Application of multiscale water and energy balance models on a tallgrass prairie

. The models presented in the previous paper (Famiglietti and Wood, this issue) are applied at their appropriate scales for evapotranspiration modeling at the First International Satellite Land Surface Climatology Project Field Experiment (FIFE) site. The local soil-vegetation-atmospheric transfer scheme is applied at five flux measurement stations in the northwest quadrant of the FIFE site. Simulations were performed for three of the four FIFE golden (cloud-free) days with good results. The spatially distributed model was applied at the ! 1.7-km 2 King's Creek catchment, also located in the northwest quadrant of the FIFE site, during FIFE Intensive Field Campaigns (!FCs) 1-4. Simulated catchment average evapotranspiration was compared to an average of observations made at the five aforementioned measurement stations with good results. The macroscale formulation was applied to both the King's Creek catchment and the entire !5-km FIFE site for evapotranspiration simulations. Macroscale model simulations for King's Creek were nearly identical to the spatially distributed results, implying that at this location and at this scale, the assumptions invoked in the development of the macroscale formulation are reasonable. The macroscale model was also employed to simulate evapotranspiration from the entire 15-km site for the four golden days. Simulated evapotranspiration rates show reasonably good agreement with the 22-station average of observations. However, it is suggested that at 15-km and larger scales, simulation error may arise as a result of the macroscale assumptions of areally averaged atmospheric forcing, vegetation parameters, soil parameters, and the methods by which these data and other flux observations are aggregated. A methodology to combat these problems at larger scales is reviewed. the use of nonlocal forcing data, and errors identified as sources of error in the


Introduction
This paper presents the application component of a body of research which addresses aggregation and scaling issues in multiscale hydrological modeling. In the first paper [FamigIietti and Wood, this issue], a methodology was proposed to aggregate local process physics across scales. A spatially distributed modeling framework was proposed for use at the catchment scale, and at the macroscale, a statistical-dynamical framework was presented. The macroscale formulation is intended for use as a land parameterization in regional and global atmospheric models. The purpose of this study is to apply the models of the previous paper [Famig!ietti and Wood, this issue] on a temperate grassland at their appropriate scales. A second goal of this work is to investigate some of the simplifying assumptions utilized in the development of the macroscale formulation using observed field data.
The site of these applications is the tallgrass prairie of eastern Kansas (United States). The area includes rolling hills and shallow soils, with roughly 50 m of elevation from Copyright 1994 by the American Geophysical Union.
Paper number 94WR01499. 0043-! 397/94/94 WR-01499505.00 stream bottoms to ridge tops, and is representative of the strip of native tallgrass prairie, 50-80 km wide, that extends from Kansas to Nebraska to Oklahoma. In the summers of 1987 and 1989, the First International Land Surface Climatology Project (ISLSCP) Field Experiment (FIFE) was conducted on a 15 x 15 km region of tallgrass prairie located near Manhattan, Kansas. FIFE was a large-scale field experiment whose purpose was to develop relationships between satellite measurements and hydrologic, climatic, and biophysical variables at the !and surface [Sellers et al., 1992]. A second goal of the experiment was to collect ground-based data to validate these relationships, and to validate simulation models of land surface processes from the point scale to scales compatible with remotely sensed observations. During the summers of 1987 and !989, multiscale ground-based and remotely sensed water and energy balance data were collected simultaneously. Thus the FIFE data set affords unique opportunities to observe the scaling behavior of hydrological processes and to test modeling strategies over a range of spatial scales.
In this paper, the local soil-vegetation-atmosphere transfer scheme (  Net radiation and wind speed observed at individual  Table 3 summarizes these vegetation parameters. Note that the LAI was treated as a tuning parameter in some of the local SVATS simulations and macroscale model simulations of the entire site. Therefore some of the LAI values shown in Table 3

