1. Introduction

Key issues in meteorology, climatology, and resource management deal with water in its various forms in the global water cycle. To investigate the water and energy budgets at various spatial and temporal scales, much research has been conducted on land–atmosphere exchanges in the Global Energy and Water-Cycle Experiment Continental-Scale International Project. Built on detailed surface physical processes, the advanced Oregon State University Land Surface Model (Mahrt and Ek 1984; Marht and Pan 1984; Pan and Mahrt 1987) was implemented by Chen and Dudhia (2001) in the fifth-generation Pennsylvania State University–National Center for Atmospheric Research (NCAR) Mesoscale Model (MM5) to describe land–atmosphere exchanges. Because of relatively large errors in modeled precipitation and unrealistic descriptions of land surface features, numerical models like this are used mostly for sensitivity studies, rather than for detailed investigations of water and energy budgets.

In recent decades, with the advancement of remote sensing technology and data processing, retrieval of water and energy fluxes from land surface models driven by satellite and surface observations has been more successful. For example, Anderson et al. (1997), Mecikalski et al. (1999), and Anderson et al. (2005) detailed a method for evaluating fluxes of sensible and latent heat at the land surface by using the Atmospheric Land Surface Exchange Inverse (ALEXI) model. This model is driven by Geostationary Operations Environmental Satellite (GOES)-derived surface brightness temperature changes and solar insolation, advanced very high resolution radiometer (AVHRR)-derived land cover properties, and synoptic weather and radiosonde data. The fluxes are computed at the end of a given day when GOES observations of surface radiometric temperature become available. Because the ALEXI model is driven primarily by data that can be obtained from remote sensing instruments, clear-sky condition, especially in the morning, is very important for the daily energy flux simulation. In addition, the assumption of a constant evaporative fraction during the day in ALEXI may not hold when weather changes during the day. Furthermore, observations from synoptic weather and radiosonde network are required for the atmospheric correction and model input. However, intensive and frequent radiosonde data are only available during expensive field experiments. Therefore, long-term as well as near-real-time modeling of surface water and energy fluxes at regional or large scale limit the model input to routine observations, which should be easily obtainable without lengthy processing.

In this study, our purpose is to address two major questions that are important for the current continental-scale water cycle study: 1) Can regional- and large-scale evapotranspiration, energy fluxes, and root-zone soil moisture over heterogeneous land surfaces be simulated accurately with routine surface and satellite observations over multiyear time scales and in nearly real time? 2) How can modeled evapotranspiration at the regional scale be evaluated reasonably across multiple years? The parameterized subgrid-scale surface flux (PASS) model, detailed in Song et al. (2000b), does not rely heavily on daily satellite input. It is driven primarily by routine, continuous surface meteorological observations, requiring remote sensing composite normalized-difference vegetation index (NDVI) values only at biweekly intervals. Thus, the PASS model enables the estimation of water and energy fluxes under all possible weather conditions.

Root-zone available soil moisture plays a key role in surface hydrological balances because it directly influences the surface evapotranspiration rate. Variations in evapotranspiration and runoff into streams and rivers cannot be fully assessed without knowledge of root-zone available moisture. However, obtaining accurate estimates of this quantity over large terrestrial areas can be difficult, because large temporal and spatial variability results from the unevenness of precipitation and the diversity of vegetation. Hirabayashi et al. (2001) proposed a method for estimating root-zone moisture content from surface soil moisture, which in turn is estimated by microwave remote sensing in the Global Soil Wetness Project. However, values need to be corrected for time lag and intensity of precipitation. Alternatively, the PASS model can be applied directly to estimate spatial and temporal variations in both evapotranspiration rate and root-zone available moisture (Song et al. 2000b).

