Evaluating runoff simulations from the Community Land Model 4.0 using observations from flux towers and a mountainous watershed



[1] Previous studies using the Community Land Model (CLM) focused on simulating land-atmosphere interactions and water balance on continental to global scales, with limited attention paid to its capability for hydrologic simulations at watershed or regional scales. This study evaluates the performance of CLM 4.0 (CLM4) for hydrologic simulations and explores possible directions of improvement. Specifically, it is found that CLM4 tends to produce unrealistically large temporal variations of runoff for applications at a mountainous catchment in the northwest United States, where subsurface runoff is dominant, as well as at a few flux tower sites spanning a wide range of climate and site conditions in the United States. Runoff simulations from CLM4 can be improved by (1) increasing spatial resolution of the land surface representations and (2) calibrating model parameter values. We also demonstrate that runoff simulations may be improved by implementing alternative runoff generation schemes such as those from the variable infiltration capacity (VIC) model or the TOPMODEL formulations with a more general power law-based transmissivity profile, which will be explored in future studies. This study also highlights the importance of evaluating both energy and water fluxes in the application of land surface models across multiple scales.

1. Introduction

[2] Freshwater resources are vulnerable to climate and land use change, with wide-ranging consequences for human societies and ecosystems [Bates et al., 2008]. To understand and predict water sustainability, analysis and modeling must be performed at a wide range of scales spanning the global and regional hydrologic cycles that drive changes in freshwater resources as well as local and regional scales on which human decisions are made. While much has been learned in the last two decades from studies using off-line hydrologic and water management models with prescribed atmospheric forcings capturing different climate change scenarios, the need to assess different climate change mitigation and adaptation strategies calls for more integrated approaches to modeling both human and Earth system processes across multiple scales [Famiglietti et al., 2010; Janetos et al., 2009; Moss et al., 2010].

[3] This study aims at assessing the potential to apply land surface models that are used to simulate biosphere-atmosphere energy and water exchange in global or regional climate models for the purpose of hydrologic modeling at the local to watershed scales. We selected the Community Land Model (CLM) for this initial effort to identify potential modeling issues that must be addressed when simulating hydrological processes across multiple scales. The focus of this study is on modeling runoff, which is closely related to soil moisture that plays an important role in land-atmosphere interactions in climate models through their controls on water and energy fluxes and streamflows that represent freshwater supply.

[4] The Community Land Model (CLM) is the land component within the Community Earth System Model (CESM) (formerly known as Community Climate System Model (CCSM)) [Collins et al., 2006; Gent et al., 2010; Lawrence et al., 2011]. The CLM has also been tested as the land surface component in an initial effort to develop a regional earth system model based on the Weather Research and Forecasting (WRF) model [Leung et al., 2006]. Rooted from the climate modeling community, the CLM has been designed and used for studies of interannual and interdecadal variability, paleoclimate regimes, and projections of future changes of the global climate system [Gent et al., 2010; Lawrence et al., 2011]. Compared with previous versions, CLM4, the latest version of CLM, represents extensive modifications in its model parameterizations and structure, including runoff generation, soil hydrology thermodynamics, the snow module, and albedo parameters [Lawrence et al., 2011]. In principle, the CLM can be run at any resolution; however, validations were mostly conducted at large river basins, on continental or global scales [Lawrence et al., 2011; Lawrence and Chase, 2007; Niu and Yang, 2006; Niu et al., 2005, 2007; Oleson et al., 2008; Wang et al., 2008]. Limited case studies have been conducted at flux tower sites [e.g., Lawrence et al., 2011; Stöckli et al., 2008; Wang et al., 2008] for validating energy budget simulations and at small watersheds for assessing streamflow simulations using earlier versions of the CLM [e.g., Niu et al., 2005].

[5] This study evaluates the capability of the CLM on runoff simulations at fine spatial scales. More specifically, we investigate how the runoff generation scheme in CLM4 performs at multiple spatial resolutions when compared with observations and with another popular scheme in watershed hydrology, the formulations from the variable infiltration capacity (VIC) model [Liang et al., 1999, 1994, 2003; Cherkauer and Lettenmaier, 2003].

[6] The runoff generation scheme in CLM4 is based on the TOPMODEL formulations [Niu et al., 2005, 2007]. The application of such a scheme is inherently constrained by several factors. First, estimation of the topographic index (TI) from a coarse-resolution digital elevation model (DEM) such as the 1 km resolution DEM [Verdin and Greenlee, 1996] used in CLM4 is problematic. As pointed out by Beven [1997], TIs derived from DEMs with grid cell sizes larger than the length of hillslopes (i.e., typically on the order of 100 m) [Beven, 1982; Troch et al., 2004] would lose their physical information. Therefore, a derivation of TIs using higher-resolution DEMs is highly recommended for applications of the TOPMODEL approach [Sørensen and Seibert, 2007; Wang et al., 2008; Wolock and Price, 1994; Wolock and McCabe, 2000; Zhang and Montgomery, 1994]. Second, a major assumption of the TOPMODEL formulations is the dominant control of the saturation area on runoff generation and hence the topographic control of lateral soil water redistribution [Beven, 1997], which is not always true. For example, in regions with an arid climate, saturated areas seldom exist, so runoff generation is mainly controlled by the interaction between local rainfall intensity and soil infiltration capacity. The assumption of topographically controlled runoff generation is violated in areas with flat terrain, thick soils, or deep groundwater. Instead of topography, runoff generation could be driven by head difference in the groundwater system, microtopography, or perched water tables because of heterogeneity in the soil layers [e.g., Beven, 1997; Maxwell and Miller, 2005; Miguez-Macho et al., 2007].

[7] In comparison, the parameterizations in the VIC model may provide a more general representation of runoff generation under different conditions because the aforementioned assumptions used in TOPMODEL have been significantly relaxed in the VIC formulations. That is, VIC assumes that surface runoff generation is controlled by the mean soil moisture capacity as well as by its spatial heterogeneity over a watershed or a grid cell, and subsurface runoff generation can be approximated by a nonlinear relationship as a function of deep-layer soil moisture [Liang et al., 1994]. These assumptions are generally valid under wide-ranging site or climatic conditions, although the calibration of model parameters for optimal model performance is generally recommended, and there are more parameters that can be calibrated than the original TOPMODEL scheme (see the discussions byHuang and Liang [2006]). Following Wang et al. [2008], who implemented the VIC surface and subsurface runoff parameterizations to CLM3, we updated the code changes to CLM4 in order to evaluate the impacts of the VIC parameterizations in CLM4 that include important changes compared with CLM3, as discussed above.

