Integration of the variable infiltration capacity model soil hydrology scheme into the community land model

Transcription

1 JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 113,, doi: /2007jd009246, 2008 Integration of the variable infiltration capacity model soil hydrology scheme into the community land model Aihui Wang, 1 Kaiyuan Y. Li, 1 and Dennis P. Lettenmaier 1 Received 2 August 2007; revised 7 December 2007; accepted 13 February 2008; published 10 May [1] Enhancements to the soil hydrology scheme in the NCAR Community Land Model version 3 (CLM3) are described, which are intended to improve the ability of CLM to represent land surface hydrologic processes. Specifically, the CLM3 soil hydrology scheme has been replaced with the scheme used in the Variable Infiltration Capacity model (VIC). While the modified model incorporates VIC soil hydrology, the original model structure in CLM3 is unchanged in most other respects, such as the representation of vegetation and its controls on surface energy exchanges. The modified version of CLM3 makes direct use of VIC soil parameters that have been developed for off-line regional, continental, and global simulations. The performance of the new model and CLM3 are evaluated through a comparison with observations at three flux tower sites: ABRACOS, a pasture clearing in the Brazilian rain forest, HAPEX-MOBILHY, a midlatitude agricultural site in France, and Valdai, a midlatitude grassland in Russia, as well as two large river basins (Arkansas-Red and Colorado in the United States). Overall, the results show that the merged model better reproduces observed soil hydrological variability, in particular, the seasonal evolution and amplitude of soil moisture, as compared to CLM3. The river basin simulations show that in the merged model, total runoff is generally less than in CLM3 and better agrees with observations. Owing to the interaction of runoff and soil moisture, the simulation of total evapotranspiration (or latent heat) is also improved in the merged model. Citation: Wang, A., K. Y. Li, and D. P. Lettenmaier (2008), Integration of the variable infiltration capacity model soil hydrology scheme into the community land model, J. Geophys. Res., 113,, doi: /2007jd Introduction [2] The land surface acts as the lower boundary of the atmosphere and controls fluxes of both energy and water across this boundary. Modern climate and weather forecast models use land surface process parameterizations [e.g., Garratt, 1993; Betts et al., 1996; Bonan et al., 2002; Zeng et al., 2002] to represent land surface controls on these fluxes, including the partitioning of net radiation into latent, sensible, and ground heat fluxes, and of precipitation into evapotranspiration and runoff. Land surface parameterizations (LSPs) have increased considerably in complexity, from simple bucket models [e.g., Manabe, 1969] to physically based soil-vegetation-atmosphere-transfer-schemes (SVATS) [e.g., Dickinson et al., 1993; Liang et al., 1994; Dai et al., 2003; Oleson et al., 2004]. Current physically based land surface models differ in their complexity, performance, and physical realism with which they represent different processes. Model intercomparison projects such as Project for Intercomparison of Land Surface Parameterization Schemes (PILPS) [see, e.g., Henderson-Sellers et al., 1993, 1995] have shown that even when as many model 1 Department of Civil and Environmental Engineering, University of Washington, Seattle, Washington, USA. Copyright 2008 by the American Geophysical Union /08/2007JD parameters as possible are prescribed to reduce degrees of freedom, different LSPs still have large disparities in performance. These differences among models are usually due to the complicated parameterizations and interactions among the model components, which differ from model to model. Studies like PILPS have shown that while the strengths and weaknesses of different models vary, model performance is often consistently better for some groups of models than for others, and that poor model performance can often be attributed to structural errors in specific parameterizations that may be relatively easy to correct. [3] One such parameterization, which critically affects LSP performance, because it directly impacts both water and energy fluxes, is soil hydrology. In LSPs, soil moisture depends on how the model partitions precipitation into canopy evaporation, infiltration and direct runoff, and in turn how the infiltration is partitioned into evapotranspiration and slow runoff. Soil moisture variation therefore depends on the interplay between the runoff and evapotranspiration formulations [Koster and Milly, 1997]. Evapotranspiration is governed by the LSP s representation of vegetation processes, whereas runoff is controlled by soil hydrology and precipitation. Runoff, therefore, is linked, albeit indirectly, to evapotranspiration, and hence to surface energy fluxes. As argued by Wood [1991] in a review of LSPs, most SVAT schemes focus more on the representation of vertical complexity in the vegetation interactions with the 1of15

