., 2011], it is essential to analyze the impact of ENSO-induced precipitation anomalies on interannual variation of land E. However, due to large uncertainties of current land precipitation data sets [Mueller et al., 2013], precipitation trends should be interpreted with caution especially when deriving from a single precipitation data set [Jung et al., 2010].

[8] To reduce the error of single Pr data set, we built an ensemble of Pr including six Pr member data sets that further produces six E ensemble members through Air Relative Humidity-Based Two-Source (ARTS) E model, which makes it possible to analyze the impact of Pr on E and finally give a reasonable estimation of ensemble average E with reduced error associated with input Pr of ensemble members. Thus, Pr uncertainty was mainly captured in this study while the rest of driving data sets came from a single reanalysis data set and not an ensemble.

[9] This study is the first long-term diagnostic analysis of the ENSO impact on global land E. An ensemble of global land E was generated at a monthly temporal scale and a 0.5° spatial scale for 1982–2011 using ARTS E model driven with six precipitation climate products and NASA MERRA reanalysis data. To provide a general picture of E variation associated with El Niño and La Niña events, interannual variation of E was analyzed against precipitation variation and two acceptable ENSO climate indices showing the strength of ENSO. Such knowledge will facilitate improved understanding of impact of global climate change on land water cycle.

2 Methods

[10] The Air Relative Humidity-Based Two-Source (ARTS) E model [Yan et al., 2012] was adopted to estimate global land E. With assumption of no water stress, ARTS E model first calculates total E (E0) as a sum of vegetation transpiration Ec and soil evaporation Es. Similarly, the available energy A is partitioned into two parts: canopy part (Ac) and the soil part (As). Further correction of E0 for soil water stress is conducted by using a soil water balance model. Evaluation against eddy covariance measurements at 19 flux sites, representing a wide variety of climate and vegetation types, indicated that monthly estimated E has an error statistics of root-mean-square error = 0.59 mm d−1, bias = −0.05 mm d−1, and R2 = 0.77. These are values comparable to other E models [Yan et al., 2012]. More detailed description of ARTS E model can be found in Yan et al. [2012].

2.1 Canopy Transpiration Ec and Canopy Conductance Gc

[11] The canopy transpiration (Ec) model is based on the Penman-Monteith model [Monteith, 1965], but the available energy (A) and surface conductance (Gs) terms are replaced by the canopy-absorbed available energy (Ac) and canopy conductance (Gc):

display math(1)
display math(2)

where Ac is the canopy available energy, Δ is the gradient of the saturated vapor pressure to air temperature, γ is the psychrometric constant, ρ is the density of air, Cp is the specific heat of air at constant pressure, Ga is the aerodynamic conductance, Gc is the canopy conductance accounting for transpiration from the vegetation, and D = es − ea is the vapor pressure deficit of the air, in which es is the saturation water vapor pressure at air temperature and ea is the actual water vapor pressure, Rh is the air relative humidity, and gsmax is the maximum stomatal conductance assumed to have a value of 12.2 mm s−1 [Kelliher et al., 1995].

2.2 Soil Evaporation Es

[12] Es equation is modified from the air relative humidity-based model of evapotranspiration (ARM-ET) [Yan and Shugart, 2010]:

display math(3)

2.3 Total Evapotranspiration E0 for Well-Watered Surface

[13] E0 represents evapotranspiration for well-watered surface:

display math(4)

2.4 Soil Water Correction Using Soil Water Balance Model

[14] As E0 equals actual E only for a well-watered surface, a correction to E0 is required for a water-stressed surface. Thus, a soil water balance model developed by Thornthwaite and Mather [1955] is adopted in ARTS E model to scale E0 to actual E.

3 Data Sets and Preprocessing

3.1 MERRA Reanalysis Data

[15] Modern Era Retrospective Analysis for Research and Applications (MERRA) was developed by NASA using a major new version of the Goddard Earth Observing System Data Assimilation System Version 5 (GEOS-5), which focuses on historical analyses of the hydrological cycle aided by the NASA modern Earth Observing System (EOS) suite of satellite observations in a climate framework. MERRA produces temporally and spatially consistent analyses of atmosphere, land surface, and ocean surface variables at a spatial resolution of 0.5° latitude × 0.7° longitude from 1979 to present with significant improvements in precipitation and water vapor climatology [Reichle et al., 2011; Bosilovich et al., 2011]. It is ideal for investigating climate variability [Rienecker et al., 2011].

3.2 Six Global Land Precipitation Data Sets

[16] The Global Historical Climatology Network (GHCN) monthly precipitation data set was created for climate monitoring at National Climatic Data Center (NCDC) of National Oceanic and Atmospheric Administration (NOAA). Monthly precipitation anomalies with respect to the 1961–1990 climate value were calculated from over 20,590 stations from 1900 to present. Station anomalies were then averaged within each 5° by 5° grid box to obtain the gridded GHCN precipitation product [Peterson and Vose, 1997; Menne et al., 2012].

[17] The Global Precipitation Climatology Centre (GPCC), operated by National Meteorological Service of Germany under the auspices of the World Meteorological Organization (WMO), has generated a Full Data Reanalysis Product covering the period from 1901 to 2010 with a resolution of 0.5° by using an empirical interpolation method SPHEREMAP [Willmott et al., 1985] from an available GPCC station database (67,200 stations with at least 10 years of data) compiled from all available sources [Rudolf et al., 2011].

