A Global Intercomparison of Modeled and Observed Land–Atmosphere Coupling

Craig R. Ferguson Department of Hydrology and Water Resources Engineering, Institute of Industrial Science, University of Tokyo, Tokyo, Japan

Search for other papers by Craig R. Ferguson in
Current site
Google Scholar
PubMed
Close
,
Eric F. Wood Department of Civil and Environmental Engineering, Princeton University, Princeton, New Jersey

Search for other papers by Eric F. Wood in
Current site
Google Scholar
PubMed
Close
, and
Raghuveer K. Vinukollu Swiss Re, Armonk, New York

Search for other papers by Raghuveer K. Vinukollu in
Current site
Google Scholar
PubMed
Close
Full access

Abstract

Land–atmosphere coupling strength or the degree to which land surface anomalies influence boundary layer development—and in extreme cases, rainfall—is arguably the single most fundamental criterion for evaluating hydrological model performance. The Global Land–Atmosphere Coupling Experiment (GLACE) showed that strength of coupling and its representation can affect a model’s ability to simulate climate predictability at the seasonal time scale. And yet, the lack of sufficient observations of coupling at appropriate temporal and spatial scales has made achieving “true” coupling in models an elusive goal. This study uses Advanced Microwave Scanning Radiometer for Earth Observing System (AMSR-E) soil moisture (SM), multisensor remote sensing (RS) evaporative fraction (EF), and Atmospheric Infrared Sounder (AIRS) lifting condensation level (LCL) to evaluate the realism of coupling in the Global Land Data Assimilation System (GLDAS) suite of land surface models (LSMs), Princeton Global Forcing Variable Infiltration Capacity model (PGF–VIC), seven global reanalyses, and the North American Regional Reanalysis (NARR) over a 5-yr period (2003–07). First, RS and modeled estimates of SM, EF, and LCL are intercompared. Then, emphasis is placed on quantifying RS and modeled differences in convective-season daily correlations between SM–LCL, SM–EF, and EF–LCL for global, regional, and conditional samples. RS is found to yield a substantially weaker state of coupling than model products. However, the rank order of basins by coupling strength calculated from RS and models do roughly agree. Using a mixture of satellite and modeled variables, a map of hybrid coupling strength was produced, which supports the findings of GLACE that transitional zones tend to have the strongest coupling.

Supplemental information related to this paper is available at the Journals Online website: http://dx.doi.org/10.1175/JHM-D-11-0119.s1.

Corresponding author address: Craig R. Ferguson, Ce407, Institute of Industrial Science, University of Tokyo, 4-6-1 Komaba, Meguro-ku, Tokyo 153-8505, Japan. E-mail: cferguso@rainbow.iis.u-tokyo.ac.jp

This article is included in the Exchanges of Energy and Water at the Land-Atmosphere Interface special collection.

Abstract

Land–atmosphere coupling strength or the degree to which land surface anomalies influence boundary layer development—and in extreme cases, rainfall—is arguably the single most fundamental criterion for evaluating hydrological model performance. The Global Land–Atmosphere Coupling Experiment (GLACE) showed that strength of coupling and its representation can affect a model’s ability to simulate climate predictability at the seasonal time scale. And yet, the lack of sufficient observations of coupling at appropriate temporal and spatial scales has made achieving “true” coupling in models an elusive goal. This study uses Advanced Microwave Scanning Radiometer for Earth Observing System (AMSR-E) soil moisture (SM), multisensor remote sensing (RS) evaporative fraction (EF), and Atmospheric Infrared Sounder (AIRS) lifting condensation level (LCL) to evaluate the realism of coupling in the Global Land Data Assimilation System (GLDAS) suite of land surface models (LSMs), Princeton Global Forcing Variable Infiltration Capacity model (PGF–VIC), seven global reanalyses, and the North American Regional Reanalysis (NARR) over a 5-yr period (2003–07). First, RS and modeled estimates of SM, EF, and LCL are intercompared. Then, emphasis is placed on quantifying RS and modeled differences in convective-season daily correlations between SM–LCL, SM–EF, and EF–LCL for global, regional, and conditional samples. RS is found to yield a substantially weaker state of coupling than model products. However, the rank order of basins by coupling strength calculated from RS and models do roughly agree. Using a mixture of satellite and modeled variables, a map of hybrid coupling strength was produced, which supports the findings of GLACE that transitional zones tend to have the strongest coupling.

Supplemental information related to this paper is available at the Journals Online website: http://dx.doi.org/10.1175/JHM-D-11-0119.s1.

Corresponding author address: Craig R. Ferguson, Ce407, Institute of Industrial Science, University of Tokyo, 4-6-1 Komaba, Meguro-ku, Tokyo 153-8505, Japan. E-mail: cferguso@rainbow.iis.u-tokyo.ac.jp

This article is included in the Exchanges of Energy and Water at the Land-Atmosphere Interface special collection.

1. Introduction

Land surface–atmosphere interaction (henceforth, coupling), or degree to which anomalies in the land surface state (i.e., soil wetness, soil texture, surface roughness, temperature, and overlying vegetation composition and structure) can affect (through complex controls on the partitioning of surface turbulent fluxes) the planetary boundary layer (PBL) and, in extreme cases, rainfall generation, is an important—if not the single most fundamental—criterion for evaluating hydrologic and atmospheric model performance (Betts 2004, 2009). By linking the surface, PBL, and cloud processes, coupling encompasses complex cross-scale interactions that determine the climate state. Coupling strength varies on local (5–10 km) to regional scales (400 km) and temporally on daily to weekly time scales (Betts 2004; Koster et al. 2003; Taylor and Ellis 2006), modulated by background synoptic weather (i.e., convergence/divergence, monsoons, and cloud fields) and larger scale (i.e., ocean–atmosphere) dynamics. Conceptually, coupling implies that in certain regions the land surface (by inducing persistence or climate “memory”) may act to reinforce meteorological extremes (i.e., flood and drought). Indeed, coupling is believed to have played an important role in the European heat waves (Fischer et al. 2007); projected changes in temperature variability in Europe are largely due to enhanced coupling as well (Seneviratne et al. 2006). It follows that the strength and/or variability of coupling may play a large role in the predictability of regional climate (i.e., temperature and precipitation). For example, potential rainfall forecast skill from land may be high where coupling is strong and invariable and low where coupling is either consistently weak or strong but highly variable.

In the past, there have been many attempts to evaluate coupling with models (e.g., Beljaars et al. 1996; Betts 2004; Cook et al. 2006; Delworth and Manabe 1989; Dirmeyer 2011; Dirmeyer et al. 2009; Findell et al. 2011; Guo et al. 2006; Koster et al. 2004, 2006; Wei et al. 2010a,b). The first phase of the Global Energy and Water Cycle Experiment (GEWEX) Global Land–Atmosphere Coupling Experiment (GLACE-1; Guo et al. 2006; Koster et al. 2006) revealed large disparities in the strength of coupling inherent to each of the 12 participating atmospheric general circulation models (AGCMs; Dirmeyer et al. 2006b). Despite the markedly different features of the models, some geographical hot spots of coupling emerged where consensus among the participating models was high. Unfortunately, because of a lack of relevant globally available on-the-ground observational data at sufficient spatial and temporal scales, GLACE was unable to address the realism of these modeled hot spots and, moreover, to provide any guidance for future model enhancement (Koster et al. 2006). The interpretation and application of other modeling results have been similarly impaired by the absence of observational data.

Dirmeyer et al. (2006b) used nine Energy Balance Bowen Ratio (Cook 2005) stations and six micrometeorological flux towers (Baldocchi et al. 2001) to evaluate the GLACE coupling strength over a limited temporal and spatial domain. For these limited sites, they were able to effectively diagnose on a model-by-model basis the realism of coupling represented. Nevertheless, the global availability and distribution of long-term collocated in situ measurements required for such intercomparisons is likely to remain scarce. The fact that coupling strength and signal may vary across scales (Hohenegger et al. 2009) implies that flux tower–based estimates of coupling, generally representative of an area no greater than 1–2 km2, no matter the density of the network, will be nonrepresentative and insufficient. Clearly, there is a mounting need to bring emerging satellite datasets to bear on this issue.

One particularly relevant satellite platform for coupling studies is the National Aeronautics and Space Administration (NASA) Earth Observing System (EOS) Aqua launched in May 2002. Aqua carries on board a suite of six sensors, including Atmospheric Infrared Sounder (AIRS), Advanced Microwave Sounding Unit (AMSU), Advanced Microwave Scanning Radiometer for EOS (AMSR-E), Clouds and the Earth’s Radiant Energy System (CERES), Moderate Resolution Imaging Spectroradiometer (MODIS), and the Humidity Sounder for Brazil (HSB; failed February 2003), which provide complimentary, coregistered (in time and space) information on the state of the land surface (e.g., soil moisture, soil temperature, vegetation fraction, land cover type, leaf area index, and albedo), atmospheric profile (e.g., temperature, humidity, and precipitation), and surface radiation forcing, at spatial resolutions ranging from 250 m (MODIS) to 50 km (AIRS/AMSU). Aqua is set in a sun-synchronous orbit with equatorial crossing times of ~1330 and ~0130 local time (LT) on ascending and descending nodes, respectively, with a global repeat on the order of 2–3 days. Accordingly, Aqua’s overpass times are well suited for coupling studies: the descending node provides a sampling of the stable, nocturnal boundary layer conditions, while the ascending node captures the near peak of the daytime PBL growth.

Reanalyses, which are loosely constrained by the observations they assimilate, represent a secondary means to investigate coupling. Reanalyses are particularly attractive as compared to satellite remote sensing because they provide long-term (30–60 year), multivariate, spatially and temporally consistent and complete records of the global land surface and multilevel atmospheric state. Spatial resolutions vary from 2.0° to 0.3°—roughly comparable with those of remote sensing—with the general trend being toward higher resolution in modern reanalyses. Temporal resolutions vary from 3 to 6 hourly, with some reanalyses providing 1-hourly updates of surface and near-surface variables. Because of their relative ease of use (particularly relative to satellite data) and widespread applicability, reanalyses are increasingly being used in place of observations (e.g., as forcing for land surface models), as well as in merged/assimilated products to quantify the global water and energy cycle (Bosilovich et al. 2011; Pan et al. 2012; Sahoo et al. 2011; Trenberth et al. 2007; Troy et al. 2011; Vinukollu et al. 2012). The accuracy of any given reanalysis’s representation of coupling is complexly dependent on four components: 1) the quality and coverage (both spatial and temporal) of the assimilated observations; 2) the assimilation system itself, including the forward radiative transfer model for direct assimilation of raw satellite radiances (if applicable); 3) the background-assimilating AGCM; and 4) the coupled land surface model. Modeled precipitation (i.e., bias and diurnal cycle error) and related radiation/cloud forcing deficiencies serve as limiting factors in the accuracy of many reanalyses (Dai 2006; Reichle et al. 2011; Ruane 2010a,b). A well-known problem with reanalyses is that the water and energy budgets are not closed (conserved) by construct (Saha et al. 2010; Trenberth and Smith 2005). In general, reanalyses overestimate both surface evaporation and precipitation, leading to accelerated global hydrologic cycling (Trenberth et al. 2009, 2011). Trenberth et al. (2011) recently assessed the global hydrological cycles from eight reanalyses. Based on their findings, they recommend users exercise “great caution” when using surface fluxes. An intercomparison of the coupling strength inherent to each of the reanalyses has yet to be carried out. Whether a reanalysis can be quantitatively wrong in capturing the mean magnitude of water and energy budget terms and still correct in terms of coupling strength also remains an open question.