2 atmosphere and less on the spatial variability that controls runoff. However, because of the above linkages, failure of LSPs to reproduce runoff processes properly has a ripple effect to energy fluxes as well. [4] Currently, runoff generation schemes in most LSPs are highly conceptualized. Observed streamflow at the river basin scale provides the best way to evaluate model performance in the reproduction of runoff. Comparison of modeled and observed runoff is best accomplished by transforming model runoff (a grid cell quantity) to streamflow through use of a river routing scheme which allows direct comparison of modeled streamflow with observations. Lohmann et al. [1998a] compared model-simulated runoff routed to the location of stream gauges in the Red- Arkansas River basin using a common routing scheme and found that even for models that reproduce annual runoff variability reasonably well, seasonal hydrographs vary greatly. On the basis of results of the two PILPS experiments that have evaluated LSP performance relative to observed streamflow, PILPS-2e (Arkansas-Red River basin; midcontinental North America) and PILPS-2e (Torne-Kalix River basin; high-latitude European), the runoff generation schemes that have performed the best are mostly derivatives of either the Variable Infiltration Capacity (VIC) model [Liang et al., 1994] or TOPMODEL [Beven and Kirkby, 1979]. It is notable that both of these parents have explicit parameterizations of the effects of subgrid variability in land surface conditions on runoff generation. [5] TOPMODEL represents subgrid land surface heterogeneity based on a topographic index, calculated using a digital elevation model (DEM). A TOPMODEL-based soil parameterization has been incorporated into CLM3 [Oleson et al., 2004] and provides the basis for partitioning a grid cell into saturated and unsaturated areas. The saturated area fraction is estimated from a water deficit depth and the maximum saturated fraction, determined by the distribution of the topographic index. Generally, the CLM3 runoff scheme has resulted in poor simulations of streamflow as compared with observations [Bonan et al., 2002; Dickinson et al., 2006], which is traceable in part to a relative lack of dynamics in soil moisture at both seasonally and the storm timescale. As noted above, CLM3 typically has a high bias in its model-predicted runoff. For example, Bonan et al. [2002] found that global runoff from land (excluding glaciers) is overestimated by CLM3 compared to observations. Lawrence et al. [2007] reported as well that CLM3 s partitioning of total evapotranspiration is unrealistic: canopy evaporation and soil evaporation are both too high, which leads to soils that are too dry and leads to a prominent warm summer temperature bias. [6] In contrast to TOPMODEL-based schemes, the VIC model on which the interactions between soil, biosphere, and atmosphere model is also based (ISBA) [Noilhan and Mahfouf, 1996] takes account of hydrological spatial heterogeneity by assuming that infiltration capacities, and hence runoff and evapotranspiration, vary within a grid cell due to variations in topography, soil, and vegetation using a single parameter representation [Wood et al., 1992; Liang et al., 1994]. Partitioning of precipitation into direct runoff and infiltration is represented by the infiltration capacity curve, whereas base flow is produced as a function of lower zone soil moisture storage using the Arno model [Todini, 1996] nonlinear base flow recession parameterization. One major advantage of the VIC formulation is that the model can be calibrated to observed discharge in a relatively straightforward manner, and the resulting parameters can either be mapped directly, or related to surrogate observed quantities such as soil properties. Using variations of this approach, the VIC model has been applied both globally [Nijssen et al., 2001] and over entire the continental United States [Maurer et al., 2002], as well as to numerous individual river basins. [7] Various attempts have been made to improve CLM3 s TOPMODEL-based runoff parameterization. A variation of TOPMODEL (SIMTOP) was developed by Niu and Yang [2005] and reportedly performed better than the standard CLM3 parameterization. SIMTOP simplifies the standard CLM3 soil hydrology scheme, but the accurate calculation of maximum fractional saturated area requires high-resolution topographic data. [8] Our concern in using a TOPMODEL-based formulation in an LSP such as CLM that is intended for global application is that the meaning of the critical link in TOP- MODEL to topographic features of a watershed is lost for several reasons. One is simply that the use of high-resolution topographic data at large scales becomes infeasible, and often the topographic index is estimated from coarse scale topographic data, which defeats the purpose of the model. More importantly, in many environments the physical underpinnings and assumptions are not met (e.g., downslope redistribution of moisture and hence saturated area control runoff generation), and in such cases, while satisfactory performance can be achieved, the physical motivation is lost. For these reasons, we pursue a different pathway here, and seek to improve CLM3 s hydrologic performance through incorporation of key aspects of VIC s soil hydrological representations [Liang et al., 1994]. In the remainder of this paper, we describe the land surface soil hydrology representations of both CLM3 and VIC, and the new hydrology scheme for CLM3 in section 2. In section 3, the data experimental design used to test the new model is described. Results are presented in section 4, and conclusions and discussions are provided in section Model Description [9] In this study, the NCAR CLM version 3 (hereafter CLM3) is used. Detailed descriptions of CLM3 are included in the work of Oleson et al. [2004]. CLM3 is the land component of the NCAR Community Climate System Model (CCSM3) [Collins et al., 2006]. It has been extensively used to perform both offline and coupled model simulations over global land areas [e.g., Dickinson et al., 2006]. Here, we focus on the soil hydrology scheme and its linkages with runoff generation and evapotranspiration CLM3 Runoff and Soil Moisture Parameterization Schemes [10] In CLM3, the soil column from the surface to a fixed depth of 3.43 m is divided into 10 layers, the thickness of which is fixed. The thickness of each layer increases from top to bottom, for instance, the top layer depth is 1.75 cm, whereas the bottom layer thickness is 1.14 m. Moisture to satisfy bare soil evaporation demands comes entirely from 2of15

3 within the thin surface layer. Soil water is calculated based on the balance between input water (throughfall precipitation plus snowmelting), runoff, infiltration, and transpiration. Soil storage changes within a time step. Richard s equation is used to predict the movement of moisture between layers, which allows water to move up or down between layers. The root profile is specified for a plant functional type (PFT). The water removed by transpiration is proportional to the root fraction in each soil layer. [11] Total runoff is the sum of surface runoff and base flow (subsurface drainage). Surface runoff, R s, is determined by using the TOPMODEL [Beven and Kirkby, 1979] concept, in which a soil grid cell is divided into saturated, f sat, and unsaturated (1 f sat ) area according to water table level f sat ¼ w fact min½1; expð z w ÞŠ; ð1þ where w fact = 0.3 is maximum saturated fraction, and z w is the mean dimensionless water table depth computed as (see Oleson et al. [2004] for details) z w ¼ f z z h;10 X10 i¼1 s i Dz i!; ð2þ where f z =1m 1 is a water table depth scale parameter, z h,10 is the depth of lowest soil layer (3.43 m), Dz i is the soil thickness of each layer, and s i is corresponding layer soil wetness, defined as the ratio of total volumetric soil water (liquid pluses ice) to saturated volumetric soil water in that layer. [12] The top three soil layers contribute to surface runoff R s ¼ f sat q liq þ ð1 f sat Þw 4 q liq ; ð3þ where q liq is the liquid water arriving soil surface (liquid throughfall rainfall pluses snowmelt). The first term on the right-hand side means that over the saturated area, all q liq goes to surface runoff and w is the soil layer thickness weighted wetness in the top three layers. [13] Base flow, R b, consists of lateral drainage from soil layers 6 9 (from 29 cm down to 2.3 m) and drainage out of the bottom layer (layer 10) plus any adjustments required to keep the liquid water content of each layer between maximum and minimum values. The total subsurface drainage is defined as: R b ¼ f sat l b expð z w þ X10 i¼1 Þþð1 f sat ÞK b w 2Biþ3 b þ K b b Dw 10 =Dt max 0; w i w max;i ; ð4þ where l b =10 5 is a base flow parameter, w b is the averaged wetness from layer 6 to layer 9, B i is a parameter depending on soil texture [Clapp and Hornberger, 1978], K b is the saturated soil hydraulic conductivity for the bottom layer [Cosby et al., 1984], w i is the soil water content in each layer, and Dt is the time step. The first two terms are the lateral drainage from saturated and unsaturated areas, respectively; the third and forth terms are the drainage from bottom layer and the change of hydraulic conductivity due to the change in liquid water content, respectively; and the last term maintains soil moisture in each layer within the maximum and minimum values. [14] Soil water is calculated based on the water balance between input water, runoff, infiltration, transpiration, and soil storage changes within a time step. Richard s equation is used to describe this @q K i i þ 1 ; ði ¼ 1; 2;...10Þ i where y i is the soil matrix potential, which depends on the soil texture and volumetric soil water in each layer, K i is the soil hydraulic conductivity of each layer, which is a function of saturated soil hydraulic conductivity, K sat,i. The definition of K sat,i follows the TOPMODEL concept [Beven and Kirkby, 1979], depending on the soil texture. [15] After defined all the needed qualities, equations (5) are converted to a set of discrete equations. Then soil water contents in the 10 layers are solved in a set of tridiagonal equations VIC Runoff and Soil Moisture Parameterization Schemes [16] In VIC, the soil column is divided into three layers whose thicknesses (unlike CLM3) vary from grid cell to grid cell. Similar to CLM3, total runoff consists of surface runoff and base flow. The thickness of the thin upper layer is usually set to about 5 10 cm. It is underlain by a second layer of medium thickness, and a deeper third layer, which mainly contributes to the base flow. A subgrid parameterization of soil properties is represented by a spatially varying infiltration capacity curve known as the Xinanjiang formulation [see Wood et al., 1992], which represents the partitioning of throughfall (after canopy interception) into infiltration to an upper thin soil layer and direct runoff. Surface runoff is controlled by a maximum infiltration capacity, i m (which is related to the thickness of the upper layer), as well as an infiltration parameter, b, and the maximum soil moisture in the two upper layers (w 12,c ). The surface runoff, R s, is calculated as R s ¼ q liq w n 12 w 12;c ( 1 1 max 1; i ) 0 þ q liq Dt 1þb ; i m where q liq is the liquid water arriving soil surface, i 0 is the infiltration capacity at the current soil moisture (see Liang et al. [1994] for details), and w n 12 is soil water at the beginning of this time step. [17] Soil moisture in the top two layers is then computed according to the water balance equation ð6þ w nþ1 1 ¼ w n 1 þ q liq Q 1 R s Tr 1 Dt; ð7þ w2 nþ1 ¼ w n 2 þ ð Q 1 Q 2 Tr 2 ÞDt; ð8þ 3of15