[18] The Climate Research Unit (CRU) at the University of East Anglia developed CRU 3.1 monthly climatic mean and time series of terrestrial surface climate for the period 1901–2009, which comprises seven climate elements (precipitation, mean temperature, diurnal temperature range, wet-day frequency, vapor pressure, cloud cover, and ground-frost frequency) [New et al., 2000]. The spatial coverage extends over all land areas, excluding Antarctica. The construction method ensures that strict temporal fidelity is maintained. Monthly CRU time series data show month-by-month variations in climate variables and allow the comparison of variations in climate with variation in other phenomena [New et al., 2000]. A 0.5° latitude/longitude gridded data set was adopted in this study.

[19] The Global Dataset of Meteorological Forcings (GDMP) was developed for land surface modeling by the Department of Civil and Environmental Engineering at the Princeton University. The data set includes precipitation, air temperature, surface pressure, specific humidity, wind speed, and downward long wave and short wave at surface and is currently available at a 1.0° monthly resolution for 1948–2008. It combines observations with the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP-NCAR) reanalysis to correct known biases in the reanalysis precipitation and near-surface meteorology [Sheffield et al., 2006].

[20] The Global Precipitation Climatology Project (GPCP) managed by NASA Goddard Space Flight Center was established by the World Climate Research Program (WCRP). It combines available satellite estimates including microwave estimates, infrared (IR) precipitation estimates, and additional low-Earth orbit estimates with monthly GPCC Precipitation Monitoring Product into a final merged product (V 2.2) covering global land and ocean at a 2.5° × 2.5° scale from 1979 to 2010 [Huffman et al., 2009].

[21] The Delaware terrestrial precipitation monthly time series (V 3.01) was developed at the Department of Geography at the University of Delaware. Monthly total precipitation measured by rain gauge was compiled from several updated sources such as GHCN2 for the years 1900–2010 with the resultant number of stations ranging from about 4100 to 22,000 globally. Based on a relatively dense network of stations, a background precipitation climatology was built and was then interpolated to a 0.5° by 0.5° grid resolution with aid of monthly total precipitation by using climatologically aided interpolation method [Willmott and Robeson, 1995] to increase the accuracy of spatially interpolation.

3.3 Global Inventory Modeling and Mapping Studies (GIMMS) Leaf Area Index Data

[22] A new global 15 day LAI data set at 8 km spatial resolution for the period July 1981 to December 2011 was generated from advanced very high resolution radiometer Global Inventory Modeling and Monitoring Study (GIMMS) NDVI3g data set using an Artificial Neural Network (ANN) model. The ANN model for generating the LAI data set was trained with overlapping GIMMS NDVI3g and best quality Moderate Resolution Imaging Spectroradiometer LAI data. The full temporal coverage GIMMS LAI3g data set was then generated using GIMMS NDVI3g data and the ANN model. The new GIMMS LAI3g data set was evaluated through direct comparison with field data and indirectly through (a) intercomparisons with similar satellite data products at biome and site scales, (b) testing for reproducing known relationships between LAI and climatic variables (temperature and precipitation), (c) canonical correlation analysis with ENSO/Arctic Oscillation indices, and (d) comparison to simulations from multiple dynamic vegetation models. These exercises resulted in establishing the validity and uncertainty of these new data sets. Further details can be found in Zhu et al. [2013].

3.4 ISLSCP II Global Gridded Soil Data

[23] Global 1° gridded surfaces of selected soil characteristics including maximum soil available water content (MAWC) for a soil depth of 0 ~ 150 cm was developed by the International Satellite Land Surface Climatology Project (ISLSCP) Initiative II project based on the International Geosphere-Biosphere Programme (IGBP)–Data and Information Services (DIS) soil data [Global soil data task, 2000].

3.5 GSWP-2 E Data

[24] The Second Global Soil Wetness Project (GSWP-2) as a recent environmental modeling research activity of the Global Land-Atmosphere System Study (GLASS) produced first global gridded multimodel analysis of land surface state variables and fluxes spanning 10 years (1986–1995) on a 0.5° grid and a monthly time scale. The resulting analysis consisting of multimodel means and standard deviations has been applied to studies of global terrestrial energy and water balance and major components of E [Dirmeyer et al., 2006; Dirmeyer, 2011].

3.6 Multivariate ENSO Index (MEI) Data

[25] ENSO events and their strength have been monitored by using MEI index derived from six main observed variables over the tropical Pacific. These are sea level pressure, zonal and meridional components of the surface wind, sea surface temperature, surface air temperature, and total cloudiness fraction of the sky collected from the International Comprehensive Ocean-Atmosphere Data Set. Spatial filtering is initially applied to the individual fields and the first unrotated principal component (PC) of six filtered fields is regarded as MEI. Monthly MEI is computed for each of 12 sliding bimonthly periods (e.g., December/January, January/February). To keep the MEI comparable, all monthly MEI values are standardized based on a comparison with the 1950–1993 climate values [Wolter and Timlin, 1998].

3.7 Southern Oscillation Index (SOI) Data

[26] The SOI index also indicates the development and intensity of El Niño or La Niña events in the Pacific Ocean. It is a standardized anomaly of the mean sea level pressure difference between Tahiti and Darwin stations, which is usually calculated on a monthly basis at Australian Bureau of Meteorology. Further multiplication by 10 is their convention. Sustained negative values of the SOI greater than −8 often indicate El Niño episodes while sustained positive values of the SOI greater than +8 are typical of a La Niña episode [Nicholls, 1988].