The objectives of this study are twofold: first, to quantify the correspondence between satellite remote sensing (RS)-based and offline modeled or reanalysis output for three key terms in our diagnosis of coupling: surface soil moisture (SM), evaporative fraction (EF), and lifting condensation level (LCL); secondly, to produce and intercompare global maps of coupling strength derived from each of the datasets. Specifically, we use global maps of RS coupling strength (surrogate truth) to assess the capability of five offline land surface models (LSMs) and all (eight) available long-term reanalysis datasets (see Table 1) to represent real-world coupling. The convective-season SM–LCL nonparametric correlation serves as our measure for coupling strength. Over land, the LCL has been shown to provide a good approximation of the mean cloud base height in addition to a good indication of the atmospheric demand for water/likelihood of precipitation (Betts 2009). Because the SM–LCL process chain comprises two critical intermediary links—soil moisture controls on surface turbulent flux partitioning (i.e., the land segment) and the impact of surface fluxes on the overlying boundary layer (clouds) (i.e., the atmospheric segment)—we compute SM–EF and EF–LCL correlations as well. RS estimates of SM, EF, and LCL are derived from the AMSR-E, NASA–GEWEX Surface Radiation Budget (SRB; Stackhouse et al. 2000) and International Satellite Cloud Climatology Project (ISCCP; Schiffer and Rossow 1983), and AIRS sensors, respectively. Included among the offline models are the four Global Land Data Assimilation System (GLDAS; Rodell et al. 2004) LSMs: Mosaic (Koster and Suarez 1996), National Centers for Environmental Prediction (NCEP)–Oregon State University (OSU)–Air Force–Hydrologic Research Laboratory Model (Noah; Chen et al. 1996; Ek et al. 2003), Community Land Model (CLM; Bonan et al. 2002; Dai et al. 2003), and the Variable Infiltration Capacity model (VIC; Liang et al. 1996, 1994), as well as the Princeton Global Forcing–VIC (PGF–VIC; Sheffield and Wood 2007). The GLDAS models serve to isolate the impact of LSM parameterizations on coupling strength because of the fact that they share the same meteorological forcing. The PGF is considered the most statistically representative forcing dataset and will soon serve as the primary forcing in the GLDAS. The PGF–VIC model was calibrated to best match observed discharge observations at twenty large-scale hydrologic basins distributed globally. Therefore, it provides arguably our best estimate of modeled coupling. The intercomparisons are conducted on a global basis first, then conditioned on key hydroclimatic and vegetation morphological parameters, and finally by basin or GLACE hot spot. Hybrid, multimodel, and observation-model pairings of variable fields are used to establish consensus on coupling strength.

Table 1.

Summary of the LSMs and reanalyses used in this study.

Table 1.

The layout of the paper is organized such that the differences between modeled and RS estimates of land–atmosphere coupling are communicated in the most direct way possible. Accordingly, many details and results are deferred to a series of appendices (A–D) and supplemental tables and figures (S1–S9; supplemental material is available at http://dx.doi.org/10.1175/JHM-D-11-0119.s1). Section 2 summarizes the RS datasets central to our study: SM, EF, and LCL, as well as the hydroclimate and vegetation morphological parameter datasets used to conduct conditional probability analyses: mean 95% ecosystem rooting depth, leaf area index, wetness index (after Budyko 1958, 1974), and land cover type. The results, summary, and conclusions are included in sections 3 and 4.

2. Data and methods

a. Primary remote sensing data

1) AMSR-E soil moisture

AMSR-E makes dual-polarized passive radiation measurements at six frequencies (6.9, 10.7, 18.7, 23.8, 36.5, and 89.0 GHz). Each of the antenna beams scan conically about the nadir axis across a 1445-km swath, providing a near-global coverage in 2 days or less (Njoku et al. 2003). Satellite microwave remote sensing retrievals of surface (0–0.02 m) soil moisture derived from AMSR-E C (6.925 GHz) and X-band (10.65 GHz) observed brightness temperatures are available at global, twice daily (night/day) temporal resolution, and 0.25° spatial resolution from the Vrije Universeiteit Amsterdam, the Netherlands (http://geoservices.falw.vu.nl/adaguc_portal_dev/). These retrievals are the output of the Land Surface Parameter Model (LPRM; Owe et al. 2008)—a dual polarization microwave radiative transfer scheme based on the theoretical relationship between the microwave polarization difference index (MPDI), vegetation optical depth, and the soil dielectric constant (Owe et al. 2001). We use only the nighttime (descending overpass) retrievals because we are interested in assessing the influence of initial soil moisture conditions on the development of the daytime boundary layer. Nighttime retrievals successfully capture soil moisture anomalies caused by late-afternoon convective precipitation, which in water-limited regimes can be substantially eroded via evaporation by the time of afternoon overpass. During nighttime, the LPRM’s assumption of thermal equilibrium between the vegetation canopy and soil surface is also most accurate.

The sensitivity of microwave soil moisture retrievals diminishes with increasing vegetation density and microwave frequency (Vinnikov et al. 1996). The overlying vegetation canopy can significantly attenuate or mask emission from the underlying soil, resulting in soil moisture retrievals that are inconsistent with observed precipitation dynamics (Gao et al. 2006). For this reason, the LPRM uses C band as the default retrieval frequency, except in regions of identified radio frequency interference (Li et al. 2004), in which case C band is replaced with X band. Previous studies have suggested that the vegetation water content threshold for soil moisture retrievability at X and L bands (1.4 GHz) is approximately 1.5–2.0 kg m−2 and 4–5 kg m−2 (Kerr 2007; Narayan et al. 2004; Njoku et al. 2003), respectively. This implies for the conterminous United States that only retrievals west of the 97°W meridian are reliable for X band, with skill extending over the Mississippi River valley and Midwest corn belt at L band. The feasibility of C-band retrievals should fall somewhere in between. The retrieval accuracy is also sensitive to other land cover properties, including surface roughness, topographic complexity, surface water fraction, ice, and snow (Gao et al. 2006). Therefore, regions with high (>5%) surface water fraction, frozen (T < 273 K) or snow-covered ground, or dense vegetation (vegetation optical thickness > 0.8) are all assigned missing values in the LPRM. For nonmissing values, the LPRM provides a quality flag. Only estimates of “nominal” or higher accuracy are used. For additional quality control, we apply the global porosity map used in the LPRM dielectric mixing module to screen out any pixel that does not exhibit drydown (i.e., saturated state for 5 or more of the previous 7 days). Previously, the LPRM SM was shown to compare well with in situ measurements at 5-cm depth taken from a semiarid region of Spain (Wagner et al. 2007).

For this study, we regridded the quality-cleared 0.25° Level 3 LPRM product to 1.0° by box averaging.

2) Multisensor evaporative fraction

We calculate EF at 3-hourly temporal and 0.5° spatial resolution using latent heat flux (LE) estimates from a modified Penman–Monteith approach (PM; Mu et al. 2007; Vinukollu et al. 2011b), SRB net radiation (Rn), and SRB ground heat flux (G):
e1
where H is the sensible heat flux and β is the commonly applied Bowen ratio, equivalent to
e2
(Bowen 1926). EF is preferable to β because β is unbounded and thus ill-suited for use in arid regions where LE approaches zero (and β approaches infinity).
The latent heat flux LE (W m−2) is calculated from
e3
where Δ is the slope of the curve, relating saturation vapor pressure to temperature (Pa K−1); ρa (kg m−3) is the air density; cp (1005 J kg−1 K−1) is the specific heat capacity of dry air; VPD = (esate) is the vapor pressure deficit, where esat and e (Pa) are the saturation vapor pressure and vapor pressure, respectively; γ is the psychrometric constant (Pa K−1); and ra and rs (s m−1) are the aerodynamic and effective surface resistance, respectively. In calculating rs, we implemented environmental stress functions for VPD and minimum daily temperature, as described in Mu et al. (2007), but importantly, prescribe biome-specific mean potential (i.e., maximum) stomatal conductance (taken from VIC), as opposed to the fixed value of 3.3 mm s−1 used in Mu et al. (2007). The VPD-based stress function simulates stomatal response to water stress, and hence, to the extent that VPD corresponds with SM, SM and LE (EF) are dependent. The aerodynamic resistance formulation was taken from the Surface Energy Balance System (SEBS; Su 2002), based on Monin–Obukhov similarity theory (Monin and Obukhov 1954) and the work of Massman (1997). SRB net radiation Rn is computed as the sum of incident shortwave (SWin) and longwave (LWin) fluxes less the sum of outgoing shortwave (SWout) and longwave (LWout) fluxes:
e4
where SWin is obtained from SRB release 3.0, and SWout is calculated by
e5
where α is the release 3.0 quality-checked (QC) daily mean surface albedo, and LWin and LWout are also from release 3.0 QC. A QC version of SWin is not currently available at 3-hourly time step (daily or monthly only). Ground heat flux G is computed as a function of the daily mean Television Infrared Observation Satellite (TIROS) Operational Vertical Sounder (TOVS) surface skin temperature following the method of Bennett et al. (2008). The SRB radiation products above (all at 1.0° spatial resolution) serve as inputs to the latent heat flux estimation, along with a suite of meteorological and surface vegetation datasets, including TOVS-derived surface pressure, surface skin temperature, surface air temperature, and surface specific humidity (Zhang et al. 2004); PGF (Sheffield et al. 2006) surface u- and υ-wind components; Advanced Very High Resolution Radiometer (AVHRR)-derived leaf area index [LAI; Ganguly et al. 2008b; see also section 2b(2)] and fractional vegetation cover (fc); and MODIS land cover class [lc; see section 2b(3)]. The TOVS meteorological inputs are available from the ISCCP flux data–surface radiative fluxes (ISCCP–FD–SRF; Zhang et al. 2004) dataset at 2.5° spatial resolution and 3-hourly temporal resolution. We approximate the surface specific humidity from the TOVS precipitable water vapor estimate for the 200-hPa layer covering the surface by assuming a constant relative humidity (RH) profile and lapsing to the surface temperature and pressure. Over land, RS estimates of wind currently do not exist, which is why we use reanalysis-based inputs from the PGF. The fc is derived from AVHRR normalized difference vegetation index (NDVI; Pinzon et al. 2005; Tucker et al. 2005) following the method of Zeng et al. (2000). We use 15-day LAI and fc, but the land cover type was fixed to that of the 2004 classification. Sensitivity of the RS LE to the choice of input datasets and surface parameterizations over the continental United States (CONUS) is discussed in detail by Ferguson et al. (2010).