[8] In order to evaluate the CLM for hydrologic modeling across scales, we introduce a semidistributed implementation of CLM4, called DCLM hereafter, which uses watersheds as the computational units. A common feature of the CLM and most land surface models is their grid-based representation, i.e., the study area is divided into a number of regularly spaced computational grid cells. This type of representation is convenient for coupling with atmospheric models as surface flux exchange between land surface and atmosphere is mainly governed by vertical complexity, and atmospheric models are usually formulated using regular grids. For studies in which streamflow-related issues such as water resources management are of particular concern, the grid-based representation may need higher resolutions to better simulate the lateral redistribution of soil moisture and river routing over the irregular stream networks. Since streamflow is by nature bonded to watershed boundary, the semidistributed approach provides a more effective means for hydrologic modeling on the watershed scale.Koster et al. [2000]tested a catchment-based approach for land surface modeling in a global climate model. More recentlyGoteti et al. [2008]developed the Catchment-Based Hydrologic and Routing Modeling System (CHARMS) that is composed of the CLM and a river routing model that operates on a network of hydrologic catchments or watersheds. Testing of CHARMS over the humid Wabash River basin in the central United States shows good agreement of the simulated streamflow with observations.

[9] This paper evaluates the impacts of runoff parameterizations and spatial resolution on hydrologic simulations. The runoff generation schemes in the CLM4 and VIC model and the implementation of VIC formulations into CLM4 are described in section 2.1. Section 2.2 describes the flux tower sites and watersheds used to evaluate the model simulations. The DCLM implementation of CLM4 is discussed in section 2.3. Section 3 presents the modeling results and discussion. Section 4 summarizes the findings and discusses future directions.

2. Methodology

2.1. CLM and VIC Runoff Generation Schemes

[10] The runoff generation scheme in CLM4 is based on a simplified TOPMODEL-based representation [Niu et al., 2005; Oleson et al., 2010]. Both surface and subsurface runoff are parameterized as exponential functions of the water table depth. The rate of surface runoff generation is given by

display math

where Fsat is the fraction of saturated area within a grid cell, p is the effective rainfall intensity (in mm s−1) (i.e., equivalent to kg m−2 s−1 in common CLM applications), which is estimated as the sum of throughfall (rainfall and dewfall after canopy interception) and snowmelt, I is the soil infiltration capacity (in mm s−1), which is controlled by soil properties and soil moisture within the top soil layer, Fmax is the maximum possible saturated area fraction, and Cs is a coefficient. Both Fmax and Cs can be derived from the distribution function of a logarithmic topographic index defined as ln(a)/tan(β) for each model grid cell, in which a is the contributing area to a point per unit contour length, β is the local surface slope angle. Here fover is a decay factor (m−1) that represents the distribution of saturated hydraulic conductivity with depth below the soil surface and can be estimated by calibration or directly estimation from the recession curve of observed hydrograph. Here z is the depth between the ground surface and the water table (in meters), which represents the mean storage deficit [Beven, 1997; Iorgulescu and Musy, 1997]. The latter is allowed to vary in CLM4 based on the groundwater table scheme. The rate of subsurface runoff generation is given by

display math

where Rsb,max is the maximum subsurface runoff when the whole grid cell is saturated (in mm s−1) and fdrai is again, a decay factor (m−1) that represents the distribution of saturated hydraulic conductivity with depth. The total soil column in CLM4 is divided into 10 layers, with the thickness of each layer increasing from top to bottom. The total soil depth for hydrologic simulation is prescribed a uniform constant value of 3.802 m. For more details about the soil hydrology in CLM4, please refer to Niu et al. [2005, 2007], Oleson et al. [2008, 2010], and Lawrence et al. [2011]. CLM4 has a default set of runoff parameters hardcoded for general applications. These parameters were derived from a sensitivity analysis of global applications. The default values are fover = 0.5 m−1 and Cs = 0.5 for equation (1), and fdrai = 2.5 m−1 and Rsb,max = 5.5 × 10−3 mm s−1 for equation (2) [Oleson et al., 2010].

[11] It is worth mentioning that in traditional TOPMODEL applications and the original SIMTOP formulations in the work by Niu et al. [2005], fover and fdrai share the same physical meaning and value. In CLM4, fover and fdrai have been decoupled, with fover controlling the amount of infiltration into the soil column and fdrai controlling the shape of the recession curve.

[12] As discussed above, the surface and subsurface runoff parameterizations from the VIC model [Liang et al., 1994] have been implemented in CLM4 following the approach described by Wang et al. [2008] and is referred as CLM4VIC. The soil column in the VIC model is typically divided into three layers with variable soil depths. To be consistent with that of CLM4, the total soil depth is also set at 3.802 m in CLM4VIC. The original 10 soil layers in CLM4 are aggregated into three layers consistent with the VIC implementation, with thicknesses of 0.0906, 0.4024, and 3.3090 m, for modeling surface and subsurface runoff. The CLM4 soil layers for thermal calculations are kept intact. In CLM4VIC, surface runoff rate follows the VIC formulation and is given by

display math

where wm is the storage capacity of top two soil layers (in millimeters), w0 is the soil water storage in the top two layers at the beginning of a time step (in millimeters), Δt is the time interval (in seconds), i is the current soil moisture holding capacity (in mm s−1), im is the maximum moist holding capacity (in mm s−1), and b is the shape parameter controlling the spatial distribution of i. Subsurface runoff generation rate is based on the ARNO model [Todini, 1996] concept adopted for the VIC model and given by

display math

where Dsmax is the maximum subsurface flow rate (in mm d−1), Ds is a fraction of Dsmax, w3,m is the storage capacity of the third layer (in millimeters), Ws is a fraction of w3,m, and w3,0 is the soil water stored in the third layer at the beginning of a time step (in millimeters). For more details about the VIC formulations, please refer to Liang et al. [1994, 1999, 2003] and Cherkauer and Lettenmaier [2003].

[13] We stress that in CLM4VIC, only the runoff generation parameterizations of CLM4 have been replaced; representations of all other land surface processes strictly follow what was implemented in CLM4. As documented in VIC literature, the runoff generation parameters [e.g., see Huang and Liang, 2006, Table 2] b, Ws, and Ds vary between 0 and 1, and Dsmax ranges from 0 to 40 mm d−1. Also, in traditional applications of VIC, the soil depths can vary through calibration using streamflow observations. Here we chose to fix these parameters for consistency with how land surface models, such as CLM4, are used in global or regional climate models.