Local-Scale Results
In this section, the results of SVATS applications at individual stations are presented. The local SVATS is applied at stations 2, 8, 10, 12, and 14. Simulations were performed using half-hourly time steps and the data described above. Modeled evapotranspiration was compared to that observed on three of the four FIFE golden (virtually cloud-free) days (June 6, July 11, and August 15, 1987).
Observed flux data for stations 4 and 6, also located in the vicinity of the King's Creek catchment, were either unavailable or of questionable quality. Observed flux data were als0 unavailable for stations 8, 10, 12, and 14 for October 11, 1987, the final golden day of the 1987 FIFE campaign. When analyzing the local-scale results, it is important to remember the following points. First, as was discussed above, some meteorological data (e.g., air temperature and pressure) were not readily available for individual stations from the FIS. Thus unavoidable bias results from driving local simulations with nonlocal data. The situation becomes more complicated when measurement errors in meteorological data and observed fluxes are considered. Errors induced by averaging air temperature, pressure, and humidity data further compound the problem. Second, the structure of the local SVATS is greatly simplified relative to others in current use so that it can be aggregated in space. Therefore our local SVATS is designed to capture the essential dynamics of land-atmosphere interaction: there are likely more detailed SVATS in operation that can better reproduce evapotranspiration observed at individual stations. Most errors in the local simulations are attributable to these two sources and are not discussed in depth below.  Simulated evapotranspiration at station !4 agreed well with observations (see Figure 2e). Station !4 was located on an unburned, moderately sloping, east-facing site. Computed evapotranspiration rates were slightly lower than those observed at midday on June 6 (rmse 0.02 tampa).
Observed flux data were unavailable for July 11 at station 14.
Computed evapotranspiration rates for August 15 agreed well with observations (rmse 0.03 mm/h).

Catchment-Scale Results
In this section, the spatially distributed model is applied at the King's Creek catchment (see Figure 1) The bottom panels of Plate 2 show modeled midday evapotranspiration rates for the corresponding times in the top panels. Evapotranspiration rates vary from near 0 to 0.4 mm/h. These images give some indication of the degree of spatial variability in evapotranspiration rates within the catchment. Such high-frequency variability was not sampled by the flux measurement stations, since only one station was located within the catchment, and a total of 22 stations were located within the entire !5-km site. Furthermore, logistical considerations required that most stations be located on prairie toplands and moderate slopes, so that steep slopes and valley bottoms were undersampled. Thus spatial patterns such as those shown in the bottom panels of Plate 2 are

difficult to ve_rify. However, FIFE investigators HolwiIl and Stewart [1992] and Jedlovec and Atkinson [1992] showed that such high-frequency variation existed using highresolution remotely sensed imagery. The patterns shown in the bottom panels of Plate 2 are consistent with the microtopographic variation shown in the imagery produced in both of these studies.
Comparison of the top and bottom panels of Plate 2 shows a strong relationship between spatial patterns of root zone moisture content and evapotranspiration. Evapotranspira~ tion rates decrease with decreasing root zone moisture content. Wetter grid elements located along the stream network evaporate at higher rates than drier locations, such as those located near ridge tops. As was described above, IFC 4 was drier than the previous IFCs, resulting in active soil  dependent on moisture content, the correspondence between the top and bottom of Plate 2 is understandable. The larger number of dark blue and green grid elements in the bottom of Plate 2 for October 5 indicates that evapotranspiration rates were higher than on October 9. The decrease in evapotranspiration rates during that period is due to the catchment-wide decrease in root zone moisture content, which resulted in decreased exfiltration and transpiration capacities locally. To demonstrate the surface runoff capabilities of the model, the runoff-producing storm event of IFC 3 (August 13, 1987) was simulated independently of the work presented previously. In this simulation the model was tuned so that the volume of computed storm runoff matched that observed. All model parameters are given in Tables 1-3. As in the evapotranspiration simulation of IFC 4, a gamma distribution of initial root zone moisture content was assumed.
Computed streamflow was tuned to that observed by increasing the root zone depth from 0.5 to 0.75 m and by increasing the saturated hydraulic conductivities of Table 1 by a factor of 5. Such independent tuning was required owing to a lack of runoff-producing storm events during the summer of !987. To properly calibrate and verify runoffrelated model parameters, rainfall and streamflow data from a number of events would be required. Since these data were not available, no attempt was made to identify one optimal parameter set for continuous simulation. Spatially