Separate modules calculate the free water evaporation from inland water bodies and canopy-intercepted precipitation, which augments the PM time step total LE. Surface water evaporation is computed at potential rate (rs = 0) using the Penman equation (Penman 1948). Canopy evaporation is computed using a simple mass-balance interception model (Valente et al. 1997; Vinukollu et al. 2011b)—the storage capacity of which is a function of vegetated land cover fraction (fc) and LAI—forced with 3-hourly precipitation uniformly disaggregated from the National Oceanic and Atmospheric Administration/Climate Prediction Center (NOAA/CPC) unified daily gauge analysis (CPC-UNI version 1.0; Chen et al. 2008; Xie et al. 2007). All inputs, except the land cover, were regridded to 0.5° spatial resolution by box averaging. Evapotranspiration estimates were then estimated using a weighted average of total evapotranspiration (transpiration from dry canopies and direct evaporation from open water sources and canopy-intercepted water) over each land cover type (tile) within a 0.5° grid. Henceforth, we shall refer to the satellite-based EF by its atmospheric input sources SRB/ISCCP, or as simply RS–EF.

Despite using ISCCP meteorological forcings, we chose not to use ISCCP Rn. Recently, a global intercomparison of the SRB and ISCCP surface radiation products over a 24-yr (1984–2007) period was carried out by Vinukollu et al. (2011a). In that study, the authors identified large mean value differences for each of the components of Rn as well as temporal inhomogeneities (e.g., steps) in the time series due in part to changes in satellite sensors and retrieval algorithms (i.e., TOVS, 1998–2001). A review of their results (Vinukollu et al. 2011a, their Fig. 6) suggests that SRB, and particularly the combination of SRB radiation components/quality control used herein (see above), yields the most internally consistent (in terms of its components) estimate of Rn.

3) AIRS lifting condensation level

The AIRS provides high-accuracy retrievals of atmospheric temperature and humidity profiles, ozone profiles, sea/land surface skin temperature, and cloud properties twice daily (equatorial crossing at ~1330/0130 LT on ascent/descent) for most of world with greater frequency at high latitudes (Aumann et al. 2003). The physical retrieval algorithm of AIRS is unique in that it corrects for the impact of clouds on the infrared sounding by applying a “cloud-clearing” technique (Chahine et al. 2006). This enables successful soundings 47.5% and 72.3% of the time over land and ocean, respectively (Susskind et al. 2003). The accuracies of AIRS temperature and water vapor profiles have been shown, through intercomparisons with coregistered radiosonde observations, to be close to the goal accuracies of 1 K RMS in 1-km tropospheric layers and 20% RMS in 2-km tropospheric layers (Divakarla et al. 2006).

Here, we use spatially interpolated fields of near-surface air temperature (Ta), specific humidity (q), and surface pressure (Psurf) from the AIRS–AMSU Level 2 Support Product Dataset (AIRX2SUP) version 5 (from 1 September 2002 to 30 September 2007) and version 5.2 (from 1 October 2007 to 31 December 2009). The version 5 retrieval algorithm, which is the most current release, implements a radiative transfer algorithm that accounts for nonlocal thermodynamic equilibrium effects on the shortwave channels; incorporates a new, accurate cloud-clearing algorithm using AIRS spectra only; and provides for the first time case-by-case product error estimates (Susskind et al. 2010). AIRS footprint values of 13.5-km (at nadir) resolution are first quality screened and then spatially interpolated to 0.125° geographic coordinates using an inverse-distance-squared interpolation scheme with a great-circle search radius of 75 km. In the case of Ta, footprint values are elevation corrected to match the U.S. Geological Survey Global 30 Arc Second Elevation Dataset (GTOPO30) using a mean environmental lapse rate of −6.5 K km−1 prior to interpolation. The footprint Psurf estimates provided in the AIRX2SUP dataset have been interpolated in space and time from the NCEP operational Global Forecast System (GFS; Saha et al. 2006), which runs operationally at T254 horizontal resolution and 3-hourly time step. For the sake of consistency, Psurf is only included if corresponding footprint values of Ta and q satisfy quality requirements. After regridding, the 0.125° fields of Ta, q, and Psurf were used to compute the LCL following Eqs. (6)(9):
e6
e7
e8
and
e9
where Td is the dewpoint temperature in degrees Celsius, e is in Pascals, cp is in units of J kg−1 K−1, R is the specific gas constant for dry air (287.05 J kg−1 K−1), and Psurf and q are in units of Pascals and kg kg−1, respectively. There are numerous other approximations of LCL, computed from atmospheric slabs of varying thickness and origin. LCL is only calculable from LSMs (more specifically, their atmospheric forcing) by this formulation, which is the principal motivation for its use. The underlying assumption, of course, is that the near-surface layer is sufficiently mixed such that 2-m RH estimates are representative of the overlying 50–100-hPa layer. Intercomparisons between AIRS LCL and spatially interpolated fields of LCL derived from multiple surface meteorological sites within the Oklahoma Mesonet yielded an RMS error of 39.7 hPa during May–September (Ferguson and Wood 2009, their Fig. 1).

Full details regarding the data processing are provided in Ferguson and Wood (2010). In that study, daytime (~1330 LT) footprint retrievals, as well as the 0.125° gridded fields, were extensively evaluated over CONUS and Africa for a period of 6 years (2002–08) using matchups at over 2000 surface meteorological stations and additionally over CONUS using the North American Land Data Assimilation System (NLDAS; Mitchell et al. 2004) model-based forcing fields. The retrievals were carefully analyzed for sensitivity across a range of surface and atmospheric conditions including season, cloud albedo, total column precipitable water vapor/ice, elevation, topographic complexity, lc, and vegetative canopy thickness (LAI). Retrieval accuracy was shown to be highest under clear-sky conditions (cloud fraction ≤ 0.2), with mean Kendall’s nonparametric correlation with NLDAS of 0.84 and 0.70 for Ta and q, respectively. Bias in Ta (warm) and q (moist) were found to correspond with elevation: the potential result of integrated effects of changes in several variables—including lc, LAI, and humidity—that vary with altitude. The correlation of q degrades with elevation, but not Ta, although this may be the artifact of station siting (Davey and Pielke 2005) and/or the data interpolation scheme applied.

For this study, the 0.125° AIRS LCL fields were regridded to 1.0° by box averaging and separated into UTC hourly global files.

b. Secondary remote sensing data

1) AMSR-E convective season

The AMSR-E Level 2B version 10 swath product (AE_Rain; Adler et al. 2007) provides instantaneous estimates of rain rate and rain type (convective or stratiform) for all ice and snow-free land and ocean between 70°S and 70°N at a spatial resolution of 5.4 km. Using all available swaths (ascending and descending) from a 7-yr period (January 2003–December 2009), we tabulated on a 1.0° geographic grid the total number of events per calendar month characterized by a 45% convective rainfall fraction or greater (henceforth, a convective event). Then, for each 1.0° pixel we selected the set of top-ranking calendar months whose convective event sum did not exceed 80% of the total (7 yr) convective event count and that independently accounted for greater than 10% of the total convective event count. This method has been shown to provide a conservative (in duration) yet accurate portrayal of global convective season (Ferguson and Wood 2011). A convective season so defined is an improvement upon conventional meteorological seasons (e.g., December–February, March–May, June–August, and September–November), which often fail to correspond with true annual precipitation cycles, especially in regions with multiple or prolonged wet seasons.

2) AVHRR leaf area index

Ganguly et al. (2008a,b) produced a 25-yr (July 1981–December 2006) record of LAI derived from the Global Inventory Monitoring and Modeling Studies (GIMMS) AVHRR NDVI composites (Tucker et al. 2005). The retrieval algorithm includes monthly biome-specific, configurable parameters that were tuned to achieve maximum consistency (e.g., minimized RMS) with quality-cleared MODIS Terra version 5 (Yang et al. 2006) LAI over a 1-yr (January–December 2001) training period (Ganguly et al. 2008b). The AVHRR LAI dataset is available as 15-day maximum value composites at 8-km spatial resolution. For this study, we maintain the 15-day temporal resolution, but upscale the 8-km fields to 1.0° using box averaging. We compute a mean (AMSR-E derived) convective-season LAI for each pixel using data from the period of 2003–2006.

3) MODIS land cover classification

The MODIS Terra and Aqua yearly land cover type 0.05° Climate Modeling Grid (CMG) product (MCD12C1; Friedl et al. 2010) provides global land cover classification according to the following three classification schemes: International Geosphere–Biosphere Programme (IGBP), MODIS-derived LAI fraction of photosynthetically active radiation (FPAR) scheme, and the University of Maryland (UMD) scheme. We use the UMD classification, which is also used by GLDAS. Using an arbitrary (2003) classification year, we produced a global map of dominant land cover type at 1.0° resolution via voting of the constituent (n = 400) 0.05° pixels.

c. Model data

1) Land surface models and reanalyses

Estimates of coupling were derived from five offline (uncoupled) semidistributed LSMs, seven global reanalyses, and one regional reanalysis (Table 1). They are the GLDAS (Rodell et al. 2004) participant models: CLM, Mosaic, Noah, and VIC; PGF–VIC; NCEP–National Center for Atmospheric Research (NCEP–NCAR) reanalysis (NCEP-R1); NCEP–Department of Energy (DOE) Atmospheric Model Intercomparison Project (AMIP) Reanalysis (NCEP-R2); Japanese 25-yr Reanalysis (JRA-25); NASA Modern-Era Retrospective Analysis for Research and Applications (MERRA); European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis Interim (ERA-INT); NCEP Climate Forecast System Reanalysis (CFSR); NOAA–Cooperative Institute for Research in Environmental Sciences (CIRES)/University of Colorado Twentieth Century Reanalysis version 2 (NCEP–CIRES); and NCEP North American Regional Reanalysis (NARR). Of particular relevance to this study is to which LSM the forecast system of each reanalysis is coupled. NCEP-R1 is coupled with the OSU two-layer LSM (Pan and Mahrt 1987). NCEP-R2 is coupled with the OSU LSM as well, but the climatology-based soil moisture nudging scheme implemented in NCEP-R1 is replaced with a correction scheme based on the assimilation of 2.5° pentad (5 day) CPC Merged Analysis of Precipitation accumulations (CMAP; Xie and Arkin 1997). JRA-25 is coupled with the three-layer Japan Meteorological Agency (JMA)–Simple Biosphere LSM (SiB; Sato et al. 1989; Sellers et al. 1986). The Catchment LSM of Koster et al. (2000) is included in MERRA’s AGCM. ERA-INT is coupled with the four-layer Tiled ECMWF Scheme for Surface Exchanges over Land (TESSEL; Van den Hurk et al. 2000; Viterbo and Beljaars 1995). And NARR, CFSR, and NOAA–CIRES are coupled with the Noah LSM (Ek et al. 2003). The assimilated observational datasets are identical for NCEP-R1/R2 (with known human processing errors corrected in NCEP-R2) and are very similar for JRA-25, NARR, MERRA, ERA-INT, and CFSR. NOAA–CIRES assimilates surface observational data (surface pressure, sea surface temperature, and sea ice extent) only. NARR is the only reanalysis to assimilate observed precipitation, converted into latent heat (Lin et al. 1999; Rogers et al. 2001). Three-dimensional variational data assimilation system (3D-Var) analysis is most common among the reanalyses, although ERA-INT and NOAA–CIRES implement 4D-Var and ensemble Kalman filter analysis algorithms, respectively.