[14] TOPMODEL-basedequations (1) and (2) contain the following assumptions: (1) approximating the saturated zone dynamics by successive steady state representations; (2) assuming a homogeneous recharge rate to the water table over the catchment; (3) approximating the hydraulic gradient of the saturated zone by the local surface slope; and (4) assuming that the distribution of downslope transmissivity is an exponential function of storage deficit or depth to the water table.

[15] The exponential function in equations (1) and (2) was proposed based on observations of the upper soil layers over small watersheds and its attractive analytical expression [Beven, 1984, 1997]. It has been shown that such a function could have profound impacts on the shape of the simulated subsurface flow recession curves [Ambroise et al., 1996] and is oversimplified for representing the nonlinear relationship between regional subsurface flow and the storage term (e.g., aquifer depth) evident in empirical (e.g., recession curve analyses) and theoretical studies [Brutsaert and Lopez, 1998; Eltahir and Yeh, 1999; Marani et al., 2001; Tallaksen, 1995; Zecharias and Brutsaert, 1988]. In the past decades, progress has been made to generalize and relax the assumptions of the TOPMODEL formulations. Duan and Miller [1997] and Iorgulescu and Musy [1997] proposed a generalized power function to replace the exponential function in equations (1) and (2); Woods et al. [Woods, 1997; Woods et al., 1997] introduced a subsurface flow formulation for use in regions with rapid and spatially variable subsurface runoff. Huang et al. [2008] proposed a new subsurface flow formulation that incorporates spatial variability of both topography and recharge and the power law transmissivity profile. They demonstrated through theoretical derivation and case studies that the power law and exponential subsurface flow parameterizations in TOPMODEL, as well as the parameterization proposed by Woods et al. [Woods, 1997; Woods et al., 1997], are all special cases of the new formulation.

[16] Although runoff simulations from the CLM could be potentially improved by using the new TOPMODEL formulations, the extensive data processing and analyses needed to implement the formulations present practical difficulties. Specifically, the generalized power form of TOPMODEL [Duan and Miller, 1997; Iorgulescu and Musy, 1997] would replace the exponential term in equation (2) with a term of power n, and the corresponding topographic index to be used for calculating Fmax and Cs becomes a power function as (a/tan β)1/n. In TOPMODEL applications, nis first determined by separating the observed streamflow into surface and subsurface runoff components and conducting recession curve analysis. Then DEMs need to be processed to derive the power law-based topographic indices. For formulations proposed byWoods et al. [1997] and Huang et al. [2008], the DEM processing procedures are even more complex. Interested readers should refer to these papers for details. We are not aware of any processed power law-based topographic indices globally or regionally (e.g., United States) for the reasons provided above. The HYDRO1K data set used for derivingFmax and Cs by Niu et al. [2005] contains only the logarithmic topographic indices derived from DEMs at a 1 km resolution.

[17] Currently, hydrologically conditioned DEMs are available globally at fine resolutions (e.g., the HydroSHEDS data set at http://hydrosheds.cr.usgs.gov/). Analyzing the high-resolution DEMs, in combination with the improved formulations, will have the potential to improve the TOPMODEL-based runoff generation parameterizations in CLM4. However, evaluating and implementing the new TOPMODEL formulations would require substantial efforts that are beyond the scope of this paper. In addition, these improvements do not address the key TOPMODEL assumption that the hydraulic gradient of the saturated zone can be approximated by the local surface slope, which is often violated in regions of low relief or thick soils. For this reason, we also explore use of the more general VIC hydrologic model. In this paper, we discuss the implications of both original TOPMODEL and VIC runoff generation schemes based on results from the case studies described below.

2.2. Study Area and Experiment Design

[18] To evaluate the impacts of runoff parameterizations and spatial resolution, CLM4 and CLM4VIC are applied to a watershed and five selected flux tower sites where observations are available to evaluate the model. To contrast with Goteti et al. [2008], who evaluated the catchment-based approach over a relatively flat and humid river basin in the central United States, we selected a mountain watershed with large topographic gradients in a snowmelt-dominated basin to highlight potential challenges that must be addressed for hydrologic modeling across multiple scales. Application of the models over the flux tower sites allows impacts of runoff parameterizations to be assessed over a wider range of site and climate conditions.

2.2.1. The American River Watershed

[19] The American River watershed is a snow-dominated mountainous watershed located along the leeside of Mt. Rainier in the Pacific Northwest region of the United States, as shown inFigure 1. The total drainage area of the American River watershed is 205 km2, which is mainly covered by evergreen forest and shrub and underlaid by sandy soil. The elevation ranges from 850 to 2100 m. This watershed is small enough so that no river routing is needed for comparison of observed and simulated streamflows on daily time scales. Annual precipitation in the watershed is about 1850 mm, mostly occurring in the winter season (November to January) as snowfall. This is a watershed in which all the TOPMODEL assumptions hold. Using a fully Distributed Hydrology Soil Vegetation Model (DHSVM) at a 100 m grid resolution, Leung and Wigmosta [1999] realistically simulated the streamflow of this mountain watershed compared with the observed hydrograph.

Figure 1.

Map of the American River watershed. The spatial discretization has been done at three levels: Level 1 treats the watershed as a lumped unit, as indicated by the dark line; level 2 divides the watershed into three “big” subwatersheds, as indicated by the blue lines; level 3 divides the watershed into 33 “small” subwatersheds, as indicated by the red lines.

[20] In order to investigate the effects of spatial heterogeneity on hydrologic responses, the watershed is discretized at three levels based on a 30 m DEM from the U. S. Geological Survey (USGS). Level 1 treats the whole American River watershed as a lumped unit, level 2 divides it into three “big” subwatersheds with an averaged drainage area of ∼68 km2, and level 3 divides it into 33 “small” subwatersheds with an averaged drainage area of ∼6 km2, as shown in Figure 1.

[21] Hourly precipitation and surface temperature data for the period 1 October 2003 to 30 September 2010 were spatially interpolated based on elevation and lapse rates from observations at two snow telemetry (SNOTEL) stations: the Morse Lake station located within the American River watershed and the Bumping Ridge station located nearby (Figure 1). Hourly solar radiation data were generated from a meteorological model [Waichler and Wigmosta, 2003] and assumed spatial uniformity over the entire watershed. Observed daily streamflow data were obtained from a gauge station (American River near Nile, Washington: USGS12488500) at the watershed outlet (Figure 1). Land cover data were derived from the 2001 National Land Cover Database, which has a 30 m spatial resolution [Multi-Resolution Land Characteristics Consortium, 2001]. Vegetation parameters such as leaf and stem area indices were obtained from the Moderate Resolution Imaging Spectrometer (MODIS) 1 km land products [Friedl et al., 2002; Myneni et al., 2002]. Soil properties such as porosity and texture were extracted from the continental United States (CONUS) soils' database [Miller and White, 1998]. Both CLM4 and CLM4VIC are applied to all three levels of watershed delineation to evaluate the impacts of runoff parameterizations as well as model resolution.

