In headwater cells, most of the flow is toward the river. However, in low-lying reaches, flow from the river into the aquifer can occur. If sufficient flow into the aquifer occurs, the top of the aquifer can encroach into the vegetationUs root zone. When this occurs, the water is transferred from the aquifer to the soil moisture reservoir, and the leaf-water stress coefficient, GLWP, is set to its no-stress value, 1. The aquifer potentially can become so deep that it floods the surface. In this case, computation of river depth to determine downstream water flow includes the flood-water.

River flow is simulated by a set of differential equations representing each discretized river reach. Within each river element

, (11)

, (12)

Part of our analysis of model output focuses on the water budgets maintained in the CLASP. Guided by Roads et al. [1994] and references therein, Figure 1 shows the budget equations used for each of the physical domains: ATMOS, SVAT and GW/SW. The equations are spatially averaged over the CLASP's drainage basin.

For ATMOS,

(13)

is the amount of atmospheric water vapor in the column. Also in the atmospheric budget,

(14)

We use the set of FIFE observational data from the Konza Prairie Research National Area that was prepared by Betts and Ball [1998] for both calibration (1987 observations) and validation (1988 and 1989 observations). Duan et al. [1996] conclude that FIFE observations in 1987 give a good depiction of the area’s hydrologic cycle. Betts and Ball [1998] express some caution, especially for 1988, when the observations do not close the area’s water budget very well. Although there is some forest in the region at lower elevations along rivers, the observations occur in grassland. In this data set, all available FIFE observations of a field such as surface latent heat flux are averaged together at 30-minute intervals. The averaging makes no distinction between instruments, nor does it use any differential weighting such as area-weighting.

Betts and Ball also give the standard deviation of the reported values at each time step as well as the number of instruments contributing to the average. For most days used, there are typically several instruments reporting values. There are a number of reasons for the range of values among the measurements. Some are due to spatially varying cloud cover [Smith et al., 1992] and physical features of the landscape, such as a site’s terrain-slope orientation [Nie et al., 1992a], specific vegetation characteristics (burned/unburned and grazed/ungrazed grassland)[Simth et al., 1992]. Instrumentation differences enter; surface latent heat flux measurements were made by both eddy-correlation and Bowen ratio instruments, giving slightly different results [Nie et al., 1992b] . Finally, differences were also caused by measurement error [Kanemasu et al., 1992].

We have used the standard deviation among concurrent measurements to give a metric for how accurately surface fluxes have been measured. Thus, our plots of FIFE observations versus time typically show two curves, the instrument average plus/minus the standard deviation. We almost always apply a multi-day running average to flux figures to remove daily fluctuations and highlight climatological behavior. In this instance, we apply the running average to the variance among concurrent measurements.

The primary calibration targets from the FIFE data set are elements of the hydrologic cycle: surface latent heat flux, precipitation and cloud cover. Surface sensible heat flux, near-surface temperature, heat flux into the ground and surface radiation were also included in model-observations comparisons, though they were not specifically targeted for calibration.

Consistent with our reference observations, we assume that the CLASP is simulating a pure grassland in the central United States at (39 N, 96.5 W) with leaf-area index, surface roughness, and surface albedo given in Table 1. The soil is assumed to be a form of silty-clayey loam [NSSC, 1998; Famiglietti et al., 1992]. On this basis, we specify soil density, heat capacity and thermal diffusivity as functions of soil moisture (Table 2) by averaging values given by Garratt [1992] for in clay and sand. We use linear interpolation to specify these properties at soil moistures intermediate to those given in Table 2.

The NGM analysis gives the state of the atmosphere at 00, 06, 12, and 18 UTC. Although there are few upper-air observations available at 06 and 18 UTC, we have found that using analyses for this times is important for the water balance because the 06 UTC analysis will give a depiction of the nocturnal low-level jet, a strong air flow from the Gulf of Mexico into the central U.S. that is important for the region’s hydrologic cycle [Stensrud, 1996]

We have also found that the atmospheric relative humidity in the NGM analyses for this region tends to be low when compared to atmospheric soundings at nearby sites (Fig. 3) . This deficiency appears to be fairly constant throughout the year, with the observations on average having 12% larger specific humidity. As a consequence, for the simulations here, we have increased NGM atmospheric humidity by 12% on input, subject to the restriction that input relative humidity does not exceed 100%.