There are two striking distinctions among the reanalyses worth noting. First, at screen level, ERA-INT and JRA-25 are the only reanalyses to assimilate 2-m temperature and humidity—others assimilate surface pressure only. Both reanalyses handle the analysis of surface parameters separately from the main atmospheric analysis. Specifically, 6-hourly estimates of 2-m temperature and humidity are produced with a univariate 2D optimal interpolation scheme, combining surface observations with modeled estimates (obtained diagnostically) from the background-assimilating model (Douville et al. 1998). These near-surface fields are then used to update the soil moisture and soil temperature profile with the goal of improving turbulent surface fluxes and subsequent atmospheric forecasts (Viterbo et al. 2000). In terms of the soil moisture, Drusch and Viterbo (2007) found no clear benefit resulting from this method. Second, CFSR is the only reanalysis to use observed precipitation in its land analysis. In CFSR, Noah is implemented both in the fully coupled 6-hourly model to make the first guess land–atmosphere simulation, as well as in a semicoupled GLDAS Land Information System (LIS; Kumar et al. 2006) to perform the land surface analysis. The CFSR–GLDAS LIS is run once daily (at 0000 UTC) using meteorological forcing from the assimilating background atmospheric model (wind, humidity, pressure, temperature, and downward radiation) and observed, rather than modeled (biased), precipitation. The observation-based precipitation is a blend of the 2.5° merged gauge/satellite CMAP (Xie and Arkin 1997) 5-day accumulations, the 0.5° CPC-UNI (Chen et al. 2008; Xie et al. 2007) daily gauge-only product, and the CFSR background 6-hourly precipitation forecast. At the end of the 24-h CFSR–GLDAS LIS hindcast, the simulated soil moisture and soil temperature of the Noah model are inserted into the CFSR 0000 UTC to serve as the land surface initial conditions for the next CFSR 0000 UTC cycle (Saha et al. 2010).

It is well known that even when supplied with common atmospheric forcings, LSMs yield disparate results, caused by differences in both model parameterization (e.g., canopy resistance, soil texture, rooting depth, and vegetation morphological and physiological characteristics), process representation (i.e., bare soil evaporation, transpiration, canopy evaporation, and runoff generation), and specified boundary conditions (e.g., Dirmeyer et al. 2006a, 1999; Haddeland et al. 2011; Henderson-Sellers et al. 1993; Mitchell et al. 2004). For example, the rooting depth confines the availability of water for transpiration; stomatal resistance, including any incorporated Jarvis-type (Jarvis 1976) stress functions, modulates the rate of transpiration; and the canopy-interception formulation and interception reservoir capacity determine the amount of “fast” moisture flux that bypasses the soil–vegetation continuum. One important distinction between LSMs and reanalyses is that the former close the surface water and energy budget by construct while the latter do not. Related to this point, we should note that the EF from models was calculated using the sum of LE and H in the denominator of Eq. (1) rather than the difference of Rn and G.

2) Wetness index

A meteorological wetness index valid for the AMSR-E-derived convective season [see section 2b(1)] was calculated using output from the PGF–VIC for the period of 2003–07. We first calculate the dryness index (D) following
e10
where (W m−2) is the mean convective-season net radiation over the 5-yr period, (mm season−1) is the mean convective-season precipitation over the 5-yr period, and λ (J kg−1), the latent heat of vaporization, is corrected for the mean convective-season Ta (°C) following
e11
Over a majority of the earth’s surface, D exceeds unity, which implies that water not energy availability controls most land surface evaporation. By taking the reciprocal of D, we obtain a wetness index (W) that for values of D greater than or equal to 1 is bounded between 0 and 1.

3) Ecosystem rooting depth

The International Satellite Land Surface Climatology Project Initiative II (ISLSCP-2) global estimates of mean 95% ecosystem rooting depth (dr) at 1.0° resolution (Schenk and Jackson 2002b) are based on 519 point-based vertical root profile measurements taken between the years 1925 and 2001, predominantly from locations in North America and Europe (Schenk and Jackson 2002a). Grid-averaged values, which range from 0.3 to 4.6 m, were interpolated or extrapolated from these observations as a function of water availability. If soil water is available at depth and there is evaporative demand (i.e., high potential evapotranspiration and low precipitation), the model predicts deeper roots (Schenk and Jackson 2002b). Accordingly, the deepest rooting depths are found in seasonally dry tropical climates where water is stored deep in the soil during the wet season and where there are high evapotranspiration demands throughout the year.

d. Methods

Each of the datasets was resampled from their native resolution to 1.0° resolution using box averaging. We use the early morning descending node (~0130 LT) AMSR-E SM estimates, midday RS–EF, and the afternoon ascending node (~1330 LT) AIRS LCL. The time lag is essential for measuring the effect of SM on EF as well as the growth of the PBL. The PBL grows steadily in clear-sky conditions over the course of the day, reaching its maximum in the late afternoon. Note that AMSR-E SM and AIRS LCL are instantaneous estimates, whereas RS–EF represents a time average. Whenever SM and LCL are taken from models, they are sampled from the 3- or 6-h model time step that contains the respective time of Aqua’s overpass on that particular day. In all cases, EF is taken from a time step 12 h subsequent to that of SM. Only positive values of LCL were included; negative LCL implies ongoing precipitation. Previous studies, unhindered by the sampling limitations of polar orbiting satellites, have considered SM, EF, and LCL in terms of daily mean quantities (Betts 2009).

Our primary focus is on quantifying the influence of the surface state on the PBL, which we represent by the correlation strength between SM and LCL. In reality, however, EF serves as a critical intermediary control in the interaction. Therefore, SM–EF and EF–LCL correlations are considered as well. In a moisture-limited regime, variations in SM affect the partitioning of available energy (RnG) at the surface. A decrease in SM will lead to an increase in H (decrease in EF) and associated increase in LCL (by an increase in Ta and decrease in RH). Accordingly, we are interested in regions characterized by strong negative SM–LCL and EF–LCL correlation and strong positive SM–EF correlation. Consider that in humid regions, with large stores of SM, decreases in SM exert less of an impact on EF and, hence, LCL.

Technically, all links in the SM–LCL process chain are independent in RS because of the independence of the SM, EF, and LCL products. This is also partly true of LSMs because the LCL is computed from the atmospheric forcing and is not an intrinsic property of the LSM itself. The LSMs’ SM is able to affect the EF (to varying degree, dependent on root zone water availability), but the EF cannot affect the LCL. Only in reanalysis are all links able to communicate.

Our estimates of coupling are critically dependent on the accuracy of variable fields as well as their interconsistency. To account for the case that either the land or atmospheric fields, or both, are flawed in a particular model, we calculated a hybrid consensus of coupling strength. Specifically, we calculated estimates of coupling for an ensemble of multimodel and observation-model pairings that included all possible land (SM and EF)–atmosphere (LCL) source combinations.

Whenever reference to “correlation” is made, it implies the nonparametric Kendall’s tau (τ) correlation statistic (Press et al. 1992). Importantly, this metric was used to accommodate for the fact that the relationships (SM–LCL, SM–EF, and EF–LCL) considered in this study are inherently nonlinear, with an abundance of observations at the extremes of SM. Such a relationship would be poorly measured by a standard linear correlation metric such as Pearson’s R. Specifically, linear metrics would be sensitive to differences in the dynamic range (Koster et al. 2009) of the RS and model-based SM estimates. The threshold values of statistical significance at the p = 0.01 and p = 0.05 levels for τ are provided for various sample sizes in Table 2.

Table 2.

Threshold values of Kendall’s τ statistical significance (from zero) at the p = 0.01 and p = 0.05 levels for various sample sizes.

Table 2.

Not all models are directly comparable. Throughout the paper, the PGF–VIC output is treated separately. This was necessary because the PGF does not provide the true meteorological time series, but rather one that is statistically representative at a monthly time scale. GLDAS–VIC was run in water balance mode and, thus, only provides estimates of SM and LCL. Results from NARR, the only regional reanalysis considered in this study, are confined to appendix B and supplemental Tables S4–S9.

This study is temporally confined to the (static) AMSR-E-derived convective rainfall season [see section 2b(1)], which varies in length from 0 to 5 months (Fig. 1d), during the period of 2003–07. This does not preclude coupling effects on large-scale (i.e., stratiform) rainfall systems that occur (sometimes coincidentally with convective events) during the convective season, as these events are included in our study, but serves to isolate the local warm wet season when rainfall is most sensitive to surface fluxes. During this period, coupling signatures are more readily identifiable because of the frequency of wetting-up and drying-down events. In some regions, the feasibility of AMSR-E SM and AIRS LCL retrievals severely limits or impedes any meaningful analysis. Figures 1a–c illustrate the global sampling characteristics of (quality cleared) satellite-based SM, EF, and LCL retrievals over the course of the study period, as well as the knock-on effect to the respective correlation sampling characteristics (Figs. 1e–g). Soil moisture retrievals are limited because of canopy density in the Amazon, Yukon, Congo, and Indonesia and because of surface water fraction in northern Russia. AIRS retrievals are limited by clouds and aerosols. Only pixels with greater than 50 quality-cleared sample pairs were included in the analyses.

Fig. 1.
Fig. 1.

(a)–(g) Summary of the sampling characteristics underlying the 5-yr (2003–07) analysis, including the independent coverage (in days) of (a) quality-cleared AMSR-E SM on the descending orbital node (~0130 LT), (b) SRB/ISCCP EF at midday and (c) quality-cleared AIRS LCL on the ascending orbital node (~1330 LT), the (d) AMSR-E-derived convective-season length (in months), and the joint coverage (in days) of (e) panels (a) and (c), (f) panels (a) and (b), and (g) panels (b) and (c). Also shown is the (h) AIRS-derived clear-sky frequency, which was not applied. Note that a region with a 3-month convective season has a potential sample size of approximately 450 days (5 years × 3 months × 30 days).

Citation: Journal of Hydrometeorology 13, 3; 10.1175/JHM-D-11-0119.1

If the record is first screened for days with substantial early morning cloud cover (i.e., when coupling is unlikely), as defined by AIRS infrared (IR) cloud fraction ≥ 0.4, the number of available samples diminishes significantly (Fig. 1h)—down to 50% over the southern Great Plains. Rather than constrain our analysis to this limited subset of days, we decided against cloud screening. Similarly, we evaluated a rainfall threshold exceedance approach to filter atmospherically controlled days from our analysis. We found that selecting a global product-specific rainfall threshold was too arbitrary and hindered interpretation of the results by adding an additional degree of complexity/ambiguity. Therefore, no screening was performed beyond 1) the cloud screening inherent in the AIRS quality control [see section 2a(3)] and 2) the exclusion of negative LCL samples.

While reanalyses provide the advantage of a static numerical weather prediction system, they are not immune to artificial inhomogeneities caused by observing system changes, or so-called observational shocks (Thorne and Vose 2010). Three of the most substantial shocks correspond to the introduction of Special Sensor Microwave Imager (SSM/I) in July 1987 and AMSU-A in November 1998 (Onogi et al. 2007; Robertson et al. 2011) and the loss of AMSU-A in March 2008 (Rienecker et al. 2011). Coincidentally, our 5-yr study period (2003–07) avoids any known step shift.

The parameters selected for the conditional analysis (section 3c) are integrative of important underlying biophysical and meteorological controls on LE and likewise are expected to correlate with coupling strength. The parameter sources were chosen on the basis that they provide the best representation of current conditions and should not be expected to correspond with actual inputs used in the RS or model formulations. Coincidently, the LAI and lc do serve as inputs to the RS LE.