2.2.2. Flux Tower Sites

[22] To further evaluate sensitivity of simulated runoff to model parameters, we selected five flux tower sites assembled by the site synthesis team of the North American Carbon Program (NACP). These sites were selected because they span a wide range of climate and site conditions, as listed in Table 1. U.S.-ARM and U.S.-Ne3 are located in Oklahoma and Nebraska, respectively, and are covered by cropland without irrigation [Fischer et al., 2007; Riley et al., 2009; Suyker and Verma, 2009]. U.S.-Ton, U.S.-Ha1, and U.S.-Dk3 are located in northern California, Massachusetts, and North Carolina, respectively, with biome types of woody savannas (∼50% oak trees and 50% grass), broadleaf deciduous forest, and needleleaf evergreen forest (i.e., loblolly pines), respectively [Baldocchi et al., 2004; Goulden et al., 1996; Ma et al., 2007; Sun et al., 2010; Urbanski et al., 2007].

Table 1. Site Information of the Five Selected NACP Towers
Site NameLongitudeLatitudeMean Precipitation (mm)Mean Temperature (°C)FmaxLand CoverPercent ClayPercent SandΔt (min)Simulation Period
U.S.-Ha1−72.171542.5378880.747.870.34Deciduous Broadleaf6.0066.00601991–2006
U.S.-Dk3−79.094235.97821041.6414.650.38Evergreen Needleleaf13.6651.59301998–2005
U.S.-Ton−120.966038.4316545.1516.320.38Woody Savannas15.0041.00302001–2007

[23] For each site, meteorological forcing, site information such as soil texture, vegetation cover, and satellite-derived phenology, as well as validation data sets such as water and energy fluxes are provided by the NACP site synthesis team. The models were spun up by cycling the provided forcing for at least three times until all the state variables reached equilibrium (e.g., the forcing data span a six year period from 2001 to 2006 at U.S.-Ne3. The model was spun up by cycling the six year data three times, equivalent to 18 years, which is the minimum number of spin-up years over the five sites). Although data availability varies across sites, the site with the shortest data coverage, U.S.-Ne3, includes at least three years. A detailed description of the NACP site synthesis data set is provided bySchwalm et al. [2010]. For each site, three sets of numerical experiments were conducted, namely, CLM4D, CLM4F, and CLM4VIC, with one control and five sensitivity experiments (i.e., CON and SEN1–5, respectively, hereafter) within each set.

[24] Configurations of the numerical experiments are given in Table 2. For the CLM4D set, the CON experiment used default parameter values given in the CLM4 official release (see Fmax extracted from the global data set in Table 1 and the default runoff parameter values listed in section 2.1). The SEN experiments were designed to perturb the subsurface flow generation parameters one at a time. The default value of fover (i.e., 0.5) is used in CLM4D simulations. Such a value would result in a larger saturated fraction at the land surface and therefore more surface runoff and less infiltration. One might suspect that if fover is raised to a higher value, more water will infiltrate into the soil so that the model could produce subsurface runoff over flat terrain. Therefore, in CLM4F simulations, we increased fover to 2.0 and repeated all the experiments. The CLM4F simulations were designed to evaluate the potential impact of surface runoff generation parameters on subsurface flow. For CLM4VIC simulations, we assumed that spatial heterogeneity and topographic relief at the flux tower sites is small, so that b and Dsmax are small (i.e., b = 0.1, Ws = 0.5, Ds = 0.05, Dsmax = 2 mm d−1). For all three sets of experiments, other model parameters were fixed, but values of the two parameters that control the subsurface flow generation in each model were adjusted to evaluate the impacts of model structure and parameter values on simulated streamflow (Table 2).

Table 2. Control and Sensitivity Experiments Conducted at the Flux Tower Sites
ExperimentsCLM4D and CLM4FCLM4VIC
Rsb,max (mm s−1)fdrai (m−1)Dsmax (mm d−1)Ws
CON5.5 × 10−
SEN15.5 × 10−42.547.500.5
SEN21.0 × 10−42.58.640.5
SEN35.5 × 10−348.640.7
SEN45.5 × 10−382.00.7
SEN55.5 × 10−3147.500.7

2.2.3. A Semidistributed Implementation of CLM

[25] A distinct feature of this work compared with previous studies with the CLM is that we applied CLM4 and CLM4VIC at the American River watershed using subwatersheds as the fundamental spatial units instead of grid cells. To take advantage of multiprocessor computers, the subwatersheds were organized into a pseudo-grid form with the input data sets arranged and read by the model as if each subwatershed is an individual model grid cell. Interactions between the subwatersheds could be mediated through a realistic river network. For application at the American River watershed, however, channel routing is not included because the residence time of runoff within the river network of the small watershed is much less than one day, the finest temporal resolution of observed streamflow available for public access. To compare with the observed daily streamflow at the outlet, the simulated surface and subsurface runoff from all subwatersheds is aggregated to obtain the simulated daily streamflow.

3. Results and Discussion

3.1. Results From the American River Watershed

[26] Runoff generation in the American River watershed is, by and large, snow dominated. Hence, it is important to evaluate the performance of the snow module before assessing the impacts of the runoff parameterizations. We first applied CLM4 in a single-point mode to the Morse Lake SNOTEL station, where observed meteorological forcing and daily snow water equivalent (SWE) data are available for the study period. Land cover at the site was set to be 100% grassland, and soil properties were assumed to be the same as those of a subwatershed from the level 3 discretization, to be consistent with the site condition of the SNOTEL station. Compared with observations, the model underestimated the accumulation during 2005–2010 (as shown inFigure 2), possibly because of the errors related to the meteorological forcing or observations. Nevertheless, the simulated SWE compares well enough (the root-mean-square error (RMSE) is 233 mm andR2 is 0.88) with observations to provide the basis for comparing the observed and simulated streamflows.

Figure 2.

Simulation of daily snow water equivalent (SWE) at the SNOTEL station.