3.8 Data Preprocessing

[27] All model forcings including six precipitation data sets, GIMMS Lai, MERRA reanalysis meteorological data (i.e., net radiation, air temperature, specific humidity, wind speed, roughness length, and displacement height), and maximum soil available water content were resampled to a 0.5o × 0.5o grid resolution by using a bilinear interpolation method and were then applied to driving the ARTS E model on a monthly time scale. GHCN Pr data sets only include anomalies and there is no available gridded GHCN Pr climate value. Thus, the CRU Pr climate value for the period 1961 to 1990 was instead added to GHCN Pr anomalies to build GHCN monthly Pr.

[28] To analyze interannual variation of E, monthly values of E were summed to yearly values and hence annual mean (equation (5)) and yearly bias (equation (6)) of E were calculated for six ensemble members, respectively. Further sum of averaged annual mean E and yearly bias E of six ensemble members produced a yearly series of ensemble average E over past 30 years (equation (7)), which was then evaluated with GSWP-2 E data and previous studies:

display math(5)
display math(6)
display math(7)

where Ny is 30, 29, 28, 27, 29, and 29 years indicating the length of GHCN, GPCC, CRU, GDMP, GPCP, and Delaware Pr-derived E data sets, respectively. As a result, member number Nm is 1, 4, 5, and 6 for year 2011, 2010, 2009, and 1982 to 2008, respectively. Similarly, the ensemble of six precipitation data sets was processed.

[29] To analyze impact of atmosphere evaporation demand on E as a driving factor, Ep was calculated according to Priestley and Taylor [1972] (PT) equilibrium E model:

display math(8)

where variables have the same meanings to ARTS E model in section 2.2.

4 Results

4.1 Temporal and Spatial Variations of ARTS E

[30] The interannual variation of global ARTS ensemble average E from 1982 to 2011 (Figure 1a) clearly shows a significant increase with a trend of 0.46 mm yr−1 (p < 0.01), which coincides well with an increasing linear trend of 0.41 mm yr−1 (p < 0.01) given by MTE E model but for a shorter period of 1982–2008 [Jung et al., 2010]. Zeng et al. [2012] reported a higher, also increasing rate of 1.1 ± 0.2 mm yr−1 (p < 0.01) for E from 1982 to 2009.

Figure 1.

Interannual variation of (a) global land ARTS ensemble average E and (b) six ensemble members' E derived from respective driving precipitation data sets and corresponding slope k of linear trend from 1982 to 2011.

[31] E variation can be explicitly split into two periods (Figure 1a). During the first period (1982-1997), E had an increasing trend of 0.4 (p = 0.14) and 0.71 mm yr−1 (p < 0.01) indicated by ARTS and MTE models, respectively. Similarly, Yan et al. [2012], Vinukollu et al. [2011], and Zeng et al. [2012] reported an increasing trend of E over the 1980s and the 1990s. Whereas, during the second period (1998–2011) ARTS E shows no significant change due to two higher positive anomalies of 12.97 and 13.67 mm yr−1 occurred in the beginning (1998) and the end (2010) of the period, respectively (Figure 1a).

[32] However, a decreasing trend of 0.2 (p = 0.66) and 0.16 mm yr−1 (p = 0.75) was found over the interval of 1998–2008 by ARTS E and MTE E (Figure 1a). Earlier, Jung et al. [2010] based on an analysis of satellite microwave TRMM soil-moisture data found this trend and attributed it to limited moisture supply. However, the decreasing trend was not significant (p > 0.60) as shown here by both E models and other studies [Vinukollu et al., 2011; Zeng et al., 2012]. Thus, it can be regarded as a fluctuation accompanied with a lower negative E anomaly of about −6.0 mm yr−1 that occurred in 2002. Beyond this interval, ARTS E shows continued increase with a positive anomaly of 13.67 mm yr−1 in 2010. The decrease of E over 1998–2008 was temporary and does not reverse the increasing trend of E since 1982.

[33] Six E models, derived from respective Pr driving data set, all display similar interannual variation but different magnitude of anomalies (Figure 1b). Similar to the ARTS ensemble average E (Figure 1a), five E ensemble members had a significant increasing trend for the period of 1982–2011 (p < 0.05) except EGPCP driven with GPCP Pr. The climatic trend of six E members ranged from 0.3 mm yr−1 for EGPCP member to 0.59 mm yr−1 for ECRU member. In addition, the increase from 1982 to 1997 and decrease from 1998 to 2011 of E were all insignificant for six ensemble members similar to ARTS ensemble average E (Figure 1a).

[34] With regard to spatial pattern of E for the whole research interval of 1982–2011 (Figures 2a and 2b), the ARTS ensemble average E had an overall increasing trend for most of the global land area while significant decreasing trend still existed in western North America, Amazon, Middle East, Northeast of China, etc. However, during the first period of 1982–1997, most land area shows an insignificant trend of E (Figures 2c and 2d). Conversely, during the second period of 1998–2011, more land areas (e.g., Australia and Amazon) had a decreasing trend of E while tropical Africa showed an increasing trend of E (Figures 2e and 2f).

Figure 2.

Distribution of global trend of ARTS ensemble average E and its significance of linear trend for period of (a, b) 1982–2011, (c, d) 1982–1997, and (e, f) 1998–2011, respectively.

4.2 Analysis of Driving Factors Resulting in the Intensified E

[35] In ARTS E model, E0 represents total evapotranspiration under plentiful supply of water which is further scaled to actual E by using a soil water balance model. Thus, E0 and Pr determine the actual E. On a global scale, both E0 and ensemble average Pr had a close linear correlation (p < 0.01) with ARTS ensemble average E from 1982 to 2011, and furthermore, E0 and Pr can explain 73% and 40% variation of E, respectively (Figure 3). In other words, global land E was predominantly controlled by E0 rather than Pr. As E0 actually included contributions of atmosphere demand and vegetation Lai in ARTS E model, E0 can substitute for Ep and Lai as a comprehensive variable combined with Pr for analyzing impacts of model forcings on E.