3. Results

a. Global analysis

Figure 2 provides a summary of the multimodel agreement with RS estimates of SM, EF, and LCL as measured by the median τ. The results of each separate intercomparison are provided in appendix A. In general, SM estimates diverge over high latitudes and densely vegetated forests. This pattern corresponds well with the known limitations of the AMSR-E SM retrievals [see section 2a(1)]. EF estimates vary widely over the tropical rain forests and Pacific Northwest United States. We hypothesize that this is due in part to differences in radiation forcing and rainfall interception formulations. Agreement among sources of LCL varies with radiosonde launch site density, with the poorest correspondence in the tropics. Considering the strong correlation between AIRS and modeled LCL over the United States and western Europe, we argue that high confidence should be assigned to AIRS elsewhere at midlatitude where ground-based observations are lacking. Confidence at midlatitudes, however, does not imply confidence in the tropics because AIRS retrieval skill is known to degrade with increasing cloud cover (Ferguson and Wood 2010; Susskind et al. 2003). To truly resolve the issue of AIRS retrieval skill over the tropics, more ground truth observations are required.

Fig. 2.
Fig. 2.

Median Kendall’s τ between RS and modeled estimates of (a) SM, (b) EF, and (c) LCL. The median is computed from 3-hourly models only, a set of eight SM (GLDAS–CLM/Mosaic/Noah/VIC, CFSR, ERA-INT, MERRA, and NOAA–CIRES), seven EF (GLDAS–CLM/Mosaic/Noah, CFSR, ERA-INT, MERRA, and NOAA–CIRES), and five LCL (GLDAS, CFSR, ERA-INT, MERRA, and NOAA–CIRES) datasets. The global mean of the medians is denoted on each panel. Missing pixels result from either a lack of coverage (n < 40) and/or lack of convective season (i.e., Sahara Desert). The contributing intercomparisons are illustrated separately in Figs. A1A3.

Citation: Journal of Hydrometeorology 13, 3; 10.1175/JHM-D-11-0119.1

Figures 35 illustrate the global distribution of τSM–LCL, τSM–EF, and τEF–LCL derived from RS (Fig. 3, top row) and the offline models and reanalyses. Each of the distributions is in turn summarized by a boxplot diagram in Fig. 6. We have also included box plots derived from the time-averaged 6-hourly version of all 3-hourly datasets in Fig. 6. It is shown that the effect of using 3-hourly/6-hourly data is minor compared to the considerable intermodel differences in global median and interquartile range (variability). For each of the links, the ranges of global median τ (model) are summarized here: τSM–LCL ≈ −0.44 (NOAA–CIRES and NCEP-R1/-R2) to −0.10 (PGF–VIC), τSM–EF ≈ +0.52 (CFSR and NOAA–CIRES) to +0.02 (RS), and τEF–LCL ≈ −0.51 (CFSR, MERRA, and NOAA–CIRES) to ±0.04 (GLDAS–Mosaic/Noah and RS). The interquartile ranges in τ (model) varied from ~0.29 (JRA-25, MERRA, and RS) to ~0.14 (GLDAS–VIC and PGF–VIC) for τSM–LCL, ~0.54 (GLDAS–Mosaic/Noah) to ~0.14 (NCEP-R2 and RS) for τSM–EF, and ~0.49 (GLDAS–Mosaic and PGF–VIC) to ~0.16 (CFSR and RS) for τEF–LCL.

Fig. 3.
Fig. 3.

The (a) τSM–LCL, (b) τSM–EF, and (c) τEF–LCL for (top)–(bottom) RS, PGF–VIC, GLDAS–CLM, GLDAS–Mosaic, GLDAS–Noah, and CFSR. The global mean is denoted on each panel. Missing pixels result from either a lack of coverage (n < 40) and/or lack of convective season (i.e., Sahara Desert). The statistical significance at each pixel is dependent on sample size (see Table 2).

Citation: Journal of Hydrometeorology 13, 3; 10.1175/JHM-D-11-0119.1

Fig. 4.
Fig. 4.

As in Fig. 3, but for (top)–(bottom) ERA-INT, JRA-25, MERRA, NCEP-R1, NCEP-R2, and NOAA–CIRES models. Note that JRA-25 and NCEP-R1/-R2 have a temporal resolution of 6 h (as opposed to 3 h).

Citation: Journal of Hydrometeorology 13, 3; 10.1175/JHM-D-11-0119.1

Fig. 5.
Fig. 5.

The τSM–LCL computed for GLDAS–VIC.

Citation: Journal of Hydrometeorology 13, 3; 10.1175/JHM-D-11-0119.1

Fig. 6.
Fig. 6.

Boxplot summary of the global distributions of (a) τSM–LCL, (b) τSM–EF, and (c) τEF–LCL. The dotted line in each subplot denotes the median of the RS τ. Time-averaged, 6-hourly versions of 3-hourly datasets (plotted in light shading) are denoted with a “6” in parentheses. The sample size is approximately 13 000 pixels. The whiskers extend to the 5th and 95th percentiles. The figure is provided in tabular form in supplemental Tables S1–S3.

Citation: Journal of Hydrometeorology 13, 3; 10.1175/JHM-D-11-0119.1

Figures 36 underscore the fact that RS offers an account of coupling that is consistently (and substantially) weaker than models, with the exception of τSM–LCL from PGF–VIC, which exhibits very weak coupling even for water-limited regimes. In Fig. 6, from left to right, a general trend of enhanced coupling from RS to LSMs to reanalyses is shown. This is not entirely unexpected considering the absence of intervariable dependencies in RS, only limited dependencies in the SM–EF link for LSMs, and the coupling of variables in reanalyses.

PGF–VIC has a global median τEF–LCL equivalent to +0.20, which is of opposing sign to most other models (Fig. 3). A closer look at Manaus, Brazil (2.61°S, 60.21°W), where PGF–VIC estimates of τEF–LCL are strongly positive (+0.48), revealed that PGF–VIC EF steadily increases over the course of the convective season (March–May), relative to GLDAS–CLM, which exhibited a larger dynamic range but no trend (not shown). Over the same period, the limited positive PGF–VIC LCL samples (remaining after screening) also trend slightly upward, driven by increasing Ta (not shown), hence the positive PGF–VIC τEF–LCL.

We could not compare GLDAS–VIC and PGF–VIC τSM–EF or τEF–LCL directly because GLDAS–VIC was run in water balance mode. However, GLDAS–VIC does provide an estimate of τSM–LCL (Fig. 5) that was found to vary substantially from that of PGF–VIC (Fig. 3). Given that the VIC versions are nearly identical (J. Sheffield 2011, personal communication), it follows that the divergence in their coupling can only be attributable to either model forcing and/or model calibration, which was carried out for PGF–VIC and not for GLDAS–VIC. Calibration by disaggregated runoff could potentially skew model performance skill toward more humid regimes where that data is available (see Troy et al. 2008, their Fig. 1) and, ultimately, play a role in favoring energy-limited mean states in VIC.

The fact that RS–EF is insensitive to SM (Fig. 3) is not surprising because SM controls were only indirectly (through VPD) implemented in the RS–LE algorithm. The lack of a water budget (i.e., SM or precipitation) constraint on RS–LE is a well-known limitation of standalone estimates of RS–EF in water-limited regimes (Ferguson et al. 2010). Among the GLDAS models, CLM yields the strongest coupling for all process links; it is also most responsive to atmospheric forcing according to intercomparisons of the SM and EF coefficient of variation (CV; Fig. D1). Despite sharing a common LSM, GLDAS–Noah and CFSR produce substantially different estimates of coupling, as accentuated by differences in τEF–LCL. On one hand, GLDAS–Noah τEF–LCL (+0.002) is statistically insignificant (see Table 2), and on the other, CFSR τEF–LCL (−0.50) is stronger than the set of all models, besides MERRA.

Figure 7 shows the consensus view of 23, 24, and 21 estimates of the τSM–LCL, τSM–EF, and τEF–LCL, respectively. Here, consensus is measured in terms of the number of ensemble members for which the given pixel lies in the lowest quartile [(−)τ] for τSM–LCL and τEF–LCL or upper quartile [(+)τ] for τSM–EF. Importantly, the strength criterion (i.e., first or third quartile) is calculated from each ensemble member’s inherent global distribution. As detailed in Tables C1C3, the ensembles comprise both model-only and mixed model–observation pairings (i.e., RS–SM plus modeled–LCL, or vice versa, and so on), hence we refer to them as “matched+mixed-model” ensembles. To maintain physical consistency within pairings, only 3-hourly model outputs were included. Accordingly, JRA-25 and NCEP-R1/-R2 were not considered. PGF–VIC was also omitted because its forcing dataset does not conserve the precise date or intensity of precipitation events. We identified four regions of complete matched+mixed-model consensus with regard to the strength of SM–LCL coupling: Iberian Peninsula, northeastern Brazil (i.e., Bahia and Piaui), northwestern United States (eastern Washington, southern Idaho, southeast Oregon, and northern Nevada and Utah), and scattered parts of India. We were intrigued to find that only in the case of northwestern United States were comparable levels of consensus not found in the intermediary links (SM–EF and EF–LCL). Of the four regions, northwestern United States has the distinction of having the poorest correspondence between RS and modeled EF (see Fig. A2). Accordingly, RS and RS-model ensemble members serve to weaken the consensus coupling there. Additional regions for which consensus was reached among 20 or more (of 24) ensemble members for τSM–LCL include southwest United States (Arizona and New Mexico), northern Mexico, northern and Western Australia, south and central Asia (including India), Turkey, southern and eastern Africa, and parts of the Sahel. Note that the areas of strongest coupling are confined to latitudes south of 50°N and north of 40°S, extending approximately from the Canada–U.S. border south to the southernmost tip of Australia.

Fig. 7.
Fig. 7.

The consensus regions of (a) τSM–LCL, (b) τSM–EF, and (c) τEF–LCL derived from an ensemble of matched+mixed-model experiments. Here, the scale represents the number of experiments that yield a coupling strength (τ) in the lowest [upper for (b)] quartile of their global probability distribution. To ensure that all pixels have the same maximum sample size, all fields were first masked according to the respective RS coverage. Twenty three, 24, and 21 ensemble members contributed to panels (a),(b), and (c), respectively (see Tables C1C3 for ensemble member composition).

Citation: Journal of Hydrometeorology 13, 3; 10.1175/JHM-D-11-0119.1

Dirmeyer (2011) cautioned that the use of correlation metrics as an indicator of coupling could be misleading. For example, in certain regions the EF may be highly sensitive to SM, but the SM itself varies unsubstantially, or vice versa. We have plotted the multimodel median CVSM, CVEF, and CVLCL in Fig. 8. The figure shows that CVLCL does exhibit low values in many places where consensus coupling is high and so might mislead the coupling interpretation. The CVLCL exceeds the 25th percentile (Q1) of the global distribution in 24% (EF–LCL) to 49% (SM–EF) of the consensus area (arbitrarily taken to be 16 or more models in Fig. 7 for all links; Fig. S1). In comparison, CVSM and CVEF exceed their global Q1 in only 0.3% (SM–EF) to 3.7% (EF–LCL) and 1.2% (SM–EF) to 7.7% (SM–LCL) of the consensus subarea, respectively (Fig. S1). Using the Q1 of the multimodel median CVLCL as an exceedance criterion, we postfiltered the areas of strong consensus and the following areas were eliminated: Western Australia, the Horn of Africa, central Asia, and parts of western United States and northeastern Brazil (Fig. S2). In other words, the most arid regions were filtered out; an annual rainfall exceedance criterion of 0.4 m yr−1 would have achieved similar results. In the case of these regions, we caution that elevated coupling strength (τSM–LCL and τEF–LCL) may be an artifact of the analysis technique employed.