[27] We first applied CLM4 with the default parameter values to the American River watershed. The simulated monthly streamflow time series at the three levels of discretization are shown in Figure 3. It is interesting to note that the subsurface runoff completely dominates over the surface runoff. This is not surprising, given the large fraction of sandy soil within this watershed, and is consistent with simulations using the Distributed Hydrology Soil Vegetation Model (DHSVM) for the same watershed [Leung and Wigmosta, 1999]. A numerical base flow separation procedure by Arnold et al. [1995]has also been applied to the observed streamflow, resulting in a mean base flow to streamflow ratio of ∼0.7, which confirms to some degree the dominance of subsurface flow. However, we caution the use of such an empirical technique in snow-dominated areas because snowmelt is a relatively slow and gradual process; numerical base flow separation techniques are usually not capable of separating the contribution to surface runoff from snowmelt entering into the channel from the subsurface runoff [Sinclair and Pitz, 1999; Wolock, 2003]. Figure 3 shows that the simulated total runoff is overestimated in spring and early summer. In May, the mean monthly runoff is overestimated by about 30%, and the simulated streamflow generally peaks about a month earlier than the observed mean monthly runoff. The model performance improves from level 1 to level 3, as shown in Figure 3 (and more obviously shown in Figures 4, 5, and 6), where the overestimation of runoff is reduced, especially in April to June, and the timing of runoff matches better with observations. From level 1 to level 3, the watershed is divided into more subwatersheds, so the spatial heterogeneity of both climatic forcing and landscape properties is more explicitly represented in the model. This helps to better capture the temporal variability that is due to the strong interaction between spatial and temporal variability of hydrological processes.

Figure 3.

Seasonal variations of hydrological responses predicted by DCLM with the default parameters. Green dots are the observed mean monthly streamflow. Blue lines are the simulations from the lumped watershed (i.e., level 1). Red dashed lines are the averaged simulations over the three big subwatersheds. Yellow dashed lines are the simulations averaged over the 33 small subwatersheds.

Figure 4.

Taylor diagram of daily streamflow simulations with the observed daily streamflow as reference.

Figure 5.

Simulated monthly total runoff from (a) DCLM4 using default parameter values, (b) DCLM4 using calibrated parameter values, (c) DCLM4VIC using “default” parameter values, and (d) DCLM4VIC using calibrated parameter values.

Figure 6.

Similar to Figure 5, but for subsurface runoff simulations. Note that the observed streamflow is included here for comparison (see section 3.1 for details).

[28] The default parameter values derived from global applications are not necessarily suitable for local or regional applications. In CLM4, fover, fdrai, Cs, and Rsb,max in equations (1) and (2) have significant impacts on the water table depth and soil moisture profile and thus on the temporal variation of runoff generation. To determine the sensitivity of the streamflow simulations to model parameters, a manual calibration was conducted, aimed at improving the seasonal variation of the total runoff. As discussed by Niu et al. [2005], soil hydrology parameters can affect the shape of the hydrograph, but they have no significant control on the annual water balance when the evapotranspiration (ET) process is dominated by the meteorological forcing, i.e., ET is energy limited rather than limited by the available soil water storage. It appears that at such a basin the annual water balance is mostly sensitive to the meteorological forcing such as precipitation, solar radiation, and so on, based on simulations from previous versions of CLM, which are consistent with our findings from this study based on simulations from CLM4. For simplicity, calibration was conducted for the lumped mode only, and then the same set of parameters were applied to the other two levels of spatial discretization, i.e., three big subwatersheds and 33 small subwatersheds. After calibration, we obtained fover = fdrai = 1.0 m−1 and Cs = 0.5 for equation (1) and Rsb,max = 2.6 × 10−3 mm s−1 for equation (2). The daily observed and simulated streamflow time series are compared and summarized by the Taylor diagram [Taylor, 2001] shown in Figure 4. Roughly, the closer the points of model simulations are to the standard radius (dashed line), the better the timing (or phase) has been captured. If drawing a line from a point of model simulation to the origin, the smaller the angle between this line and horizontal line, the better the amplitude of variation has been captured. If the point of model simulation overlaps with the intersection of the standard radius line and horizontal line, the timing and amplitude are perfectly reproduced by the model simulation. Figure 4 shows that the runoff simulations have notably improved in terms of both timing and amplitude after calibration. Because of the dominance of subsurface runoff, the calibration has focused more on the variation of subsurface runoff generation. As shown in Figures 5 and 6, the CLM with the default parameter values tends to produce unrealistically high peaks of subsurface runoff. In order to better illustrate the effect of calibration on subsurface runoff prediction and thus the total runoff, in Figure 6 the simulated subsurface runoff is plotted against the observed streamflow, which is used as a surrogate of the actual total runoff. One would expect that the subsurface runoff from a reasonable prediction should always be lower than the total runoff according to mass conservation. Any subsurface runoff peaks higher than the total runoff peaks are obviously unrealistic. Figure 6 shows that with the default parameters, CLM4 tends to produce unrealistically high peaks of subsurface runoff. The reason for these high peaks of subsurface runoff has been briefly discussed in section 2.1 and will be further investigated in section 3.2. Clearly, the major effect of calibration is to bring down the unrealistically high peaks of subsurface runoff.

[29] CLM4VIC was also applied to the American River watershed at all three levels of discretization. In order to demonstrate the effects of parameter calibration for comparison with the CLM4 simulations, two sets of parameter values were used to conduct CLM4VIC simulations at this watershed. However, we stress that a “default” parameter set is uncommon in the VIC model because parameter calibration is always recommended for any VIC application. A default parameter set was chosen for comparison purposes only. To be consistent with the flux tower application, we kept Ws = 0.5, Ds = 0.05, Dsmax = 2 mm d−1 in equation (4), but adjusted b to be 0.4 based on the large spatial heterogeneity in the watershed. For comparison, another set of parameters taken from an application of the VIC model at the Columbia River basin (b = 0.4, Ws = 0.7, Ds = 0.05, Dsmax = 8 mm d−1) from the University of Washington (http://www.hydro.washington.edu/Lettenmaier/Models/VIC/) is used as the calibrated parameter set. Figures 4, 5, and 6 show that parameter calibration also improves the timing and amplitude of the simulated streamflow by improving the subsurface runoff simulation as in CLM4. With the default parameter set, CLM4VIC underestimated subsurface runoff because the default parameters were suitable only for relatively homogeneous soil properties and flat topography. With the “calibration” parameter set, CLM4VIC was able to produce reasonably realistic subsurface runoff.