To address the water cycle, the U.S. Department of Energy (DOE) initiated a 3-yr pilot study in the Walnut River Watershed (WRW), which encompasses an area of about 5000 km² in southeastern Kansas. The Whitewater Subbasin lies in the northwest region of the WRW. Work at the Argonne National Laboratory, one of five primary participants in the pilot study, included simulation of root-zone available moisture content and evapotranspiration in the WRW with the PASS model over multiyear period. In contrast, most short-term field experiments and modeling activities, such as the First International Satellite Land Surface Climatology Project (ISLSCP) Field Experiment (FIFE)-87 (Sellers et al. 1988); Cooperative Atmosphere–Surface Exchange Study (CASES)-97 (LeMone et al. 2000; Song et al. 2000b; Yates et al. 2001); Southern Great Plains (SGP-97) Hydrological Experiment (French et al. 2001), International H₂O Project (IHOP2002) (Weckwerth et al. 2004); Soil Moisture Atmospheric Coupling Experiment (SMACEX02; Anderson et al. 2003), have concentrated on the summer growing season, which lasts for only a few months, and at selected environmental conditions. Surface models verified with data from intensive experiments might perform less well in other seasons and during prolonged drought or wet conditions. The Atmospheric Boundary Layer Experiments (ABLE) site in southern Kansas, the first long-term site supported by DOE for studies of regional-scale water and energy budgets, has enabled the current multiyear modeling study of the water and energy budgets. One goal for long-term modeling of surface hydrological components at regional scales or at a grid scale suitable for high-resolution global climate models is to describe seasonal variations in surface hydrological variables accurately, even when the surface is spatially very heterogeneous. Another goal is to resolve the subgrid-scale variability of precipitation, soil moisture, and vegetation to the extent that biases are not introduced by the scheme of aggregating to computationally manageable grid cell sizes. To address these goals and to aid in the study of the interannual variability of key surface hydrological components, research continues on the ability of the PASS model to simulate available soil moisture and evapotranspiration in the WRW over the seven-year period of 1996–2002.

Field observations of soil moisture content are typically too limited to provide the spatial resolution and coverage required to describe the spatial heterogeneity of soil moisture adequately over extended areas. Song et al. (1997) modeled the influence of heterogeneous soil moisture on latent and sensible heat fluxes and found that simulated regional-scale latent heat fluxes tend to be higher and air temperatures lower under uniform surface conditions than under spatially heterogeneous conditions. Pitman et al. (1993) demonstrated that possible biases associated with the under representation of regional land surface heterogeneity within climate models might explain the propensity of climate models to overestimate grid-cell evapotranspiration and to underestimate runoff. This suggests that detailed information on the spatial distribution of the surface conditions appears to be necessary to simulate evapotranspiration accurately over extended areas. However, estimating surface fluxes at resolution higher than that of the climate model's grid cell can be difficult, because remote sensing data emphasize local land surface conditions, while the surface fluxes can be strongly influenced by surface–atmosphere interactions over substantially larger areas (Friedl 1996). The PASS model overcomes these limitations by coupling NDVI from satellite AVHRR data (at the climate model's subgrid-scale resolution) with surface meteorological observations in the study region (at the climate model's grid-scale resolution) to infer surface temperature, root-zone available moisture content, and evapotranspiration (at the climate model's subgrid scale, which is equivalent to the pixel resolution of the satellite used here).

The PASS model couples spatially sparse yet continuous data from meteorological stations with temporally sparse yet spatially highly resolved satellite remote sensing data to retrieve updated surface hydrological components over extended areas. Heterogeneities in surface conditions are spatially resolved to an extent determined primarily by the satellite data pixel size. Previous studies evaluated the model's ability during a relatively short intensive observation period (Song et al. 2000a, b). The present study evaluated long-term (multiyear) simulation of surface hydrological components by PASS modeling with continuous input of data from satellites, radars, and meteorological stations.