For the calibration simulations, we use NGM analyses from a three-year period 1 January 1985 — 31 December 1987 to drive the CLASP. In order to arrive at appropriate initial conditions, we cycle through these three years repeatedly until soil moisture changes by less than 1% from the start of one cycle to the next. Aquifer water evolves more slowly than soil moisture, but the focus of calibration is surface-atmosphere coupling contained in the FIFE observations. To simplify this stage of the computation, we allow water to spill from the soil to the aquifer but do not permit any aquifer enchroachment into the root zone and ignore the fate of water once it enters the aquifer. This assumption appears reasonable for the FIFE observations, which have very few valley bottom sites [Smith et al., 1992] and not likely to receive significant soil moisture replenishment from river-aquifer coupling to the root zone.

The years 1985 — 1987 do not show any extreme climatic behavior in atmospheric driving data, so we assume that the state to which the model evolves during spin-up is a representative initial condition. Prior to spin-up, initial coarse calibration was performed through a sequence of 80-day sensitivity simulations spanning 1 July — 8 September 1987. The model was then spun-up to create an initial state assumed to apply to 1 January 1985. Final calibration was performed by running the model from 1 January 1985 through 31 December 1987, using the spun-up state as initial condition for soil and atmospheric water.

Four aspects of model construction emerged as important calibration factors: effective relative humidity to initiate stable (large-scale) condensation, effective length of the growing season, amount of available soil water, and cloud cover parameterization.

(c.1) Precipitation

To initiate condensation, the model’s original treatment was to allow water condensation when relative humidity exceeded 100% at the end of a time step. This resulted in deficient precipitation. Many models (find examples [AMIP?] or modify this statement) use a simple assumption that the effective saturation relative humidity, RH_eff , is less than 100% . The reasoning is that spatial variability in humidity may cause parts of a layer in a gridbox to become saturated even when gridbox-average relative humidity is less than 100%. However, using RH_eff < 100% essentially a correction for bias in the model. The relatively coarse grid of the driving analysis (165 km x 165 km) compared to the FIFE domain (approximately 15 km x 15 km) may be one source of bias because atmospheric convergence will be smoothed somewhat on the coarser grid, resulting in less intense moisture convergence. In addition, atmospheric mass convergence will be less intense, resulting in weaker vertical motion and thus less vertical moisture advection.

Some additional adjustment was made to the convection scheme, where the fraction of precipitation falling outside the convective cloud, CP, may be adjusted (Emanuel, private communication). The parameter CP was reduced to one sixth its original value. This also yeilded an small increase in precipitation. In addition, this adjustment retained an approximately even split between stratiform and oonvective precipitation. While real world precipitation is obviously much more complex than a simple split between two types of precipitation, this partitioning is similar to typical behavior of mesoscale models over the U.S. (e.g., Pan et al. [1998]). Adopted values for RH_eff and CP appear in Table 3.

(c.2) cloud cover

The primary adjustment of the cloud cover parameterization was to adjust the dependence of stratiform cloud cover on relative humidity to be consistent with the onset of stratiform precipitation. Thus, stratiform cloud cover is 100% when relative humidity = RH_eff.

(c.3) Available soil moisture

The model’s original prescription for grassland rooting depth (RD) was 1.25 m, with the maximum available water capacity equal to 13% of RD, or 0.162 m. For the period 28 May — 15 October 1987, when both evapotranspiration E and precipitation P measurements were made, the amount of water lost from the soil was at least (E-P) = 0.185 m, so the original water capacity of the soil was too small versus observations. We adjusted both the rooting depth and maximum available water fraction (AWF) to help give adequate surface latent heat flux. Final values of these parameters appear in Table 3.