4 where w n+1 1 (or w n 1 ) is the water content at the end (beginning) of time step in the top layer. T ri is the transpiration through vegetation root in each layer, and Q i is the drainage from layer i to layer i + 1, which is defined as w n i w li i;r Q i ¼ K sat;i ; ði ¼ 1; 2Þ ð9þ w i;c w i;r where soil saturated hydraulic conductivity, K sat,i, is computed from the Brooks and Corey [1964] relation, w i,r is the residual moisture content, and l i is the pore size distribution index. The calculations of both l i and K sat,i depend on soil textures in layer i [Maidment, 1993]. [18] The base flow, R b, is computed using the Arno model concept, in which a nonlinear mechanism is used to produce the base flow [Todini, 1996], R b ¼ d sd m w n 3 w s w þ max 0; w 3 w s w 3;c 3;c w 3;c w s w 3;c d m d sd m w s ; ð10þ where d m is maximum base flow, d s is the fraction of d m where nonlinear base flow begins, w s is the fraction of field capacity, w 3,c. Those three parameters are estimated on a site specific basis. The third layer soil moisture is then computed as w nþ1 3 ¼ w n 3 þ ð Q 2 R b Tr 3 ÞDt: ð11þ [19] In the above water balance calculations, whenever the lower layer soil moisture content exceeds that layer s field capacity, the extra water is retained in the upper layer. If the top layer water exceeds its field capacity, the excess water goes to R s. On the other hand, if the soil water declines below the residual moisture, water is taken from the corresponding layer s Q i (for the top two layers) and R d (for the third layer) to make up the deficit. The scheme assures that the soil water content in each layer is always between field capacity and residual soil moisture. For each vegetation type, the root fraction is specified in each defined root zone, the thickness of which maybe differs from the soil layer thickness New Runoff and Soil Moisture Scheme for CLM3 [20] To implement equations (6) (11) into CLM3, we first redivide the 10 CLM3 soil layers into three layers (referred to as moisture layers), which correspond to the three VIC layers as described above. However, we retain the ten CLM3 layers for soil thermal calculations (we refer to these as thermal layers). As in VIC, the top two moisture layers generate surface runoff (using equation (6)), and the third layer contributes base flow (equation (9)). Soil liquid water is calculated following equations (7), (8), and (10). In the following context, CLM3 with the new hydrology scheme is referred to as CLM-VIC, and the original model is referred to as CLM3. [21] Following the above procedure, we need to map soil moisture between the three VIC moisture layers and the ten thermal layers, since soil thermal calculations are carried out on 10 soil layers and the soil thermal conductivity calculation requires the soil water content in each thermal layer. After the soil thermal calculations, soil liquid water, and ice content [e.g., Cherkauer and Lettenmaier, 1999] are redistributed within each layer due to the unmatched new layer temperatures with old liquid/solid distribution. The nature of this adjustment depends on whether the soil temperature is (1) above or (2) equal to or below the freezing point. The rate of this adjustment is assessed from the energy excess (or deficit) needed to change (lower or raise) soil temperature to the freezing point. The energy excess (deficit) comes from melting ice (freezing of liquid water). Furthermore, this phase change process is the only way in CLM3 that soil ice is generated if no snowpack is present. [22] In order to map soil moisture between moisture layers and thermal layers, we first calculate the fraction of each thermal layer within each moisture layer according to the layer depth and thickness. Using this fraction, the initial soil moisture is mapped to the thermal layer at the beginning of the model run. Here, we assume that the total moisture layer depth is shallower than total thermal layer depth so that the volumetric soil water in those thermal layers deeper than the third moisture layer are taken to be the same as that in the third moisture layer. This algorithm is similar to the approach used in the VIC model to map moisture to its thermal layers. After the thermal calculation, only the ice content is mapped back to the moisture layers for the next time step, and the soil liquid water in each moisture layer is calculated as the residual of the previous time step total soil moisture (liquid pluses ice) subtracts ice content. Total soil moisture is assumed to remain unchanged during the phase change adjustment. [23] Another issue to be addressed is the specification of root profiles within the soil column, which relates to the water sources for transpiration. Similar to VIC, the root distribution scheme in CLM3 is specified for one of 16 PFTs. In each soil layer, the root distribution of a specific PFT is a function of the soil depth and two PFT-dependence parameters derived from observed data [Zeng, 2001]. On the basis of this scheme, we redistribute the roots into the three VIC moisture layers. Then according to the root fraction, soil moisture is extracted for transpiration when the water balance equation is invoked to calculate the soil moisture in each moisture layer. 3. Data Description and Experimental Design [24] As mentioned above, both CLM3 and VIC have been tested extensively at regional, continental, and global scales. To evaluate the performance of the new soil hydrology scheme for CLM3, two types of data were used. First, meteorological data from two large river basins (the Red- Arkansas and Colorado River basins) were used to force the models, and observed streamflow was compared with the model-simulated runoff at stream gauge locations. Both basins have large variations in climate ranging from arid and semiarid in the west to humid in the east. Correspondingly, vegetation generally ranges from grassland to forest. Second, the observed meteorological data for three flux towers sites with different dominant vegetation cover were used to force both CLM3 and CLM-VIC, providing for comparison of simulated and observed moisture and energy fluxes at a point. 4of15