Figure 3.

Scatterplot of anomalies of ensemble average (a) E versus E0 and (b) E versus Pr.

[36] The interannual variations of global land Pr and E0 from 1982 to 2011 (Figure 4) also show a significant increase with a trend of 0.88 (p < 0.01) and 0.51 mm yr−1 (p < 0.01), respectively, which produced an increasing trend of E = 0.46 mm yr−1 (p < 0.01). Trends in E were consistent with increasing trends in Pr and E0. Jung et al. [2010] also showed that Pr and E had consistent trends in research domain from 1998 to 2008. Figure 4 also indicates that the trend of E was lower than that of two driving factors, especially Pr, which might be due to that ecosystem had an reduced response to severe changes of Pr through its complex ecohydrological processes (e.g., soil water bank).

Figure 4.

Interannual variations of global land E0, ensemble average Pr and ARTS E from 1982 to 2011 and corresponding slope K of linear trend.

[37] The comparisons of anomalies of ensemble average Pr, E0, and E (Figure 4) indicate that there probably existed a complementary relationship between E0 and Pr in determining E on a global scale especially when anomalies of E0 and Pr were large. For instance, annual precipitation was low in 1987 with a negative anomaly of Pr = −38.46 mm yr−1, but E0 only had a minor anomaly of −1.94 mm yr−1, which resulted in an anomaly of E = −8.52 mm yr−1, falling within anomalies of Pr and E0. In another case with plentiful precipitation in 2000, even though E0 had a negative anomaly of −5.55 mm yr−1, the complementary effect due to plentiful precipitation with a positive anomaly of Pr = 29.85 mm yr−1 produced an anomaly of E = 5.23 mm yr−1. Note that E cannot exceed E0 at any grid due to limitation by soil water balance model.

[38] Figure 4 indicates that there were three sets of years of 1997–1998, 1999–2000, and 2001–2002 representing different combination of Pr, E0, and E anomalies. The years 1997–1998 featured positive anomalies of annual mean Pr, E0, and E; 1999–2000 had positive anomalies of Pr, negative anomalies of E0, and a normal E while 2001–2002 featured negative anomalies of annual mean Pr, E0, and E. Their spatial patterns show that there was plentiful Pr in tropical regions for 1999–2000 (Figure 5e) compared with 1997–1998 (Figure 5d). Conversely, E0 shows an opposite pattern compared with Pr; there was less E0, e.g., energy limitation, in tropical areas for 1999–2000 (Figure 5h) compared with 1997–1998 (Figure 5g). As a result of energy limitation, there was less E in tropical areas for 1999–2000 (Figure 5b) compared with 1997–1998 (Figure 5a), which indicates that those regions that were contributing more to the global land Pr anomalies were located in energy-limited regions (e.g., tropical regions in Figure 6e) and that is why there was not much Pr contribution to E during 1999–2000.

Figure 5.

Pattern of (a–c) E, (d–f) Pr, and (g–i) E0 anomalies for 1997–1998, 1999–2000, and 2001–2002.

Figure 6.

Pearson's correlation significance of (a) E versus E0, (b) E versus PT Ep, (c) E versus Pr, and (d) E0 versus Pr. (e) Map where E0 demand or Pr supply limitation controls interannual variation of E. (f) Traditional map using Ep instead of E0 as demand factor.

[39] In addition, 2001–2002 shows a different spatial pattern (Figures 5c, 5f, and 5i); Pr and E0 had negative anomalies in more regions (Figures 5f and 5i) than those for 1997–1998. Thus, limited E0 demand and Pr supply resulted in more regions with negative anomalies of E for 2001–2002 (Figure 5c) compared with 1997–1998 (Figure 5a).

[40] Figure 4 shows the comparisons of anomalies of E, E0, and Pr globally. The complimentary relationship occurred in 21 years (70% of the years studied) and exceptions still existed in 9 years (30%), e.g., annual E needed not be a complimentary anomaly relative to Pr and E0. For instance, Pr and E0 all had a negative anomaly of −4.99 and −4.44 mm yr−1 in 1982, respectively. This resulted in a weak negative anomaly of E = −3.08 mm yr−1. Similarly, positive anomalies of Pr = 26.0 mm yr−1 and E0 = 16.95 mm yr−1 in 2010 favored weak positive anomalies of E = 13.67 mm yr−1. The above analysis further reveals that ecosystem tended to have a weak response of E with respect to dramatic changes of the two driving factors (i.e., Pr and E0) on a global scale.

[41] The global variations of Pr, E0, and E agreed with common knowledge that both favorable Pr and E0 will produce a positive anomaly of E, or vice versa. In addition, complementary effects might reduce the amplitude of E anomaly when one factor had an opposite anomaly.