[30] One important common feature between the CLM4 and CLM4VIC simulations (without parameter calibration) is the apparent attenuation of subsurface runoff from the level 1 to level 3 discretization. By decreasing the size of the subwatersheds, the model is able to better incorporate the spatial heterogeneity in climate forcing (e.g., precipitation and temperature) and land properties (soil and vegetation properties), leading to more diverse timing of runoff generation from different subwatersheds and hence the temporal attenuation of the runoff generation when aggregated to the whole watershed.

3.2. Results From the Flux Tower Applications

[31] Simulated monthly runoff components from CLM4D, CLM4F, and CLM4VIC at the selected flux tower sites are shown in Figures 7 and 8. It is evident from Figures 7 and 8 that the models differ substantially in how surface and subsurface runoff is generated across the sites:

Figure 7.

Simulated runoff components at the flux tower sites. Warm-colored lines are simulations from CLM4VIC, and cold-colored lines are simulations from CLM4D, respectively.

Figure 8.

Same as Figure 7, but with cold-colored simulations from CLM4F.

[32] 1. At the tower sites where subsurface flow is significant from CLM4D and CLM4F (i.e., U.S.-Ha1, U.S.-Dk3, and U.S.-Ton), it is a common feature that the subsurface hydrograph shows high peaks and short recession periods. ReducingRsb,max did not alleviate the problem. For example, in SEN2 and SEN3, Rsb,max was reduced to 5.5 × 10−4 and 1.0 × 10−4 mm s−1, respectively, equivalent to 47.5 and 8.64 mm d−1, respectively, which is comparable to the values used in CLM4VIC. However, the use of equation (2) for subsurface runoff generation causes a quick release of water into the streams when the storage term is close to one (i.e., when the groundwater table approaches the surface) and limits subsurface flow when the storage term gets smaller after the initial release as a result of the rigidity of the exponential function.

[33] In contrast, CLM4VIC simulated low peaks and long recession periods at these sites, which is more consistent with the definition of subsurface flow as the slow varying component of streamflow. When Dsmax is increased to 47.5 mm d−1 (i.e., SEN1 and SEN5), subsurface flow simulated by CLM4VIC starts to show high peaks at magnitudes comparable to that produced by CLM4. However, the long recession periods remained. Varying Ws affects the shape of the recession curve in a more flexible way compared with the exponential function in equation (2).

[34] 2. The simulated subsurface flows from CLM4D at U.S.-ARM and U.S.-Ne3 are close to zero for the control and all the sensitivity experiments except for SEN5. Note thatRsb,max in equation (2) and Dsmax in equation (4) share the same physical meaning [Liang et al., 1994; Niu et al., 2005]: the maximum regional subsurface flow. Dsmax in CON, SEN3, SEN4, and SEN5 (i.e., 2.00, 8.64, 2.00, and 47.50 mm d−1, respectively) is small compared with Rsb,max, which is equivalent to 475.2 mm d−1. It appears that the topographic-driven flow assumption failed at these sites so that no subsurface flow was generated from CLM4, while a significant amount of subsurface runoff was simulated by CLM4VIC. It has been reported that, on average, subsurface flow accounts for 30%–40% of the total streamflow for the watershed in which U.S.-ARM is located (i.e., Salt Fork of the Arkansas River) during the period of the 1940s to present [Esralew and Lewis, 2010]. At major rivers in Nebraska, including the lower Platte River watershed where the U.S.-Ne3 site is located, subsurface flow accounts for more than 50% of the total streamflow because of the High Plain aquifer beneath the state [Stanton et al., 2010].

[35] In comparison, Figure 8 shows that with a higher fovervalue, surface runoffs are significantly reduced from CLM4, and the sharpness of subsurface runoff peaks becomes more pronounced because of higher saturation levels of the soil column across all sites. At U.S.-ARM and U.S.-Ne3, where subsurface flow was obviously underestimated in the CLM4 simulations, different responses are observed. At U.S.-ARM, the model starts to produce subsurface flow when a higherfovervalue is used, but with sharp peaks and a short recession period, consistent with the patterns simulated at other sites. At U.S.-Ne3, the simulated subsurface runoff remains small but becomes noticeable. It is expected that iffover is further increased, we will start to see more subsurface runoff at this site, but the subsurface runoff would share the same characteristics of short recession period and high peaks. Furthermore, the increase in subsurface flow will be compensated by a more significant reduction in surface runoff.

[36] We note that m = 1/fdrai represents the effective storage capacity, which controls the exponential decline of transmissivity with depth (see discussions in work by Iorgulescu and Musy [1997], Beven [1997] and Kirkby [1997]). In TOPMODEL applications, m has been shown to typically range between 0.001 and 0.1 m under various site and climate conditions [Beven, 1997]. Hence, when fdrai = 1 (i.e., m = 1) in equation (2)(e.g., SEN5), the TOPMODEL-based runoff generation scheme in CLM4 is pushed to an extreme. Also notice again that theRsb,max value used in SEN5 is unrealistically high. Under such circumstances, CLM4 begins to generate some subsurface flow (i.e., the dark green lines in Figures 7b, 7j, and 8j), although such an extreme parameter value results in a significant reduction of surface runoff generation.

[37] 3. The assumptions and structure of the TOPMODEL and VIC formulations have profound impacts on runoff generation. TOPMODEL assumes that the saturation excess runoff (Dunne runoff) dominates runoff generation. That is, the water table needs to rise to the surface before surface runoff could be generated. Although a term to represent infiltration excess runoff (Horton runoff) was introduced for the unsaturated portion of the grid cell by Niu et al. [Niu et al., 2005; Oleson et al., 2010], the overall status of saturation of the soil column is still a controlling factor for both surface and subsurface runoff generation. As a consequence, from Figures 7 and 8, different combinations of parameter values for subsurface flow in CLM4 also noticeably affected surface runoff generation. In other words, surface and subsurface runoff generation mechanisms are closely coupled in the TOPMODEL framework. This close coupling is problematic when the groundwater table is sufficiently deep and has limited impacts on the surface soil layer.

[38] In the VIC model and CLM4VIC, it is assumed the first two soil layers control surface runoff generation and the third soil layer controls subsurface runoff generation. Therefore, propagation of the infiltration front, rather than the overall saturation status of the soil column, determines the generation of surface runoff when the soil column is not saturated. This is why, despite the large difference in subsurface runoff produced by varying the parameter values in equation (4), limited impact is found in the simulated surface runoff across the sites in Figures 7 and 8. Under wet climate conditions, however, when the entire soil column becomes saturated, surface and subsurface runoff generation processes start to be coupled in VIC formulations [Li et al., 2011; Li and Sivapalan, 2011].