Discussion
This discussion focuses on the implications of the simulation results of the previous section. Specifically, we are interested in the implications of the agreement between macroscale model and spatially distributed simulations of evapotranspiration for the King's Creek catchment, and how these might change as spatial scale is increased. As a first step toward understanding the aggregation and scaling properties of land surface processes, we conducted a detailed investigation of the scaling behavior of evapotranspiration from local to catchment scales (paper 3). Simulation studies such as the present work and paper 3 are important because they provide insight into the role of naturally heterogeneous land surface properties and processes and how important spatial variability can be included in macroscale hydrological formulations. These types of studies provide a framework within which land hydrology parameterizations can be continually modified and our understanding of large-scale hydrological processes improved.
The assumptions invoked in the development of the macroscale formulation include an areally averaged representation of meteorological inputs, soil parameters, and vegetation parameters. Spatial variability in the topographic-soil index, soil moisture, surface runoff, and the energy fluxes is represented statistically, rather than explicitly, as in the spatially distributed formulation. Comparison of Figures 4   and 3 shows that for the 11.7-km 2 grassland King's Creek catchment these are reasonable assumptions, since macroscale model and spatially distributed results are nearly identical. Some discussion regarding why these results are similar and when they might differ should provide insight into the restrictiveness of the macroscale assumptions at larger spatial scales and in different geographic locations.
By comparing macroscale and spatially distributed evapotranspiration equations, Famiglietti and Wood [this issue] showed that the difference between evapotranspiration computed with the two formulations would depend upon two factors. The first is related to the degree of spatial variability in model parameters and inputs. The second factor is that the macroscale formulation represents spatial variability statistically rather than with explicit spatial patterns. At the King's Creek catchment, with the exception of topography and soil moisture, the degree of spatial variability in soil properties, vegetation properties, and meteorological inputs was not significant enough to yield differences in simulations with the two models. Table 1 shows that nearly 90% of the catchment has similar soil parameters. Most of the catchment is covered by native tallgrass, which showed minimal spatial variability in LAI on unburned prairie [Schimel et al., 1991]. Analysis of the precipitation data showed that many of the spatially distributed images exhibited minimal spatial variability. Comparison of Figures 4 and 3 further suggests that a statistical representation of spatially variable topography, soil moisture, and surface fluxes is adequate at the scale of the King's Creek catchment. This issue is discussed in more detail in paper 3.
In regions of similar scale but higher degrees of spatial variability, or in larger-scale applications where increased spatial variability is encountered, model results may well diverge. For example, although the spatially distributed  Although such an approach is idealized, and there are certainly more detailed methods for modeling large-scale land-atmosphere interaction, it is important to remember the philosophy behind the macroscale formulation. The model simplifies the representation of vertical soil-vegetationatmosphere transfer so that lateral heterogeneity can be incorporated in the model structure without significantly increasing computational complexity relative to currently operational SVATS.

Summary
In this paper the models of Farniglietti and Wood [this issue] were applied at their appropriate scales for evapotranspiration modeling at the FIFE site. FIFE was a largescale field experiment held in the summers of 1987 and 1989, on a 15 x 15 km region of tallgrass prairie located near Manhattan, Kansas. During the experiment, multiscale ground-based and remotely sensed water and energy balance data were collected simultaneously, so that the data set provides unique opportunities to develop and test modeling strategies over a range of spatial scales. All model forcing, parametric, and observed flux data were retrieved from the