[42] To determine whether E0 or Pr controls E as a major limitation factor, Pearson correlation coefficients of E versus E0 and E versus Pr were calculated. Figures 6a and 6c shows that E was mainly determined by E0 on global scale and Pr controlled E primarily in water-stressed areas such as deserts. Thus, E0 was the major limitation factor of E over global land compared with Pr (Figure 6e). In contrast, it is well known that Ep often controlled E in tropical areas and Northern Hemisphere high-latitude areas (Figure 6b), which is consistent with Wang et al. [2010]. Thus, when conducting traditional forcing analysis by comparing Ep and Pr, one can derive a different map of major limiting factors (Figure 6f) that indicates more global land were controlled by Pr. There seems a paradox between Figures 6e and 6f. In fact, Figure 6f only considers the nonvegetation factors while Figure 6e reflects the impact of real vegetation on E by considering the canopy conductance; i.e., interannual variation of E was principally controlled by canopy conductance-based E0 in vegetation-covered land while in bare land, Pr controlled E through soil evaporation. In addition, intercorrelation analysis applied to E0 and Pr (Figure 6d) shows that E0 and Pr had negative relationships mainly in tropical areas and Northern Hemisphere high-latitude areas while significant positive relationships were often found in arid regions of southwest North America, southern South America, southern Africa, western Asia, and Australia.

[43] Global vegetation Lai and MERRA Ta (Figure 7) had an increasing trend of 0.04 m2 m−2 yr−1 (p < 0.01) and 0.024°C yr−1 (p < 0.01), respectively, while PT Ep had an increasing trend but insignificant in statistics (p = 0.09). Similarly, HadCRUT3 Ta data set shows a linear trend of 0.027°C yr−1 (p < 0.01) from 1979 to 2005 [Brohan et al., 2006]. Global increase of Lai (Figure 7) implied enhanced vegetation activity over the past three decades, which is consistent with previous studies of global greening since 1982 mainly because of improved critical climatic constraints to plant growth [de Jong et al., 2012; Nemani et al., 2003].

Figure 7.

Interannual variations of global land (a) GIMMS Lai, (b) PT Ep, and (c) Ta from 1982 to 2011.

[44] ARTS ensemble average E had significant Pearson correlation with vegetation Lai (R2 = 0.46, p < 0.01), Pr (R2 = 0.40, p < 0.01), and PT Ep (R2 = 0.22, p < 0.01), respectively. From the view of vegetation regulation, water supply, and atmosphere demand that determined E, all driving factors, i.e., Lai, Pr, and PT Ep, had an increasing trend in the 1982 to 2011 period (Figures 7 and 4). Thus, the recent increasing trend of global E can be primarily attributed to increasing vegetation Lai, water supply, and atmosphere evaporation demand.

4.3 Uncertainties of Precipitation Variation and Its Impact on E

[45] Figure 8a shows that four Pr ensemble members (i.e., GHCN, GPCC, CRU, and GDMP) significantly increased by 0.87 ~ 1.53 mm yr−1 for 1982–2011, while GPCP and Delaware Pr had an insignificant increase with a trend of 0.34 and 0.36 mm yr−1 (p > 0.05), respectively. Nickl et al. [2010] reported an increased trend of Pr (at rates of approximately 0.75 to 2.1 mm yr−1) over a decade from 1992 to 2002 estimated from GPCC, CRU, and Delaware Pr data sets. Figure 8a also implies that there existed large uncertainties in current Pr data sets. Six Pr ensemble members had a wide range of annual mean values from 99.2 ± 2.1 to 112.7 ± 2.2 × 103 km3 yr−1 (Table 1).

Figure 8.

(a) Interannual variation of ensemble members' Pr for 1982–2011 and corresponding slope k and significance P of linear trend and (b) STDEV of ensemble average Pr and E calculated from ensemble members. Number of used ensemble members in a specific year is shown on top of x axis.

Table 1. Ensemble Average and STDEV of E and Pr and Associated Annual Mean of Six Ensemble Members for Global Land Excluding Antarctica and Greenland (Unit: 103 km3 yr−1)
 Ensemble AverageGHCNGPCCGPCPCRUGDMPDelaware
E64.8 ± 0.865.3 ± 0.864.2 ± 0.868.3 ± 0.863.7 ± 0.864.6 ± 0.863.8 ± 0.8
Pr102.8 ± 2.1103.4 ± 2.1103.7 ± 2.4112.7 ± 2.2101.1 ± 2.6102.5 ± 2.699.2 ± 2.1

[46] Comparison of standard deviation (STDEV) of ensemble average Pr and E calculated from corresponding six ensemble members (Figure 8b) indicates that ARTS E had an increasing trend of STDEV = 0.02 mm yr−1 that was lower than that of Pr (trend of STDEV = 0.05 mm yr−1), which have been seldom addressed by previous researches. As uncertainties of Pr, shown by the trend of STDEV for Pr, increased by 0.05 mm yr−1 that was obviously lower than the increasing trend of 0.46 mm yr−1 for ARTS ensemble average E, thus it can be concluded that uncertainties in Pr ensemble members could not substantially affect the increasing trend of global land E estimation.

4.4 Evaluation of ARTS Ensemble Average E

[47] We estimated an ensemble average E of 64.8 ± 0.8 × 103 km3 yr−1, comparable to recent estimates of 65 ± 3 × 103 [Jung et al., 2010], 65.5 × 103 [Oki and Kanae, 2006], 62.8 × 103 [Mu et al., 2011], 63 × 103 [Ryu et al., 2011], 67 × 103 [Trenberth et al., 2007], 69 × 103 [Vinukollu et al., 2011], and 67.9 × 103 km3 [Miralles et al., 2011a]. Schlosser and Gao [2010] also reported a GSWP-2 model-mean value of 65.1 ± 0.8 × 103 km3 yr−1 for global land excluding Antarctic.