3.3. Impact of Runoff Parameters on Energy Fluxes

[39] Since water and energy balances are tightly coupled, it is important to determine if the large differences in soil hydrology in CLM4 and CLM4VIC influence the simulated energy fluxes. The water and energy budgets are mainly connected through evapotranspiration or the latent heat flux. Evapotranspiration consists of three components: canopy evaporation, vegetation transpiration, and ground evaporation.

[40] Figure 9 shows some differences between the latent heat fluxes from the different model simulations at the American River watershed. More detailed analyses (figure not shown) reveal that the simulations of canopy evaporation and vegetation transpiration are almost identical across the simulations. The ground evaporation, which is controlled by surface soil moisture, is lower in the CLM4 simulations (without calibration) than in CLM4VIC. The difference of the latent heat component that is due to ground evaporation is nearly compensated by the difference of sensible heat from different simulations. This results in very similar net radiations from the different simulations. Generally, as CLM4 and CLM4VIC produced different soil water vertical distributions, the partitioning between sensible heat and latent heat is different. This is consistent with the study by Leung et al. [2011]. However, in an energy-limited basin such as that of the American River, the difference in latent heat flux leads to only small differences in the net radiation, despite much larger differences in the water fluxes simulated by the two modeling approaches. It is therefore critical to examine both the energy and water balance across various spatiotemporal scales in evaluating land surface models for use in both climate models and hydrologic applications.

Figure 9.

Comparison between the energy fluxes predictions between CLM4 and CLM4VIC, both without calibration.

[41] Figure 10 illustrates the impacts of different runoff generation schemes and parameter values on the energy flux simulations based on CLM4D and CLM4VIC simulations at the flux towers:

Figure 10.

Simulated sensible and latent heat fluxes at the flux tower sites.

[42] 1. Both CLM4D and CLM4VIC simulate the latent heat and sensible heat reasonably well at all sites. The underestimation of latent heat flux and overestimation of sensible heat flux at U.S.-Dk3 is mainly due to uncertainty in the biophysical parameters of the specific type of evergreen needleleaf trees at the site: loblolly pines (LPs). It has been found that LPs tend to produce lower sensible heat flux and higher latent heat flux, possibly because of their lower albedo values compared with those of other evergreen needleleaf trees [Sun et al., 2010]. Adding a plant functional type (PFT) representative of LP in CLM will be the key to improving energy budget simulations at the site.

[43] 2. The difference in CLM4D and CLM4VIC simulated energy fluxes is more pronounced at sites such as U.S.-ARM, U.S.-Ton, and U.S.-Ne3, with shallow-rooted vegetation covers. This shows that the VIC formulations can affect energy partitioning at the surface by altering soil moisture stress in the shallow soil layers. Indeed, the VIC surface runoff formulation (i.e.,equation (3)) generates a more dynamic saturated fraction over the land surface and therefore more dynamic variations in the shallow layer soil moisture in general (figures not shown) compared with those of CLM4.

[44] 3. By varying parameter values in the runoff generation schemes in CLM4 and CLM4VIC, the simulated energy fluxes are again affected mainly at sites with shallow-rooted vegetation. However, the impacts that are due to model parameters are not as significant as those introduced by differences in model structure and formulations between CLM4 and CLM4VIC.

[45] 4. The difference between simulated energy fluxes from CLM4D and CLM4F is relatively small (therefore, figures are not shown) compared with the difference in runoff simulations. Noting that the difference between CLM4D and CLM4F is associated only with the values of fover, which controls the amount of infiltration, this indicates that most of the tower sites are more energy limited rather than water limited. The largest difference is observed at U.S.-Ton, which is a water-limited system. Summertime latent heat fluxes are obviously increased in the CLM4F simulations compared with those from CLM4D. This suggests thatfoveris an important tuning parameter for both water and energy fluxes, especially for water-limited ecosystems.

4. Conclusion and Future Work

[46] Simulations have been performed to evaluate the potential of the Community Land Model (CLM) for land surface and hydrologic modeling across multiple scales. More specifically, we compared the TOPMODEL-based and VIC surface and subsurface runoff parameterizations and tested the implementation of a semidistributed approach of applying CLM and its sensitivity to spatial resolution. A comparison of simulations in a snow-dominated mountain watershed and different flux tower sites shows that the streamflow responses from CLM4 and CLM4VIC are distinctly different and determined largely by model structures and the underlying model assumptions.

[47] We demonstrate through case studies and theoretical discussions that the TOPMODEL-based subsurface parameterization as currently implemented in CLM4, although it considers spatial heterogeneity in soil hydraulic properties and topography, suffers from limitations, as follows:

[48] 1. Problems associated with the estimation of the topographic index (TI) from the coarse-resolution digital elevation model (DEM) at the 1 km resolution. As pointed out byBeven [1997], TIs derived from DEMs with grid cell sizes larger than the length of hillslopes (i.e., typically of the order of 100 m) [Beven, 1982; Troch et al., 2004] would lose their physical information. Therefore, a derivation of TIs using higher-resolution DEMs is highly recommended for applications of the TOPMODEL approach [Sørensen and Seibert, 2007; Wang et al., 2008; Wolock and Price, 1994; Wolock and McCabe, 2000; Zhang and Montgomery, 1994].

[49] 2. Failure of the site conditions, under many circumstances, to meet the key assumption of the TOPMODEL formulations. That is, assumption of the dominant control of saturation area on runoff generation and hence the topographic control of lateral soil water redistribution [Beven, 1997] is not valid at many sites. For example, in regions with an arid climate, saturated areas seldom exist, so runoff generation is mainly controlled by the interaction between local rainfall intensity and soil infiltration capacity. The assumption of topographically controlled runoff generation is also violated in areas with flat terrain, thick soils, or deep groundwater. Instead of topography, runoff generation could be driven by head difference in the groundwater system, microtopography, or perched water tables because of heterogeneity in the soil layers [Beven, 1997; Maxwell and Miller, 2005; Miguez-Macho et al., 2007].

[50] 3. Unrealistic behavior of subsurface runoff responses that is due to the exponential form of the surface runoff parameterization in CLM4. The TOPMODEL-based exponential function was proposed based on observations of the upper soil layers over small watersheds and its attractive analytical expression [Beven, 1984, 1997]. It has been shown that such a function could have profound impacts on the shape of the simulated subsurface flow recession curves [Ambroise et al., 1996] and is oversimplified for representing the nonlinear relationship between regional subsurface flow and the storage term (e.g., aquifer depth) evident in empirical (e.g., recession curve analyses) and theoretical studies [Brutsaert and Lopez, 1998; Eltahir and Yeh, 1999; Marani et al., 2001; Tallaksen, 1995; Zecharias and Brutsaert, 1988].