5 Figure 1. Arkansas-Red River basin map. Naturalized streamflow data (anthropogenic effects removed) for the stations indicated are compared with the model-simulated values Arkansas-Red River Basin [25] This basin, covering a total area of 566,000 km 2,is located in the U.S. southern Great Plains [Wood et al., 1998]. The mean annual runoff decreases from the eastern part of the basin with a value above 400 mm/a to the western part with values as low as 5 mm/a. The basin was subdivided into 3999 grid cells at a resolution of Figure 1 shows the basin map and selected gauge stations in which the model-simulated runoff and observed streamflow data can be compared. All stations are located in relatively far downstream regions in the basin, although the upstream-most stations capture runoff from the relatively dry area to the west. The streamflow data are naturalized observations, that is, observations that have been adjusted for the effects of upstream storage changes and diversions, by the Tulsa District of the U.S. Army Corps of Engineers as reported by Lohmann et al. [1998a]. The meteorological forcing data are from Maurer et al. [2002]. Using the Maurer et al. [2002] soil and vegetation parameters (originally used in the N-LDAS), both models were run for the 19-year period The first 2 years of the simulation period were treated as spin-up (to initialize model soil moisture) and were not used in the evaluations Colorado River Basin [26] The Colorado River basin is located in the southwest United States (Figure 2) and was the second river basin studied. The observation data are for the basin above Imperial Dam, with drainage area of 650,000 km 2. Most parts of the basin are arid, and over 70% of its annual runoff is derived from mountain snowpack in the relatively small Rocky Mountain headwaters area. Annual average streamflow is only about 40 mm averaged over the drainage area. The model representation of the basin includes 4518 grid cells at a resolution of Similar to the Arkansan-Red River basin, the Maurer et al. [2002] 3-h meteorological forcing data were used. Vegetation and soils data for both models are taken from the Maurer et al. [2002]. The simulations were conducted for the 13-year period at an hourly time step (via linear interpolation of the Maurer et al. 3-hourly data). As in the Arkansas-Red River simulations, the first 2 years were removed to initialize model soil moisture. The model output total runoff (surface runoff pluses base flow) in each grid cell was routed using the Lohmann et al. [1998a] routing model to produce streamflow at the gauging station locations ABRACOS Site (10.1 S 61.9 W) [27] The ABRACOS (Anglo-Brazilian Amazonian Climate Observation Study) field site is located at Reserva Jaru Figure 2. Colorado River basin map. Naturalized streamflow data (anthropogenic effects removed) for the stations indicated are used to compare with the model-simulated values. 5of15

6 Figure 3. Comparison of routed simulated monthly runoff with observed (naturalized) streamflow for the stations shown in Figure 1 and Table 1 in the Arkansas-Red River basin. in the Brazilian state of Rondônia. It is an open tropical forest site at which field observations were made from 1990 to 1994 [Gash et al., 1996]. The observed data set includes atmospheric forcing data, neutron probe soil moisture measurements, and radiative and turbulent flux data from November 1991 to December The flux data were measured during two intensive field campaigns from August to October 1992 and from April to August Using observed soil moisture for model initial conditions, the simulations were performed from November 1991 to December HAPEX-MOBILHY Site (43.7 N 0.1 W) [28] This agricultural (soybean) site is part of the larger HAPEX-MOBILHY experiment conducted in southwest France in 1986 [Andre et al., 1986]. The atmospheric forcing data are from half-hourly reference level weather observations. Energy flux measurements (net radiation, sensible heat, and ground heat fluxes) were taken every 15 min between 28 May and 3 July 1986, while the latent heat flux was calculated as the residual of the energy balance. Soil moisture was measured weekly to 1.6 m depth at 0.1 m intervals. A more detailed site description also can be found in the work of Shao and Henderson-Sellers [1996]. We initialized models with soil moisture observations on 7 January 1986 since this was the first date on which soil moisture was observed. The atmospheric data were aggregated to an hourly time step and the models were run through the end of the calendar year Valdai Site [29] The Valdai site is a small catchment (0.36 km 2 )in central Russia, mainly covered by grassland. The data set for the period of includes atmospheric forcings at 3-hourly intervals, and monthly soil moisture measurements for the top 1 m depth (average of 11 sites within the catchment). Details of this data set are documented by Robock et al. [1995]. Runoff observations are also available. Estimates of monthly evaporation based on lysimeter measurements during summer and a simple water balance algorithm for the winter months are also available from 1966 to The simulations were conducted for the entire Figure 4. Same as Figure 3, but for the multiyear (11 years for left panel and 9 years for right panel) monthly averages. 6of15