[48] Six E ensemble members also had different annual E values with a narrow range of 63.7 × 103 to 68.3 × 103 km3 yr−1 (Table 1), which falls within the model range (49 × 103 to 82 × 103 km3 yr−1) estimated by the GSWP-2 project [Schlosser and Gao, 2010] and the recent model range (60 × 103 to 85 × 103 km3 yr−1) reported by the Water and Global Change project [Haddeland et al., 2011]. As ensemble average Pr for global land was 102.8 ± 2.1 × 103 km3 yr−1 (Table 1), the annual land E-to-Pr ratio was 0.63, which is close to reported values of 0.66 [Dirmeyer, 2011], 0.60 [Zhang et al., 2010], and 0.58 ± 0.9 [Alton et al., 2009]. ARTS ensemble average E (Figure 9) shows higher E over 1300 mm yr−1 mainly distributed in tropical American, African, and Asian areas due to plentiful supply of precipitation and heat. Whereas lower E less than 300 mm yr−1 often occurred in cold regions of Northern Hemisphere due to limited heat resources and in arid regions of Australia, Sahara, western and central Asia, etc., due to limited precipitation. Further comparison of ARTS E versus GSWP-2 E at grid scale (Figure 10) indicates a significant linear correlation (p < 0.01) with a slope k = 0.98. Significant correlation also existed on monthly comparisons (not shown). Above evaluations of ARTS ensemble average E show a reasonable spatial pattern of E.

Figure 9.

Spatial distribution of ARTS ensemble average E (1982–2011).

Figure 10.

Comparison of ARTS ensemble E versus GSWP-2 E.

5 Discussion and Conclusion

[49] According to the Monteith [1965] evaporation theory, land E is actually determined by water supply, atmosphere demand, and vegetation regulation that were expressed by Pr, Ep, and Lai, respectively, in this study, and their effects on E was further analyzed to understand the increasing trend of E since 1982. Statistics of interannual variation indicate that significant Pearson correlations were found between E versus Lai, Pr, and Ep with a determining coefficient R2 of 0.46, 0.40, and 0.22, respectively, which shows that vegetation as the dominant forcing explained 46% variation of E. However, as the relationship between E and its three controlling variables is typical of multivariable correlation analysis and Pearson correlation fits better for bivariate analysis, thus partial correlation, due to its ability of measuring the degree of association between two random variables with the influence of the remaining variables eliminated, was adopted in this study. The result indicates significant partial correlations existed between E versus Pr, Lai, and Ep with a determining coefficient R2 of 0.37, 0.33, and 0.24, respectively; i.e., these three forcings explained 95% interannual variation of global land E in which Ep only contributed 24% variation of E while Pr contributed 37% variation of E. The statistics show that global land E was slightly more sensitive to Pr than other perturbations, which is consistent with the result of Schlosser and Gao [2010] and a feedback process described by Dirmeyer et al. [2009] that Pr very strongly determines soil moisture globally and hence soil moisture moderately controls land E.

[50] As water supply to E fundamentally comes from Pr and ENSO controls the interannual variation of global (land and ocean) Pr by shifting precipitation patterns in the tropics and subtropics due to changes of sea surface temperature in Pacific [Trenberth, 2011; Curtis and Adler, 2000], variation of global land Pr can be largely attributed to ENSO activities often characterized with ENSO climate index [Gu and Adler, 2011; Dai and Wigley, 2000; Trenberth and Caron, 2000]. Interannual variation of MEI and SOI (Figure 11a) indicates that −MEI and SOI significantly increased (p < 0.05) from 1982 to 2011; i.e., El Niño impact was weakening while La Niña impact was intensifying and hence resulting in more land Pr than normal, which was proved by the increasing trend of ensemble average Pr and a significant correlation between MEI and land Pr with R2 = 0.51 (p < 0.01) shown in Figure 11a. As Pr acts as a controlling factor of the trend of E, thus it can be concluded that ENSO favored the increasing trend of global land E during the last 30 years, which has seldom been mentioned in previous studies.

Figure 11.

(a) Interannual variation of MEI multiplied by −1, SOI multiplied by 0.1, and global land Pr and scatterplot of anomalies of ensemble average (b) E versus −MEI and (c) E versus SOI × 0.1.

[51] In addition, interannual variation of ensemble average Pr agreed to common knowledge that the global land annual Pr decreases significantly in El Niño years but increases evidently when La Niña events occur [Gong and Wang, 1999; Mason and Goddard, 2001; Gu et al., 2007]. However, ENSO indices did not agree well with Pr for the strongest ENSO event during 1997–2001 (Figure 11a), which may be due to the ENSO event itself undergoing long-period variations [Gu et al., 2007; Vimont et al., 2003]. Besides, volcanic eruptions such as the Mount Pinatubo in 1991 also caused an obvious drop in land precipitation accompanying a widespread drought [Trenberth, 2011; Gu and Adler, 2011].

[52] In addition, previous studies have illustrated the regional precipitation such as the Amazonia in South America is closely associated with the cycle of El Niño and La Niña. Long-term historical climate records of Amazonia in South America shows that the “average El Niño” is drier and warmer than normal in Amazonia, while the “average La Niña” is wetter and cooler [Foley et al., 2002]. Similarly, Fu et al. [2007] reported that the average annual precipitation is 494.8 mm in La Niña years and only 408.8 mm in El Niño years with a difference of 18.8% over the long-term average in the Yellow River basin of China.

[53] However, whether ENSO affects global land E is still not clear. We found that significant correlations existed between MEI and global land E excluding two outliers in 1997 and 1998 (Figure 11b) and between SOI and E (Figure 11c). We can get to a conclusion that ENSO could be a controlling factor of the interannual variability of E.