[51] The key to improving the TOPMODEL-based runoff generation scheme in CLM4 is to relax the embedded assumptions. This could be achieved by using the power law transmissivity profile [Duan and Miller, 1997; Huang et al., 2008; Iorgulescu and Musy, 1997] and a finer-resolution DEM (e.g., the 90 m Digital Elevation Data from the NASA Shuttle Radar Topographic Mission (SRTM) or the 30 m ASTER-based Global Digital Elevation Map) to reduce the bias in estimatingFmax and Cs in equation (1).

[52] On the other hand, the parameterizations in the VIC model might provide a more general representation of runoff generation under different conditions because the aforementioned assumptions used in TOPMODEL have been significantly relaxed. The VIC runoff scheme implicitly captures the subgrid variability of land surface properties that has an impact on both runoff generation and evapotranspiration through the Xinanjiang distribution of infiltration or storage capacity, which enables it to mimic surface runoff generation, including variable contributing area dynamics [Liang et al., 1994]. Another important module of the VIC runoff scheme is the subsurface runoff generation, which is parameterized in term of soil water storage. In many parts of the world, especially steep, forested (or previously forested) landscapes, subsurface flow is an important component of the water balance. However, the spatially varying landscape information such as soil properties and topography has not been explicitly incorporated into the VIC formulation of subsurface runoff, which prevents a priori estimation of the corresponding parameters.

[53] Therefore, it is crucial to (1) evaluate the theoretical basis of current subsurface runoff parameterization for improvements to be implemented in the CLM, and (2) provide a practical approach for a priori estimations of parameters in the improved parameterization based on the improved knowledge from the existing data in real river basins.

[54] Regardless of the surface and subsurface runoff parameterizations used, successful applications on large scales require model calibration, which was shown to improve both CLM4 and CLM4VIC simulations in the American River watershed. Effective model calibration requires strategies of estimating parameter values (e.g., application of efficient calibration tools) [Duan et al., 1992; Gupta et al., 1998; Vrugt et al., 2003], regionalization of parameter values by relating them to physical characteristics of the watersheds or grid cells [Abdulla and Lettenmaier, 1997a, 1997b; Huang et al., 2003], or potential ways of reducing the number of parameters [Huang and Liang, 2006; Huang et al., 2008].

[55] Although model calibration is important for achieving realistic simulations of runoff, most land surface schemes used in climate models use a set of default or spatially uniform model parameters. Our analyses show that differences in runoff simulations that are due to model structure or parameters could potentially have important effects, not only on runoff and streamflow, but they can also influence the surface energy budgets, especially in areas with short vegetation. As different runoff formulations and parameter values can have an important influence on surface energy partitioning, they can alter the strength of land-atmosphere coupling represented by the models.Dirmeyer et al. [2006]compared the land-atmosphere coupling strength estimated based on observation and climate model simulations and found substantial differences among models and between models and observations. Such differences could lead to wide-ranging land surface feedbacks and temperature and precipitation responses to greenhouse warming [e.g.,Seneviratne et al., 2006]. In addition, because of the shallow-rooting profiles, short vegetation ecosystems may exhibit higher sensitivity to climate that modulates water availability in the soil. An accurate simulation of the runoff components, and therefore soil moisture dynamics, will be important for assessing the resilience of these ecosystems to climate variability and change.

[56] The effects of spatial heterogeneity on runoff processes have also been assessed in this study based on the American River watershed. Within the CLM modeling framework, spatial heterogeneity has significant effects on runoff even at this small watershed. Spatial heterogeneity in the American River is dominated by topography, which has large impacts on the atmospheric forcing, including temperature and precipitation. Such first-order effects must be incorporated effectively for the model to be applied across multiple scales, regionally and globally, as mountains provide disproportionate amounts of freshwater supply worldwide. The VIC model includes subgrid snow bands to capture the topographic effects on snowpack, which provides a dominant control on runoff timing through partitioning of snowfall or rainfall and snow accumulation or snowmelt. Currently the CLM includes subgrid plant functional types (PFTs) to represent spatial heterogeneity. This subgrid structure could be extended to include subgrid representations of topography by introducing additional subgrid elevation bands or by correlating subgrid elevation with subgrid PFTs so that each subgrid PFT is associated with a different surface elevation, as tested byLeung and Ghan [1998]. It is important, however, to assess how well the latter approach may work at different grid scales globally across different mountains and ecosystems.

[57] Last, comparison of CLM4 and CLM4VIC requires further efforts to evaluate the model performance across multiple scales and different climate and hydrologic regimes. Work is underway to implement the groundwater module developed by Liang et al. [2003]and tested in a coupled land-atmosphere model byLeung et al. [2011] to CLM4VIC so that the model has soil hydrologic modeling capability comparable to that of CLM4, which includes a representation of the groundwater aquifer. This way the models can be evaluated more consistently across climate and hydrologic regimes where groundwater table dynamics play a different role. In addition, efforts are being undertaken to evaluate the new modifications globally and regionally at various resolutions. We will compare simulated streamflows and other water and energy fluxes from different versions of CLM4 against observations such as those from the Model Parameter Estimation Experiment (MOPEX) database (see http://www.weather.gov/oh/mopex/mo_datasets.htm for more information), and simulations from other land surface model such as those from the North American Land Data Assimilation System (NLDAS: http://ldas.gsfc.nasa.gov/nldas) and the Global Land Data Assimilation System (GLDAS: http://ldas.gsfc.nasa.gov/gldas).


[58] This work is supported by the PNNL Integrated Regional Earth System Modeling (iRESM) Initiative and DOE projects on “Investigation of the Magnitudes and Probabilities of Abrupt Climate Transitions (IMPACTS)” and “Strengthening the Coupling between Climate and Earth System Models (ESMs) and Integrated Assessment Models (IAMs).” The authors would like to thank Teklu K. Tesfa for his help on data analysis at the beginning stage of this study, R. Stockli for sharing his MODELFARM scripts, the NACP site synthesis team for providing the flux tower data sets, and X. Liang, D. Lettenmaier, G.-Y. Niu, Z.-L. Yang, and M. Sivapalan for their comments and suggestions on this study. PNNL is operated for the U.S. DOE by Battelle Memorial Institute under contract DE-AC06-76RLO1830.