2. The PASS model

The original version of PASS model was developed by Gao (1995) and was applied by Gao et al. (1998). More recently, the model was improved by Song et al. (2000a, b) and evaluated with observations from an intensive field experiment. The PASS model uses a computationally efficient algorithm for computing subgrid-scale surface energy partitioning on the basis of an analytical solution to surface energy budget equations. An efficient computing algorithm is needed for long-term studies because of the large number of pixels associated with the satellite, land use, and soil characteristics data. A detailed description of the PASS model was provided by Song et al. (2000a, b). Here, the approach is briefly outlined, and additional requirements for long-term simulation are described.

All variables in the model are described at two scales: the regional scale of at least 100 km (a climate model's grid scale) and the satellite pixel scale (the climate model's subgrid scale). The regional-scale variables are approximated by the average of the observations made at the surface meteorological stations. Values of satellite pixel-scale surface parameters, including roughness length, surface albedo, surface conductance for water vapor, and the ratio of soil heat flux to net radiation, are estimated according to satellite-derived spectral indices and land use classes. These relationships contain empirical coefficients whose values have been derived for midlatitude areas with surface vegetation dominated by grasslands and agricultural crops. To account for the feedback of locally influenced meteorological conditions on the local atmosphere–surface exchange, pixel-specific near-surface meteorological conditions such as air temperature, vapor pressure, and wind speed are adjusted from their corresponding values at the regional scale according to local surface forcing as described below.

3. Site and data description

Located in southern Kansas, east of Wichita, the WRW is a rectangular region having an area of about 5000 km² within the Atmosphere Radiation Measurement (ARM) Climate Research Facility (Fig. 1). The WRW is representative of a global climate model's grid cell with heterogeneous land cover types. In addition to the large terrain gradient shown in Fig. 1, annual rainfall in the WRW varies from 76 cm in the west to 86 cm in the east, and average temperature varies from 0°C in the winter to 26°C in the summer (www.kwo.org/Org_People/walnut.htm). Land cover in the WRW is primarily a mix of grassland and cropland, with the cropland mostly in the western half of the basin. In addition to a few small scattered towns, Wichita suburbs expand into the western WRW. The Walnut River is the major stream, and the Whitewater River in the northwest is one of its tributaries. El Dorado Lake and Winfield Lake are major reservoirs on the river system. Surface water use for municipal, irrigation, recreational, and industrial purposes accounts for 91% of the water use in the basin (www.kwo.org/Org_People/walnut.htm). The WRW is a quasi-closed basin amenable to computing the components of the hydrological budget, and it is a tight watershed with minimum leakage into the substrata (LeMone et al. 2000). Nested within the northwest portion of the WRW is the 35 km × 30 km Whitewater Watershed (WW). These domains were selected for detailed observations and hydrological studies because of information provided by previous studies (LeMone et al. 2000).

4. Results of surface hydrological modeling

The PASS modeling was performed over a rectangular domain 73 km × 109 km, encompassing the WRW. The model resolution was 1 km (Fig. 1). Modeled outputs at each pixel within the WRW were selected for analysis. Spatial distributions of modeled evapotranspiration in 1996–2002 are displayed in Fig. 3. Within the WRW, modeled evaporative loss was largest above water surfaces (darkest areas) and smallest over urban and suburban (lightest) areas. Evapotranspiration in the WRW is very heterogeneous spatially, even when values are summed for each year. Table 2 summarizes annual values of modeled evapotranspiration and runoff for the total domain, as well as observed precipitation and streamflow for the WRW and subbasin WW. Evapotranspiration was largest in 1999, followed by 1997, and was smallest in 2002.