7 Table 1. Simulated Routed Runoff and Observed Streamflow Comparison Statistics Over the Gauge Stations Shown in Figures 1 and 2 in Two River Basins Gauge Name Lat ( N) Lon ( W) Flow area (mi 2 ) Rmse (100%) CLM3.0 CLM- VIC Relative Bias (100%) CLM3.0 CLM- VIC Average Observed Flow (m 3 s 1 ) Arkansas-Red River Basin Shreveport Fulton Farris Little Rock Dardanelle Sperry Average Colorado River Basin Green river Bluff Lees Ferry Hoover Davis Parker Imperial Cisco Average period after recycling the first year five times to initialize model soil moisture. [30] The forcing data we used at all three the flux towers are the same as in the work of Delire and Foley [1999], and similar vegetation and soil texture data were also used [see Delire and Foley, 1999, Table 1]. Other parameters, such as soil water conductivity and wilting point, were derived from model standard formulation. 4. Model Evaluations 4.1. Large River Basin Evaluations [31] Over the two large river basins, model simulated surface runoff and base flow in each grid cell were routed to the locations of the naturalized flow records shown in Figures 1 and 2 using the Lohmann et al. [1998b] routing algorithm at a daily time step. Simulated streamflow was then aggregated to the monthly time interval at which the naturalized streamflow observations were available. The time periods for naturalized streamflow observations for the Red and Arkansan Rivers are different, and for this reason the comparisons for the Arkansas River stations are for , and for the Red River stations are for Figure 3 shows the simulated monthly streamflow compared with observations. The figure shows that CLM-VIC produces lower streamflow, which is generally closer to observations than CLM3, even though the simulated CLM-VIC streamflow is still somewhat overestimated, especially peak values at some stations. CLM-VIC performs best at downstream stations, such as Shreveport and Little Rock. Although both modeled results capture seasonal variations reasonably well (Figure 4), substantial biases between simulated and observed flows remain. The greatest differences between the two model-simulated streamflows are in the wet season, during which CLM3 greatly overestimates streamflow, a bias that is much reduced by CLM- VIC. The CLM3 overestimations mainly occur because soil moisture removed by transpiration is underestimated, a known problem with CLM3 [Lawrence et al., 2007]. Following the method used by Maurer et al. [2002], the root-mean square error (RMSE), and relative bias between simulated and observed monthly streamflow for those selected stations are summarized in Table 1. The positive relative bias is indicative of the runoff overestimation, and is consistent with Figures 3 and 4. The CLM-VIC RMSEs and relative biases confirm improved performance of the model for the larger drainage areas. For instance, at Little Rock, the RMSE (50%) and relative bias (14%) for CLM- VIC are the smallest among all stations in the Arkansas-Red basin. Rather than a reflection of performance relative to drainage basin size, however, we think that this has more to do with the fact that the lower Arkansas-Red basin is much more humid than the upper basin, so downstream stations are more affected by lower basin conditions, and CLM-VIC performs better for semihumid and humid than for arid conditions. [32] Figure 5 compares simulated and observed monthly streamflow for selected stations indicated in Figure 2 over the Colorado River basin. CLM-VIC performs reasonably well in reproducing observed streamflow, while CLM3 overestimates seasonal peak streamflow during the entire comparison period (Figure 6) for all stations. CLM3 s performance in summer is generally better than over the rest of the year. There is a consistent phase shift in both CLM3 and CLM-VIC with the spring streamflow peak occurring about 1 month earlier than in the observed peak. This phase shift is seen more clearly in seasonally averaged streamflow simulations as compared with observations (Figure 6). In order to diagnose the reason for the phase shift, we also plotted VIC simulated monthly streamflow in Figure 6, which shows that the VIC simulations agree with observations quite well, and generally do not have a phase shift. Because CLM-VIC is nearly identical to VIC in terms of its runoff production mechanisms, we believe that the biases in CLM3 and CLM-VIC are attributable to 7of15

8 Figure 5. Same as Figure 3, but for the Colorado River basin. the CLM3 snow model, which was remain unchanged in CLM-VIC. [33] To evaluate this apparent bias further, we also plotted average monthly snow water equivalent (SWE) over a grid cell (43.4 N, W) in the upper portion of the basin, where most snowmelt originates. Figure 7 shows an obvious bias toward early snowmelt in both CLM3 and CLM3.5 (the newest public released version of CLM) relative to VIC. We also performed a similar comparison between the two models and observations of SWE at an NRCS SNOTEL site within the Colorado River basin (not shown) and found a similar tendency of both versions of CLM to simulate smaller SWE accumulations, and more rapid melt, than either VIC or observations. [34] Because snow albedo plays an important role in snowmelt timing, we compared the daily albedo over snow for CLM and VIC and found that the maximum snow albedo is about in CLM, which is much smaller than VIC which has a maximum albedo (for new snow) of about 0.85, and a snow aging parameterization that allows this maximum value to decay over several days to a lower value of about 0.5. We believe that the difference in snow albedo is the major reason for the apparent bias of both CLM3 and CLM3.5 toward early snowmelt. This is an issue we intend to address in detail in a subsequent publication, but it is beyond the scope of this paper. Even with the bias toward early snowmelt, however, CLM-VIC generally performs better over the Colorado basin than over the Figure 6. Same as Figure 4, but for the Colorado River basin. 8of15