Fig. 8.
Fig. 8.

Median (a) CVSM, (b) CVEF, and (c) CVLCL calculated from the set of all models at their native temporal resolution (i.e., as in Figs. 35). Accordingly, the sample set comprises 12, 11, and 11 members for SM, EF, and LCL, respectively.

Citation: Journal of Hydrometeorology 13, 3; 10.1175/JHM-D-11-0119.1

In the previous aside, our use of the Q1 multimodel median CVLCL was arbitrary and intended for the purpose of discussion only. Figures D1D3 illustrate the broad intermodel range in global mean CV for all variables: from 0.109 (ERA-INT) to 0.430 (JRA-25) for SM, from 0.291 (ERA-INT) to 0.785 (GLDAS–CLM) for EF, and from 0.292 (RS) to 0.780 (NCEP-R1) for LCL. Accordingly, it is easy to imagine that on a single dataset basis the magnitude and filtering effectiveness of the Q1 CVLCL may vary substantially. In practice, selection of an appropriate variability threshold should be left to the careful discretion of the end user because in some instances even a small absolute shift in LCL could have a significant impact on precipitation formation (Betts 2009; Ferguson and Wood 2011; Findell and Eltahir 2003a,b). For this reason, we opted against variability screening in our study. We note that Dirmeyer (2011) also chose not to recommend a threshold.

Regionally, the greatest intermodel disagreements in coupling strength occur over densely vegetated and/or energy-limited (i.e., taiga) regions. For example, over the Amazon River basin, the τEF–LCL ranges from +0.46 (PGF–VIC) to −0.74 (CFSR) (Table S6). NCEP-R2 demonstrates enhanced coupling relative to NCEP-R1 for all links (SM–LCL, SM–EF, and EF–LCL) over the Amazon and Congo River basins, as well as a distinct reduction in τSM–EF over the Pacific Northwest United States. Globally, there is a substantial reduction in CVLCL in NCEP-R2 relative to NCEP-R1 (Fig. D2). Both of these shifts are likely due in part to improvements in the radiation scheme, cloud parameterizations, and RH–cloud relationship implemented in NCEP-R2. The JRA-25 produces very weak coupling over central and eastern United States, western Europe, and the Pampas of South America. This feature is likely attributable in part to the inaccuracy of the land classification assigned to those regions: cultivated wheat fields (Sato et al. 1989, their Fig. 1; see also Sud et al. 1990, their Fig. 1). Between the two most modern reanalyses, MERRA and CFSR, MERRA yields stronger τSM–LCL over the tropics and midlatitudes, while CFSR produces stronger τSM–LCL coupling over the taiga. MERRA τSM–EF exceeds CFSR τSM–EF over South America and central Africa. In terms of τEF–LCL, MERRA is more strongly coupled everywhere except the Amazon, Southeast Asia, southern West Africa, and northern South Africa (Angola, Zambia, Mozambique, and Tanzania).

Reichle et al. (2011) recently documented errors in the amplitude and phase of MERRA’s diurnal cycle of precipitation that are further exasperated by shortcomings in MERRA’s radiation scheme. Basically, MERRA’s precipitation rates are lower than observed, tending toward long-lasting “drizzle,” and the diurnal cycle is weighted toward midday. In reality, observations often show nighttime precipitation maxima. During precipitation events, the incoming solar radiation is not sufficiently reduced. The result is that the canopy interception (evaporation) is too high and an unrealistically small balance of precipitation remains to infiltrate the soil or generate surface runoff. This highlights a situation where the intermediary link, SM–EF, can be of negative sign (SM decreases while EF and Ta increase), while the net coupling (SM–LCL) is still negative. Figure D2 shows a clear bullseye in MERRA CVEF over central Africa, where MERRA’s mean interception loss fraction exceeds 0.5 (Reichle et al. 2011, their Fig. 4), that is likely an artifact of the process deficiencies discussed above.

b. Basin-scale analysis

Twenty-six large-scale hydrologic basins and three GLACE hotspots were selected for intermodel comparisons (Fig. 9). The basins span a broad spectrum of scale (from 37 to 481 km2), convective-season length (2.0–4.4 months), meteorological wetness (0.04–0.46), and land cover composition (see Table 3). Figure 10 shows the relative ranking of the regions according to the multimodel median τSM–LCL, τSM–EF, and τEF–LCL (RS not included). In each case, Limpopo, Niger, and the GLACE hotspots [central Great Plains (CGP), India (IND), and West Africa (WAF)] fall within the top 10 strongest regions of coupling. Limpopo is one of the smallest basins and happens to be situated in an area of uniformly strong multimodel consensus (see Fig. 7). Thus, the coupling signal is not degraded by the effect of spatial averaging over heterogeneous signals as may occur in larger basins. Figure 10 reveals an interesting distinction between IND and Indus—two regions in very close geographic proximity of one another. In this case, the coupling signal is similar for τSM–EF, but varies substantially for τEF–LCL. This is likely due in part to differences in wetness index (W; Indus: 0.07; IND: 0.29) and lc (Indus: barren or sparsely vegetated; IND: crops). Tables 46 provide a summary of the multimodel statistics at each basin, including rank by median τ, rank by standard deviation of τ, 3- and 6-h model consensus counts, and RS τ. A more comprehensive summary with independent model statistics (τ and CV) for each region is available in supplemental Tables S4–S9. The largest intermodel disagreement (uncertainty) occurs for the Amazon, Pearl, Congo, and Danube, which are ranked in the top 10 by standard deviation for all links. Modeled coupling is always more profound than RS, although RS generally yields a similar basin rank order (Tables 46). For example, CGP, IND, Limpopo, Murray–Darling, Niger, and Senegal rank among the top 10 regions of coupling strength for all links based on RS.

Fig. 9.
Fig. 9.

Map of the 26 hydrologic basins and three GLACE land–atmosphere interaction hotspots selected for analysis.

Citation: Journal of Hydrometeorology 13, 3; 10.1175/JHM-D-11-0119.1

Table 3.

Descriptive hydroclimatic and vegetative characteristics of the 26 hydrologic basins and three GLACE land–atmosphere interaction hotspots selected for analysis. Only the three (or fewer) most prevalent land cover classes are listed. For land cover acronyms, refer to Table 7.

Table 3.
Fig. 10.
Fig. 10.

Regional coupling strength according to the median (a) τSM–LCL, (b) τSM–EF, and (c) τEF–LCL calculated from the set of all models at native resolution (taken from Tables 46).

Citation: Journal of Hydrometeorology 13, 3; 10.1175/JHM-D-11-0119.1

Table 4.

Summary of the τSM–LCL statistics for each of the 26 hydrologic basins and three GLACE land–atmosphere interaction hotspots (see Fig. 9). Here, the median and standard deviation (SD) are calculated using the basin mean τSM–LCL from each of the models (n = 12) at their native temporal resolution (i.e., as in Figs. 35). For a full listing of statistics by model, including NARR (where applicable), see supplemental Table S4.

Table 4.
Table 5.

As in Table 4, but for the τSM–EF statistics. The median and standard deviation of τSM–EF are calculated from 11 models (as in Figs. 34). For a full listing of statistics by model, including NARR (where applicable), see supplemental Table S5.

Table 5.
Table 6.

As in Table 4, but for the τEF–LCL statistics. The median and standard deviation of τEF–LCL are calculated from 11 models (as in Figs. 34). For a full listing of statistics by model, including NARR (where applicable), see supplemental Table S6.

Table 6.

c. Conditional analysis

As shown in Fig. 11, we have investigated the strength of coupling as a function of dr, LAI, W, and lc, of which the global distributions are illustrated in Fig. 12. We calculated the median global τ for each decile of dr, LAI, and W, and for each separate lc (sample sizes provided in Table 7). Also included in Fig. 11 are the trends in 3- and 6-hourly multimodel consensus. Because of retrievability constraints on RS products, particularly SM (see Fig. 1), the underlying RS sample population does differ from that of the models (see Table 7). We calculated the model statistics after first applying a filter for RS retrievability and found the results to differ unsubstantially, except in the case of τSM–LCL and τSM–EF for deciduous needleleaf forests (DNF; RS: n ≤ 22), for which model results converge to the RS τ values of approximately zero (not shown). Rather than focus on the intermodel differences, we will characterize and attribute the general signature of the consensus trends. It is important to keep in mind that a broad range of climate regimes (i.e., W), ranging from soil moisture to energy limited, may be encompassed in a single land cover type or decile of dr and LAI (see Figs. S3–S5).

Fig. 11.
Fig. 11.

Summary of the 60°S–70°N median of (left)–(right) τSM–LCL, τSM–EF, and τEF–LCL from RS, LSMs, and reanalyses, conditioned on (a) dr, (b) W, (c) LAI, and (d) UMD lc. Data points are defined at each decile (provided on the leftmost y axes) for (a)–(c), and in the case of (d), for each lc. The deciles and median values are calculated using the population of pixels where an AMSR-E-derived convective season is defined (see Fig. 1d). W and LAI are calculated over the 2003–07 and 2003–06 (because of data availability) convective seasons, respectively. The dr is time invariant. As detailed in Table 7, the sampling varies for RS because of retrievability. Also included (in gray) are the median number of models, either matched+mixed (solid), 3 hourly (dashed), or 6 hourly (dotted), for which τ ranks within the top quartile of their global distribution. The 3-hourly (6 hourly) model-only subsets comprise 9 (12), 8 (11), and 8 (11) members for τSM–LCL, τSM–EF, and τEF–LCL, respectively. Note that the x axis of the center (τSM–EF) column is reversed to maintain the strength convention. For the spatial distribution of each dr, W, and LAI decile, see Figs. S3–S5.

Citation: Journal of Hydrometeorology 13, 3; 10.1175/JHM-D-11-0119.1

Fig. 12.
Fig. 12.

Global hydroclimatic and vegetation morphological parameters: (a) dr, (b) convective-season LAI, (c) convective-season W, and (d) UMD lc. The lc’s are defined in Table 7.

Citation: Journal of Hydrometeorology 13, 3; 10.1175/JHM-D-11-0119.1

Table 7.

UMD dominant lc composition and RS retrievability at 1.0° resolution for all pixels with an AMSR-E-derived convective season (see Fig. 1d) between 60°S and 70°N. The urban land cover class (class No. 13) is excluded because no 1.0° pixels were so classified.

Table 7.