[54] We found that global land Ta and Pr all tended to increase since 1982 (Figures 7c and 8a), which seems to contradict the general concept of more Pr causing the decrease of Ta. In fact, Trenberth and Fasullo [2009] attribute the global warming to increasing absorbed solar radiation resulting from increasing greenhouse gases and water vapor that offsets, to a large degree, the increasing radiative emissions from global warming. Especially, water vapor in the air as the dominant greenhouse gas roughly doubles the Ta change due to its positive feedback effect [Randall et al., 2007]. Thus, global warming does not necessarily correspond with a decline of land Pr. However, Dai et al. [2004] and Trenberth [2011] point out that higher temperatures have globally increased potential evapotranspiration and hence contribute to greater evaporation assuming no water stress. Similarly, increasing trend of PT Ep was found in this study mainly arising from higher temperatures (Figure 7).

[55] In addition, we found that there was a big change for MEI, Pr, E0, and E from the first period of 1982–1997 to the second period of 1998–2011; Pr, E0, and E all had a higher average value in the second period than that in the first period, which coincided with a negative average MEI = −0.05 in the second period of 1998-2011 featuring La Niña effect and a positive average MEI = 0.55 in the first period mainly suffering El Niño impact, respectively (Figure 11a). Similar studies show that 1997–1999 ENSO cycle was unique because during the transition from the warm 1997/1998 El Niño phase into the cold 1998/1999 La Niña phase, corresponding precipitation patterns were simultaneously strong and the 1997/1998 El Niño was the strongest event over last 20 years before 1999 [Gong and Wang, 1999; Curtis and Adler, 2000; Curtis et al., 2001]. We found that in the transition year 1998 occurred the obvious positive anomaly of land E = 12.97 mm yr−1. Recent 2009–2010 ENSO cycle also demonstrated a transition from the 2009–2010 El Niño phase to the 2010–2011 La Niña phase that resulted in the highest positive anomaly E = 13.67 mm yr−1of 2010 during the last 30 years. It is our assessment that abnormal higher land E probably occurred during an obvious transition from El Niño phase to La Niña phase resulting in higher E0 and Pr.

[56] According to the threshold of SOI = ±0.8 representing ENSO episode [Nicholls, 1988] and SOI time series over past 30 years (Figure 11a), typical El Niño years (i.e., 1987, 1992, 1994, and 2002) and La Niña years (i.e., 1999, 2000, 2010, and 2011) were selected. Further calculation of average anomalies of E, Pr, Ep and E0 indicates a distinctive spatial pattern in El Niño and La Niña years, respectively. ENSO-induced land Pr anomaly (Figures 12c and 12d) had a spatial pattern consistent with previous results [Dai and Wigley, 2000; Curtis and Adler, 2003]. El Niño years (Figures 12a, 12c, 12e, and 12g) featured positive anomalies of E in southern South America, Mexico, and western Asia because of more Pr while negative anomalies of E were found in Australia and southern Africa due to decreased Pr and Ep. These areas mainly suffered water supply limitation of Pr (Figure 12c). Although tropical Amazon and Asia islands experienced a decreased Pr in El Niño years, they still had a positive anomaly of E because El Niño-induced increase of Ep and E0 satisfied its needs of heating resource. Conversely, tropical Africa rainforest had a decreasing E due to reduced Ep and E0 (Figures 12e and 12g).

Figure 12.

Pattern of (a, b) E, (c, d) Pr, (e, f) PT Ep, and (g, h) E0 annual anomalies during El Niño and La Niña events.

[57] However, La Niña years (Figures 12b, 12d, 12f, and 12h) indicate an almost reversed spatial pattern of E, Pr, Ep, and E0 compared with that in El Niño years. For instance, Amazon experienced a negative anomaly of E resulting from a decreased Ep and E0 while tropical Africa rainforest had an increasing E due to an enhanced Ep and E0 (Figures 12f and 12h). Australia and southern Africa had positive anomalies of E mainly due to plentiful Pr plus increased Ep and E0. In all, the response of E to ENSO events was essentially determined by whether its major limitation factor was Ep demand or Pr supply (Figures 6e and 6f). Besides, soil moisture memory affects E in some water-stressed areas as soil moisture memory can last up to a short period of 90 days [Dirmeyer et al., 2009].

[58] Jung et al. [2010] initially reported the recent decline in the global land E trend from 1988 to 2008 due to limited soil moisture supply. However, as Pr has a strong correlation (99% significance) with soil moisture globally [Dirmeyer et al., 2009], we focused on the impact of precipitation on the decline instead of soil moisture. Our modeled results first proved the decline in ensemble average E and Pr trend over 1988 to 2008, but we found the decline was not only due to limited water supply of Pr but also due to decreased E0 (Figure 4). In fact, E0 reached its summit in 1998 corresponding with strong El Niño while Pr reached its summit in 1999/2000 associated with strong La Niña; then both E0 and Pr kept decreasing until 2009, which jointly produced the decline of E (Figure 4). Note that E0 represents E assuming no water stress under current vegetation and atmosphere conditions.

[59] In addition, the decline was fundamentally due to natural climate variability of ENSO (Figure 11a); The longest 1999–2000 La Niña event since 1982 gave a higher positive anomaly of Pr = 26.1 mm yr−1 and the following La Niña event that occurred in 2007–2009 brought a lower positive anomaly of Pr = 10.2 mm yr−1. Between these two La Niña events were several El Niño events featuring weak land precipitation, i.e., ENSO-induced decreasing of Pr resulted in the limited water supply and hence the decline of E from 1988 to 2008. Thus, ENSO was the fundamental reason for the decline of E while Pr and soil moisture featuring limited water supply were the direct reason, which answers the concern of Jung et al. [2010] whether the decline of E is representative of natural climate variability.