9 Figure 7. Comparison of model-simulated snow water equivalent (SWE, mm) at the grid cell (43.4 N, W) in the northern of Colorado River basin. Arkansan-Red River basin. It should be noted, however, that the success of the CLM-VIC and VIC models in reproducing observed streamflows is partly because the soil parameters used (from Maurer et al. [2002]) were better calibrated for the Colorado River basin than for the Arkansas-Red River basin. [35] The results shown for CLM-VIC for the two test basins are encouraging, and we believe they can be further improved through calibration of the VIC soil parameters (in the Arkansas-Red River basin) and through resolution of apparent biases in the CLM3 snow model (in the case of the Colorado River basin). Since the streamflow observations are widely available, and errors in runoff are reflected in errors in evapotranspiration, and hence land-atmosphere energy fluxes as well, much of the attention has been focused on CLM-VIC s runoff production. In land surface models, runoff simulation is related to many processes, such as canopy-intercepted rainfall, soil evaporation, and, in cold climates, snow accumulation and ablation. For example, canopy-intercepted rainfall indirectly determines how much water arrives at the ground and can potentially contribute to runoff. Runoff also affects soil moisture and evapotranspiration. If surface runoff occurs too often, the upper layer soil has to extract moisture from deeper layers or reduce surface evaporation to supplement the water lost to runoff. Koster and Milly [1997] examined multiple land surface models (LSMs) and concluded that the compatibility between runoff and evapotranspiration schemes determines their soil moisture simulations. Soil moisture and evapotranspiration data for large basins are not easily obtained. However, flux tower measurements of water and energy fluxes offer an opportunity to test the nature of the soil moisture relationship with energy fluxes Comparison of Modeled Water and Energy Fluxes at Tower Flux Sites [36] Figure 8 shows the model-simulated and observed total column soil moisture, evapotranspiration and surface runoff at the ABRACOS rain forest site. In general, the soil in CLM3 is too dry, and the evolution of soil moisture has much less seasonal variation than observations. Similar results have been reported by Lawrence et al. [2007]. CLM-VIC does a much better job in capturing both the seasonal variation and magnitude of soil moisture over the wet seasons, although the simulated soil is slightly wetter than observed in the dry season. Comparison of simulated runoff from both models shows that the CLM3-simulated surface runoff is much larger than in CLM-VIC, and tracks precipitation more closely than CLM-VIC. Both modelsimulated estimates of evapotranspiration (ET) show similar seasonal variations as precipitation, however CLM3- simulated total ET is relatively low in the dry season. CLM-VIC simulated ET is closer to observed whereas CLM3 in general underestimates ET (Figure 8b). This result is also reflected in latent heat (Figure 9b). The timing of base flow in CLM-VIC is consistent with the timing of the maximum soil moisture storage, whereas in CLM3 it is lagged by about 2 months. The net radiation is well simulated in both models (Figure 9a), while the sensible heat is overestimated in CLM3. This is the consequence of too much runoff (Figure 9), and underestimated latent heat in the CLM3 simulation (Figure 9b). The soil heat fluxes in CLM3 are calculated as the residual of net radiation, sensible and latent heat. Poorly simulated ground heat fluxes (Figure 9d) could be the result of mischaracterization of the physics of soil heat in the model or the different representations of the measured and modeled soil heat. Furthermore, soil heat flux is difficult to measure, a problem that is exacerbated by the age (about 15 years) of the measurements. Delire and Foley [1999] has suggested possible reasons for poor simulated soil heat flux in their model, which may also be relevant here. [37] Over the HAPEX-MOBILHY site, model-simulated soil moisture (to 1.6 m), ET, surface runoff and observed precipitation are shown in Figure 10. Similar to the ABRA- COS site, the soil in CLM3 is too dry and soil moisture variations are too small, while CLM-VIC modeled soil moisture is in overall agreement with observation of both seasonal variation and amplitude, aside from slight underestimates in the wet season. Simulated surface runoff in CLM3 follows precipitation very closely, and the variations are much smaller than those in precipitation. The peak seasonal ET in CLM3 is in May, which is one month later than the peak seasonal precipitation. In contrast, the maximum of CLM-VIC simulated ET lags precipitation by about 2 months. The longer lag in CLM-VIC is likely related in part to differences in the topsoil moisture layer in the two models, and the manner by which soil evaporation is calculated. In CLM-VIC, soil moisture in the top thermal layer (about 1.75 cm) is calculated as a part of the top moisture layer (about 10 cm), while in CLM3 soil moisture in the top thermal layer is calculated directly for this (thinner) layer. Because the retention time of soil moisture in the deeper layer is longer than in the upper 9of15

10 Figure 8. Comparisons of simulated and observed soil moisture, evapotranspiration, runoff, and observed precipitation, for ABRACOS site: (a) observed (dotted) and model-simulated (solid) total column soil moisture (mm); (b) model-simulated evapotranspiration (mm); (c) model-simulated surface runoff (mm); (d) model-simulated base flow (mm); (e) observed precipitation (mm). The observed soil moisture is accumulated from the surface to 3.43 m. 10 of 15

11 Figure 9. Comparison of model-simulated with observed (a) net radiation (CLM3 stars, CLM-VIC circles); (b) latent heat; (c) sensible heat; (d) soil heat flux for ABRACOS site. The period of observation was 8 August 1992 to 4 October layer [Wang et al., 2006], the soil evaporation contributed by the deeper layers is delayed longer relative to precipitation. [38] The base flow produced by the two models is similar, except that simulated base flow from CLM3 is very high at the beginning of the simulation. Similar to the ABRACOS site, both models simulate net radiation well, but overestimate the sensible heat and underestimate latent heat (Figure 11). CLM-VIC simulated latent heat is better than in CLM3, which corresponds to too much runoff in CLM3. The soil heat is not simulated well by either model, the explanation for which is similar to ABRACOS (see above). [39] Over the Vaidai site, the annual variation of soil moisture content is well captured by CLM-VIC although the maximum storage is overestimated in most years (Figure 12a). Similar to the other sites, the variation of CLM3-simulated soil moisture and soil water storage is too low. Because total ET depends in part on soil moisture, the model with better soil moisture simulation better simulates ET (Figure 12b). Both model-simulated snow depths are in general agreement with observations, although the maximum snow depth is overestimated in some years by both models and the snowmelts too early in spring (Figure 12c) in both models. The simulated monthly runoff is comparable to the observed, except that CLM-VIC underestimates peaks in some years and CLM3 overestimates them in other years. 5. Discussion and Conclusions [40] The soil hydrology scheme in the NCAR Community Land Model (CLM3) has been replaced with the 11 of 15

12 Figure 10. For the HAPEX-MOBILHY site, (a) observed and model-simulated total column soil moisture (mm); (b) model-simulated evapotranspiration (mm); (c) model-simulated surface runoff (mm); (d) model-simulated base flow (mm); (e) observed precipitation (mm). The observed soil moisture is accumulated from the surface to 1.6 m. The data are for calendar year scheme used in the Variable Infiltration Capacity model (VIC). While the merged model incorporates VIC soil hydrology, the original model structure is unchanged in most other respects, such as the representation of vegetation and its controls on surface energy exchanges. [41] The merged model (termed CLM-VIC) was evaluated using streamflow data for two large river basins, as well as flux tower observations of land-atmosphere energy exchanges at three sites. The overall results show that the merged model improves the soil hydrology representation and in turn surface moisture and energy fluxes, especially in the reproduction of streamflow and soil moisture. The variation and amplitude of soil moisture in the new model are generally in closer agreement with observations. When the model simulated runoff is routed to stations with observed streamflow, CLM-VIC produced streamflows are generally smaller than that from CLM3 and in closer agreement with observations. In general, the reproduction 12 of 15