Seasonal variations in modeled evapotranspiration and root-zone available moisture are shown with observed precipitation in Fig. 4. The modeled root-zone available moisture shows an annual wet and dry cycle. In the summer, the large loss from evapotranspiration exceeds the moisture input from precipitation corresponding to the growing season. The wet periods at the end of each year reflect the reduction in evapotranspirative loss and the recharge from precipitation. An exception is the year 2001, when precipitation recharge was minimal after the growing season. In each year, precipitation was unevenly distributed; most rain fell in the warm growing season, except in 2002 (Fig. 4). The low modeled total evapotranspiration in 2002 is due to the timing of the rainfall as well as lower NDVI. Precipitation events in 2002 occurred mostly in spring and fall, rather than in the summer peak growing season (Fig. 4). With drier soil, vegetation did not grow well, as indicated by lower NDVI values, and thus evapotranspiration was limited. Similarly, in 1996, modeled evapotranspiration was next to lowest, because a prolonged dry period in the spring and early summer influenced both soil moisture and vegetation growth. In comparison, the largest modeled evapotranspiration, in 1999, corresponded to abundant precipitation during the entire growing season. In 1998, although annual total precipitation was greater than in 1999, modeled evapotranspiration was lower (Table 2), because relatively less rain fell during the summer growing season, leading to lower soil moisture and lower NDVI. Although the heaviest precipitation occurred later in the year 1998, evapotranspiration did not increase much after the growing season.

To evaluate modeled evapotranspiration over the extended heterogeneous region, we compared modeled runoff with the stream-gauge-measured streamflow (Fig. 5). Modeled runoff at each pixel was estimated as the excess soil water over capacity, and the values were summed for both the WRW and the WW subbasin. The Winfield stream gauge measures runoff integrated over the whole WRW; the Towanda stream gauge measures flow from the WW subbasin (Fig. 1). Figure 5 shows generally good correlation between modeled runoff and measured streamflow. At times, when the model assumptions are best satisfied, agreement is excellent. An example is in 1998, when a large amount of precipitation saturated the soil. At other times the agreement is not so good. An example of this is in early 2000 and 2001, when some of the precipitation might have occurred as snow. Modeled runoff exceeds observed streamflow in some of the high-flow periods but is less than the observed value during some low-flow periods. This may be explained by reservoir regulation and storage pond and is examined further below. The overall correlation values between modeled runoff and observed streamflow for the WRW and the WW are 0.91 and 0.89, respectively, with significance at 0.05.

Over the seven-year period shown in Table 2 for both the WRW and the WW, precipitation amounts varied greatly from dry years (e.g., 1996) to wet years (e.g., 1999). The modeled water loss due to evapotranspiration in the WRW varied from 66% (1998, a wet year) to 88% (1996, a dry year) of precipitation at the end of each year, which is reasonable for southern Kansas. The differences between the observed discharge and modeled runoff are small in comparison with precipitation amount. Modeled runoff amounts exceeded observed streamflow in years when annual precipitation was relatively low, such as 1996, 2001, and 2002. Modeled runoff was less than observed streamflow in years when annual precipitation was relatively high, such as 1998 and 1999. At the end of the seven-year period, the overall modeled runoff in the WRW was about 3% more than observed streamflow (5% for the WW subbasin).

Runoff comparisons can provide an evaluation of the regional-scale water budget, and comparisons with available field measurements can provide evidence for model performance at the pixel or local scale. In addition, the possibility that underestimation in modeled evapotranspiration resulted in increased modeled runoff during years 2000–2002 can be evaluated with in situ observations. Modeled latent heat fluxes and net radiation at the pixels nearest the WH and SM sites are compared in Fig. 6 with corresponding in situ observations during 2000–2002. Results indicate that modeled net radiation and latent heat fluxes are similar to the observations. On average, modeled net radiation is 11 (WH) and 24 W m⁻² (SM) below the observed values, but modeled latent heat fluxes are only 2 (WH) and 8 W m⁻² (SM) more than the observed values. Modeled sensible and ground heat fluxes are below the corresponding observations (data not shown), with average differences of 16 (WH) and 3 W m⁻² (SM) for sensible heat and 1 (WH) and 2 W m⁻² (SM) for ground heat.