(c.4) Effective growing season

One of the conductance coefficients in (4), G_leaf, depends on the near-surface air temperature:

, (15)

where originally T_oo = 0. We have used this function to control green-up and senescence in the model. In the original G_leaf formula, substantial evapotranspiration could occur in spring as temperatures warmed above freezing. Although FIFE surface evapotranspiration observations in 1987 begin only on DATE, model behavior suggested that simulated E was too large in spring, resulting in a soil moisture deficit and weak evaportanspiration later in the summer. We delayed the onset of substantial evapotranspiration in the model by shifting the G_leaf formula toward higher temperatures. Final values were chosen to give adequate surface latent heat flux in mid-summer and appear in Table 3.

(c.5) Canopy reservoir

The canopy reservoir depth is proportional to LAI. We set the reservoir depth to (0.1 mm 4 LAI). The depth/LAI ratio was calibrated to be between two extremes: having a relatively deep reservoir that allows too little water to reach the ground and having a canopy reservoir of negligible depth and, hence, negligible influence. For ratios of 0.1 mm/LAI and smaller, approximately 85% of water from modest to intense storms (precipitation > 10 mm/day) fell through to the surface, consistent with results reported in Dingman (1994), though here the vegetation is prairie grass. Larger ratios yielded much less fall through. The adopted ratio is also the same as used by Bonan (19__) in his detailed SVAT.

(c.6) Calibrated model

Comparisons of the calibrated model versus FIFE observations appear in Figures 4 - 10 and Table 4. Surface latent heat flux from FIFE observations and the calibrated CLASP appear in Fig. 4. The CLASP time evolution falls approximately in the middle of the two FIFE curves and thus reproduces surface latent heat flux during the 1987 period of observations quite well. The bias and temporal standard deviation between the CLASP and FIFE observations (after 31-day running average) are relatively small.

The comparison of observed and simulated precipitation (Fig. 5) shows daily values and thus does not use a running average to smooth time series curves, in order to distinguish frequency of occurrence from daily magnitude. Note that figure shows observed precipitation multiplied by —1 to separate the two curves. The model reproduces the frequency of rainfall events fairly well. Although the model tends to produce more precipitation on the days with greater observed precipitation, it does not reproduce the most intense daily values. This is not surprising, as the analysis providing the model’s boundary conditions has a grid spacing (~ 165 km) that is much coarser than the roughly 15 km width and breadth of the FIFE observation domain. As a consequence the external forcing of precipitation events will be spatially smoothed, and thus somewhat muted, compared to that experienced by the actual atmosphere in the FIFE domain. The bias in CLASP precipitation rate versus FIFE observations is relatively small. The temporal standard deviation between the two is somewhat larger, although smaller than the temporal standard deviation (8.2 mm/d) of the daily observations versus no precipitation.

Daily cloud cover fluctuates rapidly in the observations and the model (Fig. 6), the model generally matching the observations well over most of the period. We do not subject the cloud-cover time series to a running average because of a data gap from day 239 to 257 that considerably reduces the time span allowing averaging.

Other simulated surface energy fluxes (sensible heat, incident solar radiation, downward infrared radiation, heat into the soil) generally compare well will observations. Surface sensible heat flux has acceptable values during the observing period ( Fig. 7) with small bias and standard deviation versus observations (Table 4), although it does not capture a slow, monthly variability seen in the observations. Incident solar radiation at the surface agrees well with FIFE observations ( Fig. 8) except at the end of the year, when it is larger. For both the observations and the simulation, the envelope of daily values (not shown) gradually declines toward the end of the year, and the model’s envelope is nearly coincident with the observation’s. However, over the last 30 days of the year, the simulated daily values tend to lie closer to the envelope than do the observations. This indicates that the model's cloud cover is deficient during this period. The FIFE cloud cover data set unfortunately terminates before this point, preventing a cloud-cover comparison. The simulated infrared radiation from the atmosphere to the surface also compares well with FIFE observations ( Fig. 9). The deleted from the comparison are the last 30 days of the year, when observations become sporadic. Finally, the model's simulation of energy flux into the soil ( Fig. 10 ) agrees well with the FIFE observations except at the end of the observation period, when the model simulates more energy leaving the soil than given by the observations.