[60] Furthermore, Jung et al. [2010] argued whether the decline of E is permanent indicating reorganization of the land water cycle. We found that decline of E was temporary because E continued to increase to another summit in 2010 (Figure 1a) and MEI, Pr, E0, and E all featured an enhanced land water cycle during the period of 1998–2011 compared with the first period of 1982–1997. Variation of Pr indicates wetter climate in 2010 and 2011 (Figure 4) consistent with climate analysis based on GHCN and GPCC precipitation record [Blunden and Arndt, 2012].

[61] Fundamentally, ENSO is the largest signal in the interannual variation of ocean-atmosphere system [Wang et al., 1999]. It originates from the tropical Pacific but affects the global climate through teleconnection effect via changes in the circulation patterns. Numerous studies have identified ENSO-induced climate variability of precipitation, temperature, reference evapotranspiration, water balance, and drought [Yang and DelSole, 2011; Ropelewski and Halpert, 1986; Meza, 2005; Twine et al., 2005; Coelho and Goddard, 2009]. Besides, we found that ENSO impacted the interannual variation of global land E through the similar mechanism of teleconnection.

[62] Currently, large difference still exists among global estimations of E given by different studies [Jung et al., 2010; Trenberth et al., 2007; Vinukollu et al., 2011; Miralles et al., 2011a] because of uncertain forcing and model mechanism [Schlosser and Gao, 2010; Yan et al., 2012]. As the model range of global E estimates is larger than any bias caused by uncertainties in the model forcings [Vinukollu et al., 2011; Schlosser and Gao, 2010], model mechanism should be improved with first priority. For example, the E models, due to using air humidity as a surrogate to soil moisture, may not reflect the decline in E in the 2000s due to soil moisture limitation that mainly occurred in the Southern Hemisphere [Vinukollu et al., 2011]. ARTS E model as a revised Penman-Monteith model [Monteith, 1965] explicitly considers canopy conductance derived from remote sensing Lai, energy balance and water balance [Yan et al., 2012]. Although ARTS E model considers the snow melting effect, snow sublimation effect will be studied in our future work because snow sublimation, dominating E in high latitudes in winter and in the mountainous regions of midlatitudes, occupies 2% of global land E [Miralles et al., 2011b].

[63] However, Jiménez et al. [2011] argued that whether model forcing adds less uncertainty than E different parameterizations depends on what processes are analyzed and over what regions. Thus, there are currently international efforts trying to get a better understanding of the whole E estimation, such as the Global Energy and Water Cycle Experiment coordinated evaluation of E estimates by the LandFlux-EVAL initiative [Jiménez et al., 2011; Mueller et al., 2011].

[64] As water balance was actually driven by Pr data in ARTS E model and different Pr products had some differences (Figure 8a) and even different spatial pattern of trends (not shown) due to different numbers of observing stations [Blunden and Arndt, 2012], different interpolation method [Willmott and Robeson, 1995], and merging of satellite precipitation production in GPCP [Huffman et al., 2009], we adopted the ensemble method to obtain ensemble average Pr and E to reduce the impact of input error of Pr. The ensemble average Pr had an increasing trend consistent with precipitation climate analysis by Blunden and Arndt [2012].

[65] Similarly, large differences found in radiation dada affect global estimation of E [Vinukollu et al., 2011; Yan et al., 2012]. Two satellite radiation products, i.e., International Satellite Cloud Climatology Project and Surface Radiation Budget (SRB) radiation products, show temporal inconsistencies due to changes in satellite sensors and retrieval algorithms [Vinukollu et al., 2011]. In addition, the negative bias of SRB net radiation partly resulted in a lower global E estimate of 58.4 × 103 km3 yr−1 [Yan et al., 2012]. However, as interannual variation of global E was more sensitive to Pr and Lai than Ep perturbations already reflecting the impact of net radiation, uncertainties in radiation data of MERRA reanalysis cannot substantially affect the main conclusions of this study.

[66] To reduce uncertainty in E estimates, improving model and forcing continues to be essential [Mueller et al., 2011; Vinukollu et al., 2011; Yan et al., 2012]. For example, GLEAM E model [Miralles et al., 2011b] considers more processes including the rainfall interception and snow sublimation and incorporates Advanced Microwave Scanning Radiometer–EOS microwave-derived land surface temperature, vegetation optical depth to capture vegetation phenology [Jones et al., 2011], and soil moisture which may result in a better representation of the water supply process as well as evaporation stress.

[67] However, it is our assessment that to build ensemble average E, based on ensemble of forcing data and E models, was an available method to obtain a reasonable estimation of global land E with deviation statistics. This study preliminarily built an ensemble of six precipitation data sets and then analyzed an ensemble of six independent E products estimated from one single ARTS E model over the period of 1982 to 2011. Further ensemble analysis of more E models and forcings including radiation will be conducted in the next step of work.


[68] This work was supported by National Natural Science Foundation of China (41171284, 40801129), Chinese Academy of Sciences (XDA05050602-1), and China Scholarship Council Foundation and partly funded by NASA Earth Science Division, as well as by the following NASA grants to H.H. Shugart: 10-CARBON10-0068 and Climate Change/09-IDS09-116. The reviewers are thanked for their constructive remarks and suggestions.