13 Figure 11. Comparison of model simulated with observed (a) net radiation (CLM3 stars, CLM-VIC circles); (b) latent heat; (c) sensible heat; (d) soil heat flux over the HAPEX-MOBILHY site. The period of observations is from 28 May 1986 to 3 July of streamflow by the CLM-VIC model is better in humid and semihumid regions than in arid regions. On the other hand, the latent heat simulation is improved in the CLM- VIC model across the full range of climates represented by the observation stations. [42] We believe that the detailed analysis of CLM3, and improvements to its soil hydrology, will be of general interest to the broader climate modeling community given that CLM is the land surface scheme used in CCSM, one of the most widely used GCMs by the U.S. climate community. The potential for integration of some aspects of VIC land surface hydrology is of interest because of its extensive and successful use in off-line simulations, including retrospective climate assessments [e.g., Hamlet and Lettenmaier, 1999; Payne et al., 2004; Christensen et al., 2004; Christensen and Lettenmaier, 2007] and seasonal streamflow forecasting [Wood and Lettenmaier, 2006]. In general, land surface parameterizations used for hydrological applications pay more attention to subgrid variability than those that derive from other heritages, in part because of the key role that subgrid variability plays in runoff generation. In contrast, land surface schemes that come from a meteorological heritage tend to place more emphasis on vertical processes, such as surface fluxes (vertically distribution of latent/sensible heat), and soil temperature profiles, hence the simulated energy fluxes (sensible/latent heat fluxes and ET). Graham and Bergström [2000] reviewed the differences of land surface schemes in hydrological and meteorological models and also summarized the advantages of merging these two schemes through a modeling experiment. In this paper, we have accomplished such a merger. [43] We have focused here on improving the soil hydrological representation of CLM3. A recent publicly released version 3.5 of CLM contains a number of upgrades to the version 3.0 on which this paper is based, including modifications to the soil hydrology scheme, as well as some other changes, such as altered canopy interception (see CLM3_5_documentation.pdf). Version 3.5 continues to undergo revisions as this paper is written, nonetheless, based on the methods described in section 2.3, we have performed some exploratory analysis with a version of CLM3.5 which incorporates VIC soil hydrology, and the improvement in simulated streamflow (not shown) over the Colorado River basin showed similar changes relative to the results shown earlier. However, the phase shift in simulated 13 of 15

14 Figure 12. Comparison of model-simulated with observed (a) soil moisture (mm); (b) evapotranspiration, ET (mm); and (c) snow water equivalent SWE (mm); (d) total runoff (mm) for the Valdai site. The observed soil moisture is accumulated from the surface to 1 m. streamflow (which peaks too early in the year) still exists, apparently due to unresolved issues with the CLM snow model (discussed in section 4.1). Integration of VIC soil hydrology into CLM3.5 is somewhat complicated by a number of substantial differences between CLM3.0 and CLM3.5 and, given that the release of CLM3.5 occurred only shortly before this paper was completed and is still evolving, we have chosen not to conduct a detailed analysis of a combination of VIC and CLM3.5, although we intend to do so in the future. [44] Acknowledgments. The research reported herein was supported by the U.S. Department of Energy under DOE agreement DE-FG02 04ER63873 to the University of Washington. We thank Ben Livneh of the University of Washington Land Surface Hydrology Group for his comments and insights. Three anonymous reviewers are also thanked for their valuable comments. References Andre, J. C., J. P. Goutorbe, and A. Perrier (1986), HAPEX-MOBILHY: A hydrologic atmospheric experiment for the study of water budget and evaporation flux at the climatic scale, Bull. Am. Meteorol. Soc., 67, , doi: / (1986)067<0138:hahaef>2.0.co;2. Betts, A. K., J. H. Ball, A. C. M. Beljaars, M. J. Miller, and P. Viterbo (1996), The land surface-atmosphere interaction: A review based on observational and global modeling perspectives, J. Geophys. Res., 101, , doi: /95jd Beven, K. J., and M. J. Kirkby (1979), A physically based, variable contributing model of basin hydrology, Hydrol. Sci. Bull., 24, Bonan, G. B., K. W. Oleson, M. Vertenstein, S. Levis, X. Zeng, Y. Dai, R. E. Dickinson, and Z. L. Yang (2002), The land surface climatology of the community land model coupled to the NCAR Community Climate Model, J. Clim., 15, , doi: / (2002)015<3123:tlscot>2.0.co;2. 14 of 15