(d) Validation

For CLASP validation, we continued the run of the fully calibrated model through 1988 and 1989. The next set of figures were constructed in the same manner as the calibrations figures: FIFE observations are presented as the instrument average given by Betts and Ball (1998) plus or minus the standard deviation of the instrument average. In addition, Tables 5a and 5b give biases and standard deviations of CLASP output from FIFE observations for the periods that the observations are available.

For surface latent heat flux ( Fig. 11), CLASP values are relatively low compared to FIFE observations in 1988 but are close to FIFE observations in 1989. Betts and Ball (1998) note that FIFE observations for the region's water budget do not balance. The imbalance is worse for 1988 than 1987. In addition, even though the FIFE region experienced drought in 1988, the 1988 surface latent heat flux observations are close in magnitude to the 1987 observations. These features of the observation suggest some caution in using them to assess model. Nonetheless, the FIFE data represents a consensus of measurements by instruments in the field, and the model is deficient versus these observations in 1988. However, any deficiency in the 1988 simulation is not permanent, as the model reproduces the 1989 simulations well.

Precipitation simulation for 1988 and 1989 ( Fig. 12 ) reproduces features of the 1987 FIFE-CLASP comparison: the model tends to capture the frequency of wet and dry periods well, but has more difficulty matching the observed magnitude, especially for intense rain events. CLASP bias versus FIFE is acceptable in 1988 (Table 5a)., but is rather large in 1989 (Table 5b). Much of the bias in 1989 is due to very intense precipitation events that the model does not capture. The model's inability to capture this relatively extreme event further underscores the difficulty of reproducing observed precipitation for a relatively small area compared to the resolution of the driving fields.

Cloud cover observations were not available from the Betts and Ball (1998) dataset for 1988 and 1989. Surface fluxes of incident solar radiation (Fig. 13) indicate that the model produces reasonably good cloud cover with bias toward less cloud cover than actually occurred in 1988. Other energy fluxes at the surface (sensible heat flux, downward infrared radiation, heat flux into the soil) compare acceptably well (Figs. 14,15, 16), except that the CLASP sensible heat flux tends to be too large in 1988, consistent with its latent heat flux tending to be too small. As with the latent heat flux, the surface sensible heat flux returns to values similar to the FIFE observations in 1989.

Overall, CLASP does a reasonably good job of simulating climatic behavior of water and energy cycles for the two-year validation period when judged against FIFE observations. Two years after the calibration period, there do not appear to be any persistent trends in the model that cause permanent bias versus observations. Indeed, the model’s surface sensible and latent heat fluxes appear to recover well in 1989 from whatever bias they had in 1988.

Ar river cross-sectional area

Cat atmospheric conductance

Ccan canopy conductance

Cdh surface drag coefficient for sensible heat

Cdm " " " for momentum

Cp specific heat of air, at constant pressure

Cs soil (root zone) heat capacity

Cdeep deep-soil heat capacity

Dat contribution to atmospheric water budget by damping term

Dd downstream discharge from a river reach

Du upstream discharge into a river reach

Fu,Fv,Fq,Fq ATMOS physical processes computed from parameterizations

GHD conductance coefficient for absolute humidity deficit

GLEAF " " " leaf temperature

GLWP " " " leaf-water potential

GSOL " " " net solar radiation

Hdeep deep-soil heat flux

Qriver aquifer's base flow

Rhyd hydraulic radius of river

Rnet net radiative flux at the surface

Ro groundwater recharge

Rp rainfall directly onto a river

Ta temperature of the lowest atmospheric layer

Ts soil-surface temperature

Tdeep deep-soil temperature

Va-r volume of water in combined aquifer-river network

Vriver volume of water in a river element

Wavail available soil moisture

DXriver length of a surface cell's river element

cmax maximum leaf conductance

cp heat capacity of air at constant pressure

ea vapor pressure of the lowest atmospheric layer

es saturation vapor pressure

f Coriolis parameter, = 2Wsin(latitude)

hriver depth of river

nm Manning's roughness coefficient

wriver width of river

za mid-level of the lowest atmospheric layer

zo surface roughness

W rotation rate of the Earth

g psychrometric constant

q atmospheric potential temperature

k R/Cp

r air density

f soil porosity