1) Rooting depth

Most models show a direct correlation between τSM–EF (τSM–LCL and τEF–LCL) in the interquartile range (third quartile) of dr. For the highest deciles of dr, the relationship reverses in sign with the deepest rooted vegetation exhibiting muted coupling strength. The trend in coupling strength tends to rise and fall with the proportional composition of open shrublands (OS) and barren or sparsely vegetated (BAR) lc’s (see Fig. S6), except for the lowest three deciles, which populate high-latitude energy-limited evapotranspiration regimes (i.e., limited growing and convective-season lengths; see Figs. 1 and S3). In regions where energy constraints on evapotranspiration predominate, we do not expect to observe a strong land signal; this is confirmed by the models. We also do not expect a strong land signal where rooting depths are large because in these areas transpiration is either insensitive to daily meteorological forcing (because of access to deep soil moisture stores) or the rainfall is insufficient to have any lasting effect on evapotranspiration. On the other hand, we do expect strong coupling in semiarid regions where moderate rooting depths are found and indeed it is. Specifically, the eighth decile of dr comprises OS and savannas (SV) of southwestern United States, Mexico, southern Africa, and central Australia (Fig. S3). In these regions, plants use deep soil moisture stores for transpiration, enabling long-term climate memory (i.e., evapotranspiration decay time scales) between precipitation events. RS estimates of τSM–LCL share a similar correspondence with dr, but RS τSM–EF and RS τEF–LCL do not vary with dr. This makes sense, considering that only the LSMs and reanalyses maintain a soil column/root distribution profile.

2) Wetness index

There are dual peaks in the consensus τSM–LCL and τSM–EF at the second and eighth deciles of W and a single peak in the consensus τEF–LCL at the eighth decile. The double peak trend generally follows the fractional composition of grassland (GS) across deciles in W (see Fig. S7), although the trend is mostly driven by the diminishing fraction of OS and BAR with increasing W, as well as the large fraction of SV and evergreen broadleaf forests (EBF) composing the highest deciles of W. A summary of the dominant lc’s at each decile of W is instructive (see also Figs. S4 and S7): 1st decile: midlatitude soil moisture-limited OS and BAR; 2nd decile: midlatitude soil moisture-limited and northern high-latitude energy-limited OS, BAR, and GS; 3rd and 4th deciles: midlatitude semiarid and northern high-latitude energy-limited OS, GS, and croplands (CP); 5th through 8th deciles: midlatitude energy-limited OS, CP, and mixed forests (MF); 9th decile: energy-limited woody savannas (WS), SV, and tropical EBF (primarily the Congolese rain forest); and 10th decile: energy-limited EBF (primarily Amazonia). Accordingly, we expect coupling strength to diminish with increasing W, which is what the consensus shows, except for deciles of moderate wetness (fifth through eighth) in which case zonal sampling and CP fractional composition play a role in increasing consensus coupling strength. RS τSM–LCL follows the consensus trend, but neither RS τSM–EF nor RS τEF–LCL were found to vary with W. This behavior highlights the lack of a direct soil moisture constraint on RS–LE.

The peak in τEF–LCL at the second decile is observed in some models, but it does not come through in the consensus. In midlatitude soil moisture-limited regions, LCL exhibits markedly lower variability than EF (Fig. 8). As discussed previously (section 3a), it is possible in these regions that the magnitude of evapotranspiration is insufficient to substantially impact regional temperature and humidity. Notably, the intermodel spread in τEF–LCL more than triples from dry to wet extreme. This is most likely attributable to differences in the precipitation and radiation schemes in the case of reanalyses as well as the water holding capacity and canopy interception formulations of all models.

3) Leaf area index

The consensus coupling strength peaks at the second decile of LAI and steadily declines thereafter with increasing LAI (reduced fractional coverage of BAR, CP, and OS; see Fig. S8). The lowest decile of LAI comprises BAR approximately correspondent with the lowest decile of W, where the impact of soil moisture anomalies is short lived (see Figs. S4–S5). The second through fifth deciles of LAI comprise midlatitude semiarid OS, GS, and SV, where soil moisture is the primary control for evapotranspiration, hence the strong coupling. The sharp decline in coupling strength between the second and fifth deciles can be explained by the diminishing proportional composition of OS and GS (Fig. S8). The sixth through ninth deciles of LAI comprise northern high-latitude taiga and temperate forests that are predominantly energy limited. The uppermost decile of LAI comprises temperate deciduous forests and EBF. RS τSM–LCL varies similarly (inversely) with LAI. However, RS τSM–EF and RS τEF–LCL are found to be insensitive to LAI. LAI is used in the RS–LE formulation to scale the stomatal conductance, but these monotonic increases in LE (and EF) with increasing LAI has no effect on SM–EF and EF–LCL coupling strength.

4) UMD land cover

According to both models and RS, the strongest coupling occurs for closed shrublands (CS) and SV. Both land cover types occupy semiarid climate niches that are narrowly defined. Closed shrublands occur in areas that are more humid than open shrublands and less humid than grasslands or croplands, correspondent to only 0.6% of the area considered (Table 7). Savannas, defined by a coexistence of shallow rooted herbaceous and deep rooted woody vegetation, occupy tracts of land between the equatorial rain forests and midlatitude deserts, covering 6.4% of our study area (Table 7). Savannas have shown elevated sensitivity to changes in climate forcing as manifested by dramatic changes in structure and function (Scanlon and Albertson 2003; Williams and Albertson 2005, 2006). By contrast, other lc’s are less constrained by climate. For example, open shrublands occupy areas that span the dry half of the global W spectrum. Croplands occupy areas of widely varying W as well. The inclusive dry and wet extremes act to reduce the median coupling strength (i.e., contribute to a bell-shaped rather than skewed distribution of τ). Alternatively, intermodel lc classification differences (because not all models apply the UMD classification scheme) can serve to temper consensus trends, particularly in the case of cropland versus forest classification. Forests are characterized by the weakest coupling strength. As expected, the τSM–EF is strong over barren or sparsely vegetated areas where in the absence of plants there is no mechanism to access deeper soil moisture stores and, thus, sensitivity to short-term forcing anomalies only.

4. Summary and conclusions

In this study, remote sensing was used to assess the ability of five offline LSMs and eight reanalyses to represent the land segment of land–atmosphere coupling, as measured by the daily correlation between SM–LCL, SM–EF, and EF–LCL. The reanalyses span the gamut of modernity, from 1995 (NCEP-R1) to 2008 (NOAA–CIRES), and in this regard the study can be considered to provide an account of the evolution of state-of-the-art understanding of coupling. Model age encompasses feature differences such as spatial and temporal resolution, forecast system, assimilation system, assimilated observations, and LSM. Nevertheless, no clear trend in coupling strength with model age was found. Inclusion of the GLDAS LSMs provided an opportunity to isolate the effects of model parameterization from the effects of model forcing on coupling strength. Using a matched+mixed-model ensemble, global maps of the consensus on coupling strength were generated for each process link (SM–LCL, SM–EF, and EF–LCL). These maps (Fig. 7) support the findings of GLACE (Guo et al. 2006; Koster et al. 2006) that transitional zones between arid and humid climates (typified by shrublands, grasslands, croplands, and savannas) tend to have the strongest coupling. Coincidentally, this is also where RS retrievals of SM, EF, and LCL are reasonably skillful (section 2).

Remote sensing yields a substantially weaker coupling strength over the globe relative to LSMs and reanalyses, particularly for links SM–EF and EF–LCL (Tables 46). For example, over the central Great Plains, the RS (CFSR) estimates of τSM–LCL, τSM–EF, and τEF–LCL are −0.35 (−0.43), +0.10 (+0.66), and −0.13 (−0.50), respectively. In fact, the range of RS and modeled τSM–LCL overlap only for the lowest four deciles of W and LAI (Fig. 11). Interestingly, the rank order of basins by coupling strength calculated from RS and models do roughly agree (section 3b).

The weak coupling between RS SM–EF and EF–LCL may not come as a surprise considering that the RS–LE (EF) formulation [section 2a(2)] does not include a direct soil moisture constraint. To further evaluate the accuracy of RS and modeled τSM–EF, we analyzed in situ measurements of SM and EF from six flux towers for which data was available over our study period to non-Fluxnet investigators: three located in evergreen needleleaf forests (ENF), two located in grasslands, and one located in cropland. We found that the RS τSM–EF does underestimate coupling strength at the grassland and cropland sites but either overestimates or roughly matches the observed τSM–EF at the ENF sites. On the other hand, models (with little exception) were found to overestimate the coupling for all sites (see Fig. S9). These findings, while limited in transferability, support the main conclusion of this study, which is that coupling in LSMs and reanalyses is too strong. An obvious follow-up investigation is to expand the tower analysis to all stations globally.

The LSMs and reanalyses produce vastly different accounts of coupling, as we have demonstrated through global, regional, and hydroclimate/surface parameter conditioned intercomparisons. In a few cases, we have provided informed hypotheses on the attribution of intermodel differences. It is encouraging to find common directional shifts in τSM–LCL, for example, with respect to dr, W, LAI, and lc between the RS and models (Fig. 11). Conversely, the finding that VIC–PGF τSM–LCL does not vary with any of the above parameters is rather alarming. A comprehensive attribution of the intermodel differences extends beyond the scope of our study.

Previously, Wei et al. (2010b) showed that the land–atmosphere coupling signal could be decomposed into two components: the impact of low-frequency external forcing (i.e., precipitation) and the impact of soil moisture (i.e., land surface memory). They determined that the majority of the coupling strength was associated with intraseasonal (20–84 days) precipitation variation. Because AGCMs (e.g., the relaxed Arakawa–Schubert convection scheme) have the tendency to overestimate the variability (i.e., premature triggering of convection; Sun et al. 2006; Wilcox and Donner 2007) in this frequency band, they consequently tend to overestimate the coupling strength as well. These findings support our conclusion that modeled coupling is too strong. They also provide one cause of the step increase in coupling strength between LSMs and reanalyses (Fig. 6). As discussed in section 2(d), another cause of the step increase is that SM and LCL are linked in the reanalyses unlike in the case of RS and LSMs where SM and LCL are independent.

The large range in model coupling strength came as no surprise considering the multimodel spread in water and energy budget partitioning revealed by past intermodel comparison efforts (e.g., Dirmeyer et al. 2006a, 1999; Haddeland et al. 2011; Henderson-Sellers et al. 1993; Mitchell et al. 2004). The dichotomy between this study and those, of course, is scale. Coupling strength expresses the degree of influence that the land surface exerts on the local-scale growth of the diurnal PBL, whereas terrestrial water and energy budgets relate to the ability of models to match seasonal-to-annual aggregations of fluxes at basin scales. The inability of reanalyses to accurately match the observed diurnal cycle of precipitation over land has detrimental consequences to a measure of coupling, but the accuracy of their seasonal or annual mean precipitation could still be high (Bosilovich et al. 2011). Accordingly, for a model to perform well in traditional land hydrology assessments says nothing about the model’s ability to represent real-world coupling. Likewise, a model could represent observed coupling well, but provide biased estimates of the water and energy budget components as a function of its forcing.

Currently, there is a gap in understanding how τSM–LCL, τSM–EF, and τEF–LCL, or variability thereof, translate in physical terms to the hydrological cycle (e.g., LE, precipitation, and runoff). Making this connection is critical in the case of seasonal forecast and climate projection initiatives. For example, what is the sensitivity of rainfall forecast skill to shifts in coupling strength (caused by climate and global change)? Or, how well can intermodel differences in LE be prescribed by differences in coupling? Our preliminary assessment, using outputs of LE from the GLDAS LSMs suggest that intermodel differences in coupling strength (ΔτSM–LCL, ΔτSM–EF, and ΔτEF–LCL) do not correlate well with intermodel differences in LE (not shown). For a given LSM, changes in LE with respect to changes in τ (all links) varied broadly in terms of both magnitude and sign, even in basins of similar climate. For the same basin, even the sign of the shift could not be assumed transferable across LSMs. This is a research theme that deserves continued attention.