The above evaluation of in situ latent heat fluxes indicates that the reason for greater modeled runoff than observed streamflow during years 2000–02 was not due to the underestimation of evapotranspiration within the watershed. In fact, evapotranspiration may have been overestimated slightly when biweekly composite NDVI was used instead of daily NDVI. To explain the possible cause of less observed than modeled runoff during the peak flow periods shown in Fig. 5, we further examined surface water usage in the WRW and found that water stored in the two major reservoirs in the WRW during the high-flow episodes was later withdrawn for municipal and irrigation purposes (Miller et al. 2005). The larger El Dorado Reservoir (Fig. 3) especially dominates water flow in the lower section of the Walnut River, because the river is fed primarily by water released from the reservoirs. At a smaller scale, collection of detailed sets of precipitation, stream, and soil water samples from a small tributary of the Whitewater River (the Rock Creek drainage basin) in May of 2002 (Miller et al. 2005) showed that the upper 30% of the drainage basin feeds into a stock pond that limits flow to the lower part of the creek to periods following precipitation events. Thus, in addition to snow accumulation, which can delay surface runoff in the spring, runoff storage in reservoirs and stock ponds in the upper part of the river drainage can cause observed streamflow to be less than modeled flow during a high-flow period but more than modeled flow during a low-flow period, as Fig. 5 shows. In addition, groundwater recharge is not modeled or measured, leaving the accounting for water loss incomplete.

Interannual and seasonal variations in modeled root-zone available moisture have shown reasonable trends within the WRW. Evaluation is problematic because of a lack of observations at root-zone depth, although near-surface (at 2.5 cm) soil moisture has been measured at the WH and SM sites to account for heat storage for a ground heat flux correction. As a result, large differences are expected when comparisons are made between modeled root-zone available moisture at the pixels nearest the WH and SM sites and in situ surface soil moisture measurements (Fig. 7). The first reason for the difference between modeled and observed values is that unlike the observed total soil moisture content, modeled values are available moisture content, which does not include soil water that is not accessible at wilting point. In contrast, the observed values are not limited to available soil water. Second, modeled values represent the root-zone condition, which tends to be less extreme than surface conditions and does not respond to dryness and wetness as quickly as the observed surface condition. Third, because heterogeneity in soil texture and soil moisture is large, sampling even at five points within an area of about 4 m² at each site might not adequately represent root-zone available moisture modeled at 1-km resolution. Fourth, heterogeneous soil texture requires calibration functions for moisture sensors to be generated by a field scientist for each site. Soil at SM soil is silty clay (bulk density 1.22, porosity 54%), while soil at WH is clay loam (bulk density 1.33, porosity 50%). The soil moisture values are adjusted according to nonlinear functions fitted to collected soil samples in a limited moisture range. Despite the difference between the modeled and observed values, temporal variations in the monthly averaged measurements shown in Fig. 7 are similar to those in modeled root-zone values, which sometimes lag behind the measured surface values because response is slower at root-zone depth.

5. Summary and conclusions

Estimates of evaporative water loss and root-zone available moisture in the WRW and WW were made with the PASS model and applied in an evaluation of hydrological balance components for 1996–2002. Satellite remote sensing data in biweekly composite, 1-km-resolution NDVI data products were used to describe the surface vegetative conditions. In addition, 4-km-resolution radar-based estimates of precipitation constituted major input. Surface radiation and basic meteorological data provided the driving force for modeling evapotranspiration.

The small differences between modeled runoff and observed streamflow, which is less than 5% for the seven-year period, indicates that the regional-scale surface evapotranspiration modeled by PASS is realistic, with errors of 10% or less. Seasonal changes in modeled evapotranspiration in years 2000–02 matched fairly well with the in situ flux measurements, with slight overestimates during certain seasons. Overall, modeled water and energy fluxes closely reflected observed spatial and long-term temporal variations in the region. The results suggest that a highly parameterized but relatively simple surface model like PASS can be exercised efficiently to estimate both real-time and long-term surface energy fluxes and hydrological components. Moreover, with its subgrid-scale resolvability, PASS can provide scale linkage for evaluating each grid cell of global climate models with regional streamflow measurements and field observations. Work continues on selection of proper root-zone depths for various types of vegetation.