15 Brooks, R. H., and A. T. Corey (1964), Hydraulic properties of porous media, Hydrol. Pap. 3, 27 pp., Colo. State Univ., Fort Collins. Cherkauer, K. A., and D. P. Lettenmaier (1999), Hydrologic effects of frozen soils in the upper Mississippi River basin, J. Geophys. Res., 104, 19,599 19,610, doi: /1999jd Christensen, N. S., and D. P. Lettenmaier (2007), A multimodel ensemble approach to assessment of climate change impacts on the hydrology and water resources of the Colorado River basin, Hydrol. Earth Syst. Sci., 11, Christensen, N. S., A. W. Wood, N. Voisin, D. P. Lettenmaier, and R. N. Palmer (2004), Effects of climate change on the hydrology and water resources of the Colorado River basin, Clim. Change, 62, , doi: /b:clim f. Clapp, R. B., and G. M. Hornberger (1978), Empirical equations for some soil hydraulic properties, Water Resour. Res., 14, , doi: / WR014i004p Collins, W. D., et al. (2006), The community climate system model version 3, J. Clim., 19, , doi: /jcli Cosby, B. J., G. M. Hornberger, R. B. Clapp, and T. R. Ginn (1984), A statistical exploration of the relationships of soil moisture characteristics to the physical properties of soils, Water Resour. Res., 20, , doi: /wr020i006p Dai, Y., et al. (2003), The Common Land Model, Bull. Am. Meteorol. Soc., 84, , doi: /bams Delire, C., and J. A. Foley (1999), Evaluating the performance of a land surface/ecosystem model with biophydical measurements from contrasting environments, J. Geophys. Res., 104, 16,895 16,909, doi: / 1999JD Dickinson, R. E., A. Henderson-Sellers, and P. J. Kennedy (1993), Biosphere-Atmosphere Transfer Scheme (BATS) version le as coupled to the NCAR Community Climate Model, NCAR Tech. Note, NCAR/TN- 387+STR, Natl. Cent. for Atmos. Res., Boulder, Colo. Dickinson, R. E., K. W. Oleson, G. Bonan, F. Hoffman, P. Thornton, M. Vertenstein, Z.-L. Yang, and X. Zeng (2006), The Community Land Model and its climate statistics as a component of the Community Climate System Model, J. Clim., 19, , doi: /jcli Garratt, J. R. (1993), Sensitivity of climate simulations to land-surface and atmospheric boundary-layer treatments A review, J. Clim., 6, , doi: / (1993)006<0419:socstl>2.0.co;2. Gash, J. C., C. Nobre, J. Roberts, and R. Victoria (1996), An overview of ABRACOS, in Amazonian Deforestation and Climate, edited by J. H. C. Gash et al., pp. 1 14, John Wiley, New York. Graham, L. P., and S. Bergström (2000), Land surface modelling in hydrology and meteorology Lessons learned from the Baltic Basin, Hydrol. Earth Syst. Sci., 4, Hamlet, A. F., and D. P. Lettenmaier (1999), Effects of climate change on hydrology and water resources in the Columbia River basin, J. Am. Water Resour. Assoc., 35, Henderson-Sellers, A., Z.-L. Yang, and R. Dickinson (1993), The project for intercomparison of land-surface parameterization schemes, Bull. Am. Meteorol. Soc., 74, , doi: / (1993) 074<1335:TPFIOL>2.0.CO;2. Henderson-Sellers, A., A. Pitman, P. Love, P. Irrannejad, and T. Chen (1995), The project for intercomparison of land-surface parameterization schemes (PILPS): Phase 2 and 3, Bull. Am. Meteorol. Soc., 76, , doi: / (1995)076<0489:tpfiol>2.0.co;2. Koster, R. D., and P. C. D. Milly (1997), The interplay between transpiration and runoff formulations in land surface schemes used with atmospheric models, J. Clim., 10, , doi: / (1997)010<1578:TIBTAR>2.0.CO;2. Lawrence, D. M., P. E. Thornton, K. W. Oleson, and G. B. Bonan (2007), The partitioning of evapotranspiration into transpiration, soil evaporation and canopy evaporation in a GCM: Impact on land-atmosphere interaction, J. Hydrometeorol., 8, , doi: /jhm Liang, X., D. P. Lettenmaier, E. F. Wood, and S. J. Burges (1994), A simple hydrologically based model of land surface water and energy fluxes for general circulation models, J. Geophys. Res., 99, 14,415 14,428, doi: /94jd Lohmann, D., et al. (1998a), The Project for Intercomparison of Landsurface Parameterization Schemes, PIPLS) Phase 2(c) Red-Arkansas River basin experiment: 3. Spatial and temporal analysis of water fluxes, Global Planet. Change, 19, , doi: /s (98) Lohmann, D., E. Raschke, B. Nijssen, and D. P. Lettenmaier (1998b), Regional scale hydrology: I. Formulation of the VIC-2L model coupled to a routing model, Hydrol. Sci. J., 43, Maidment, D. R. (1993), Handbook of Hydrology, 1424 pp., McGraw-Hill, New York. Manabe, S. (1969), Climate and ocean circulation: I. The atmospheric circulation and the hydrology of the earth s surface, Mon. Weather Rev., 97, , doi: / (1969)097<0739:catoc>2.3.co;2. Maurer, E. P., A. W. Wood, J. C. Adam, and D. P. Lettenmaier (2002), A long-term hydrologically based dataset of land surface fluxes and states for the conterminous United States, J. Clim., 15, , doi: / (2002)015<3237:althbd>2.0.co;2. Nijssen, B., P. Schnur, and D. P. Lettenmaier (2001), Global retrospectie estimation of soil moisture using the variable infiltration capacity land surface mode, , J. Clim., 14, , doi: / (2001)014<1790:greosm>2.0.co;2. Niu, G.-Y., and Z.-L. Yang (2005), A simple TOPMODEL-based runoff parameterization (SIMTOP) for use in global climates models, J. Geophys. Res., 110, D21106, doi: /2005jd Noilhan, J., and J.-F. Mahfouf (1996), The ISBA land surface parameterization scheme, Global Planet. Change, 13, , doi: / (95) Oleson, K. W., et al. (2004), Technical description of the community land model (CLM), NCAR Tech. Note NCAR/TN-461+STR, 174 pp., Natl. Cent. for Atmos. Res., Boulder, Colo. Payne, J. T., A. W. Wood, A. F. Hamlet, R. N. Palmer, and D. P. Lettenmaier (2004), Mitigating the effects of climate change on the water resources of the Columbia River basin, Clim. Change, 62, , doi: /b:clim d6. Robock, A., K. Vinnikov, C. A. Schlosser, N. A. Speranskaya, and Y. Xue (1995), Use of midlatitude soil moisture and meteorological observations to validate soil moisture simulations with biosphere and bucket models, J. Clim., 8, 15 35, doi: / (1995)008<0015: UOMSMA>2.0.CO;2. Shao, Y., and A. Henderson-Sellers (1996), Validation of soil moisture simulation in land surface parameterization schemes with HAPEX data, Global Planet. Change, 13, 11 46, doi: / (95) Todini, E. (1996), The ARNO rainfall-runoff model, J. Hydrol., 175, Wang, A., X. Zeng, S. S. P. Shen, Q.-C. Zeng, and R. E. Dickinson (2006), Timescales of land surface hydrology, J. Hydrometeorol., 7, , doi: /jhm Wood, A. W., and D. P. Lettenmaier (2006), A testbed for new seasonal hydrologic forecasting approaches in the western U.S., Bull. Am. Meteorol. Soc., 87, , doi: /bams Wood, E. F. (1991), Global scale hydrology: Advances in land surface modeling, Rev. Geophys., 29, Wood, E. F., D. P. Lettenmaier, and V. G. Zartarian (1992), A land surface hydrology parameterization with subgrid variability for general circulation models, J. Geophys. Res., 97, Wood, E. F., et al. (1998), The Project for Intercomparison of Land-surface Parameterization Schemes, PIPLS) Phase 2(c) Red-Arkansas River basin experiment: 1. Experiment description and summary intercomparisons, Global Planet. Change, 19, , doi: /s (98) Zeng, X. (2001), Global vegetation root distribution for land modeling, J. Hydrometerol., 2, , doi: / (2001) 002<0525:GVRDFL>2.0.CO;2. Zeng, X., M. Sharikh, Y. Dai, R. E. Dickinson, and R. Myneni (2002), Coupling of the common land model to the NCAR community climate model, J. Clim., 15, , doi: / (2002) 015<1832:COTCLM>2.0.CO;2. D. P. Lettenmaier, K. Y. Li, and A. Wang, Department of Civil and Environmental Engineering, University of Washington, Box , Seattle, WA , USA. 15 of 15