Finally, it should be restated that our study has focused on the land segment of land–atmosphere coupling, which is a necessary but not sufficient condition for SM–precipitation feedback. Accordingly, our results could be used to screen studies that incorporate the precipitation feedback. For example, our Fig. 7 (also RS-only results: Fig. 3, top row), which does not show strong consensus in the eastern United States, casts doubt on the realism of EF–precipitation feedbacks reported for that region by Findell et al. (2011). The atmospheric segment is much more uncertain than the land segment (see Seneviratne 2010 for a review).

Acknowledgments

Initial analyses were conducted at Princeton University while the lead author was supported under NASA Earth and Space Science Fellowship NNX08AU28H: Understanding Hydrologic Sensitivity and Land–Atmosphere Coupling through Space-Based Remote Sensing. The manuscript was completed at the University of Tokyo Institute of Industrial Science supported by the lead author’s Japan Society for the Promotion of Science Postdoctoral Fellowship for Foreign Researchers P10379: Climate Change and the Potential Acceleration of the Hydrological Cycle. Publication fees were partially supported by the Catastrophe Perils group at Swiss Re. The following reanalyses were obtained from the Research Data Archive (RDA; http://dss.ucar.edu), which is maintained by the Computational and Information Systems Laboratory (CISL) at the National Center for Atmospheric Research (NCAR), and sponsored by the National Science Foundation (NSF): NOAA–CIRES 20th Century V2.0 (ds131.1), NCEP-R1 (ds090.0), NCEP-R2 (ds091.0), CFSR (ds093.1), JRA-25 (ds625.0), and NARR (ds608.0). GLDAS and MERRA data were obtained from the Global Modeling and Assimilation Office (GMAO) at NASA Goddard Space Flight Center through the NASA GES DISC online archive. ECMWF ERA-Interim data “for research and educational applications” was obtained through the ECMWF data portal (http://data-portal.ecmwf.int/data/d/interim_daily). Justin Sheffield provided the PGF–VIC model output.

APPENDIX A

Intercomparison of Modeled and Observed Variables

We evaluate each of the LSMs and reanalyses outputs against the remote sensing products. In total, there are 11 sources of SM, 10 sources of EF, and 8 sources of LCL. The results are shown in Figs. A1A3. All intercomparisons were carried out using 3-hourly outputs, with the exception of JRA-25 and NCEP-R1/-R2, for which output was only available 6 hourly. In the case of the latter, the only difference is that the reanalysis EF was matched against a 6-hourly linearly rescaled version of RS–EF. We should underscore here that temporal mismatches underlie the SM and LCL intercomparisons; we are comparing 3- or 6-hourly averages with instantaneous estimates. Accordingly, it is possible for precipitation events (either occurring within the time step, or having occurred earlier but are still being carried forward by model memory) or large cloud systems to be captured in the time-averaged model outputs, but missing from the instantaneous satellite-based record. On the other hand, the model forcing and reanalysis assimilating fields should not be viewed as perfect, particularly in data-sparse regions.

Fig. A1.
Fig. A1.

Kendall’s τ correlation between AMSR-E SM and model-based surface layer SM estimates. JRA-25 and NCEP-R1/-R2 are 6 hourly; the rest are 3 hourly.

Citation: Journal of Hydrometeorology 13, 3; 10.1175/JHM-D-11-0119.1

Fig. A2.
Fig. A2.

Kendall’s τ correlation between RS–EF and model-based EF. The JRA-25 and NCEP-R1/-R2 statistics were computed at 6-hourly temporal resolution (with RS–EF linearly rescaled); all other statistics were computed from 3-hourly data matchups.

Citation: Journal of Hydrometeorology 13, 3; 10.1175/JHM-D-11-0119.1

Fig. A3.
Fig. A3.

Kendall’s τ correlation between AIRS LCL and modeled LCL. The temporal resolution of JRA-25 and NCEP-R1/-R2 is 6 hourly; for all others it is 3 hourly.

Citation: Journal of Hydrometeorology 13, 3; 10.1175/JHM-D-11-0119.1

In Fig. A1, the global mean correlation between AMSR-E SM and model-based estimates of surface layer SM is shown to range from 0.175 (NCEP-R2) to 0.270 (CFSR). Not surprisingly, the weakest agreement is between AMSR-E and the 6-hourly models. NOAA–CIRES is an outlier among 3-hourly models, with a τSM that falls within the range of the 6-hourly models. Of the GLDAS models, CLM and Noah are in better agreement with AMSR-E than Mosaic and VIC, which suggests that strength of comparison does not necessary hinge on the (in)consistency of soil layer representative depths: AMSR-E: 2 cm, CLM: 1.8 cm, Noah: 10 cm, Mosaic: 2 cm, and VIC: 10 cm. Retrievals over the western United States, Brazilian Highlands, Iberian Peninsula, southern Africa, central Asia, and Australia yield the highest correlations. In general, agreement is strongest in semiarid regions and weakest over the tropics and northern boreal forests and peatlands. This pattern aligns with the well-documented trend of microwave soil moisture retrieval sensitivity, which significantly degrades over densely vegetated canopies and regions of high surface water fraction (Gao et al. 2006). However, within regions where AMSR-E skill is expected to be high, agreement with models is shown to vary and could serve as an important tool for identifying model deficiencies.

Figure A2 illustrates the correlation between RS–EF and model or reanalysis estimates of EF, which ranges from 0.065 (NCEP-R1/-R2) to 0.115 (ERA-INT). All models and reanalyses compare poorly over the Pacific Northwest United States, central Australia, and the Congolese rain forest. Of the GLDAS models, CLM provides the best agreement over Europe and the Yangtze River basin in China, but is the most weakly correlated GLDAS model over the Amazon. CLM’s spatial correlation structure is shared by the reanalyses, which yield correlations that are broadly higher (except NCEP-R1/-R2).

Of all variables considered (i.e., SM, EF, and LCL), estimates of LCL were found to agree most strongly (Fig. A3). The global mean correlation ranged from 0.269 (NCEP-R1) to 0.356 (CFSR). NOAA–CIRES τLCL (0.270) is the least correlated with AIRS of the 3-hourly subset of models (and quite nearly among all models). This is significant because NOAA–CIRES is the only reanalysis not to assimilate raobs. The implication is that reanalyses may be getting the correct near-surface (2 m) properties by assimilating everything except near-surface (2 m) observations themselves. Also noteworthy is the fact that radiosonde-assimilating reanalyses achieve better agreement with AIRS than the surface-based forcing GLDAS. Spatially, the weakest correlations are observed over the rain forest of the Amazon, Congo, and Southeast Asia, which is likely the result of degraded AIRS retrievals in the presence of clouds (Ferguson and Wood 2010; Susskind et al. 2003). Over Europe and the Pampas/Parana basin of South America, ERA-INT yielded distinctively stronger correlations with AIRS relative to the other datasets. Figure A3 shows a striking falloff in τLCL at the boundary of eastern Europe and Russia in all reanalysis except NOAA–CIRES. This is most likely attributable to the lack of radiosonde observations over Eurasia and China being integrated into the forcing datasets of GLDAS and assimilated into the reanalyses. This phenomenon could also explain the poor correlations over much of Africa. If so, the fact that AIRS compares well with GLDAS and the reanalyses over the developed regions of the world where in situ contributions to the models are sufficiently dense in time and space gives us confidence in AIRS retrievals elsewhere, especially outside of the tropics.

APPENDIX B

Coupling in NARR

APPENDIX C

Ensemble Member Descriptions

  • Table C1.

    Table C1.

    SM–LCL matched+mixed-model ensemble member composition.

    Table C1.

  • Table C2.

    Table C2.

    SM–EF matched+mixed-model ensemble member composition.

    Table C2.

  • Table C3.

    Table C3.

    EF–LCL matched+mixed-model ensemble member composition.

    Table C3.

APPENDIX D

Dataset Dispersion Properties

REFERENCES

  • Adler, R., Wilheit T. , Kummerow C. , and Ferraro R. , 2007: AMSR-E/Aqua L2B Global Swath Rain Rate/Type GSFC Profiling Algorithm V10-Stage 2. National Snow and Ice Data Center, Boulder, CO, digital media. [Available online at http://nsidc.org/data/ae_rain.html.]

  • Aumann, H. H., and Coauthors, 2003: AIRS/AMSU/HSB on the Aqua mission: Design, science objectives, data products, and processing systems. IEEE Trans. Geosci. Remote Sens., 41, 253264.

    • Search Google Scholar
    • Export Citation
  • Baldocchi, D., and Coauthors, 2001: FLUXNET: A new tool to study the temporal and spatial variability of ecosystem-scale carbon dioxide, water vapor, and energy flux densities. Bull. Amer. Meteor. Soc., 82, 24152434.

    • Search Google Scholar
    • Export Citation
  • Beljaars, A. C. M., Viterbo P. , Miller M. J. , and Betts A. K. , 1996: The anomalous rainfall over the United States during July 1993: Sensitivity to land surface parameterization and soil moisture. Mon. Wea. Rev., 124, 362383.

    • Search Google Scholar
    • Export Citation
  • Bennett, W. B., Wang J. F. , and Bras R. L. , 2008: Estimation of global ground heat flux. J. Hydrometeor., 9, 744759.

  • Betts, A. K., 2004: Understanding hydrometeorology using global models. Bull. Amer. Meteor. Soc., 85, 16731688.

  • Betts, A. K., 2009: Land-surface-atmosphere coupling in observations and models. J. Adv. Model. Earth Syst., 1 (4), 118, doi:10.3894/JAMES.2009.1.4.

    • Search Google Scholar
    • Export Citation
  • Bonan, G. B., Oleson K. W. , Vertenstein M. , Levis S. , Zeng X. , Dai Y. , Dickinson R. E. , and Yang Z.-L. , 2002: The land surface climatology of the community land model coupled to the NCAR community climate model. J. Climate, 15, 31233149.

    • Search Google Scholar
    • Export Citation
  • Bosilovich, M. G., Robertson F. R. , and Chen J. , 2011: Global energy and water budgets in MERRA. J. Climate, 24, 57215739.

  • Bowen, I. S., 1926: The ratio of heat losses by conduction and by evaporation from any water surface. Phys. Rev., 27, 779787.

  • Budyko, M. I., 1958: The heat balance of the earth’s surface. U.S. Department of Commerce, Weather Bureau, 259 pp.

  • Budyko, M. I., 1974: Climate and Life. Academic Press, 508 pp.

  • Chahine, M. T., and Coauthors, 2006: Improving weather forecasting and providing new data on greenhouse gases. Bull. Amer. Meteor. Soc., 87, 911926.

    • Search Google Scholar
    • Export Citation
  • Chen, F., and Coauthors, 1996: Modeling of land surface evaporation by four schemes and comparison with FIFE observations. J. Geophys. Res., 101 (D3), 72517268.