A Multiscale Remote Sensing Model for Disaggregating Regional Fluxes to Micrometeorological Scales

APRIL 2004 343A N D E R S O N E T A L .

q 2004 American Meteorological Society

A Multiscale Remote Sensing Model for Disaggregating Regional Fluxes toMicrometeorological Scales

MARTHA C. ANDERSON AND J. M. NORMAN

Department of Soil Science, University of Wisconsin—Madison, Madison, Wisconsin

JOHN R. MECIKALSKI

Space Science and Engineering Center, University of Wisconsin—Madison, Madison, Wisconsin

RYAN D. TORN

Department of Atmospheric Sciences, University of Washington, Seattle, Washington

WILLIAM P. KUSTAS

Hydrology and Remote Sensing Laboratory, USDA-ARS, Beltsville, Maryland

JEFFREY B. BASARA

Oklahoma Climatological Survey, Norman, Oklahoma

(Manuscript received 18 June 2003, in final form 9 October 2003)

ABSTRACT

Disaggregation of regional-scale (103 m) flux estimates to micrometeorological scales (101–102 m) facilitatesdirect comparison between land surface models and ground-based observations. Inversely, it also provides ameans for upscaling flux-tower information into a regional context. The utility of the Atmosphere–Land ExchangeInverse (ALEXI) model and associated disaggregation technique (DisALEXI) in effecting regional to localdownscaling is demonstrated in an application to thermal imagery collected with the Geostationary OperationalEnvironmental Satellite (GOES) (5-km resolution) and Landsat (60-m resolution) over the state of Oklahomaon 4 days during 2000–01. A related algorithm (DisTrad) sharpens thermal imagery to resolutions associatedwith visible–near-infrared bands (30 m on Landsat), extending the range in scales achievable through disag-gregation. The accuracy and utility of this combined multiscale modeling system is evaluated quantitatively incomparison with measurements made with flux towers in the Oklahoma Mesonet and qualitatively in terms ofenhanced information content that emerges at high resolution where flux patterns can be identified with rec-ognizable surface phenomena.

Disaggregated flux fields at 30-m resolution were reaggregated over an area approximating the tower fluxfootprint and agreed with observed fluxes to within 10%. In contrast, 5-km flux predictions from ALEXI showeda higher relative error of 17% because of the gross mismatch in scale between model and measurement, high-lighting the efficacy of disaggregation as a means for validating regional-scale flux predictions over heterogeneouslandscapes. Sharpening the thermal inputs to DisALEXI with DisTrad did not improve agreement with obser-vations in comparison with a simple bilinear interpolation technique because the sharpening interval associatedwith Landsat (60–30 m) was much smaller than the dominant scale of heterogeneity (200–500 m) in the scenesstudied. Greater benefit is expected in application to Moderate Resolution Imaging Spectroradiometer (MODIS)data, where the potential sharpening interval (1 km to 250 m) brackets the typical agricultural field scale. Thermalsharpening did, however, significantly improve output in terms of visual information content and model con-vergence rate.

1. IntroductionThe set of instruments deployed on the current fleet

of satellite remote sensing platforms provides unique

Corresponding author address: M.C. Anderson, 1525 ObservatoryDr., University of Wisconsin, Madison, WI 53706.E-mail: [email protected]

possibilities for monitoring the biotic and hydrologiccondition of terrestrial ecosystems and their response toenvironmental changes. In particular, these instrumentsallow mapping of evapotranspiration (ET) and moisturestress at a wide range in spatial resolution, coveringareas from local to regional to global scales. The breadthof resolution afforded by these combined instruments

344 VOLUME 5J O U R N A L O F H Y D R O M E T E O R O L O G Y

FIG. 1. Schematic demonstrating the DisALEXI flux-disaggregation approach. Regional-scale flux predictions at 5-km resolution over the state of Oklahoma are disaggregated to 30-m resolution at grid cells coincident with surfaceflux towers in the OASIS network. The disaggregated fluxes are reaggregated within the source footprint associatedwith each tower (white cross) and can be compared directly to observed fluxes.

is crucial to engendering a broad understanding of landsurface response. High-resolution (101 m) data can helpto pinpoint the cause of remotely sensed phenomena,while lower resolution (104 m) imagery are better suitedfor assessing impacts over large areas.

Norman et al. (2003) presented a nested flux-disag-gregation technique that has demonstrated utility in in-tegrating multiscale remote sensing data to effect ups-caling and downscaling in surface fluxes (see Fig. 1).Disaggregation provides a potential means for bridgingextant gaps in spatial scale that have traditionally posedchallenges to surface-flux-monitoring programs: phys-ical gaps encountered in upscaling ground observations

to landscape and regional scales (or, equivalently, down-scaling regional model predictions to the ground-ob-servation footprint scale) and logistical gaps in coverageprovided by current satellites, which have either high-temporal/low-spatial resolution or low-temporal/high-spatial resolution.

New-generation flux-monitoring campaigns are rec-ognizing the need to establish connectivity between in-tensive surface observations, typically collected at apoint or a distributed set of points, by placing themwithin a regional context using modeling supported byremote sensing (Running et al. 1999). The North Amer-ican Carbon Program, for example, is adopting a mul-


TABLE 1. Nominal satellite pixel resolutions for Vegetation Indices(VIs) and radiometric surface temperature (TRAD) and approximatecoverage repeat interval.

Instrument/satellite

VI pixelresolution

(m)

TRAD pixelresolution

(m)Coverage repeat

interval

ASTER* (Terra)AVHRRGOESLandsat-5Landsat-7MODIS (Terra, Aqua)

1511001000

3030

250

9011004000

12060

1000

On demand1 day15 min16 days16 daysTwice daily

* Advanced Spaceborne Thermal Emission and Reflection Radiom-eter.

titiered sampling scheme (Wofsy et al. 2002) in whichgeneral surface biophysical and land-use properties willbe monitored over continental scales using remote sens-ing, with more intensive sampling of carbon stores andfluxes occurring on the ground at sites selected to rep-resent the endemic range in spatial variability. Large-scale carbon flux networks, such as AmeriFlux (Bal-docchi et al. 2001) and EuroFlux (Valentini 2003), willrequire robust methodologies for upscaling and inte-grating observations made at individual towers to beable to draw regional inferences regarding terrestrialcarbon cycles (Fan et al. 1998).

From a modeling standpoint, the regional-scale soil–plant–atmosphere models required to support these ups-caling efforts are notoriously difficult to validate be-cause of the gross scale mismatch between the modelgrid cell (typically 10–100 km) and the surface footprintsampled by a ground-based micrometeorological tower(0.1–1 km). The problem of validation is exacerbatedover heterogeneous landscapes, such as those depictedin Fig. 1, where great care must be taken to select rep-resentative ground sampling sites. Improved estimatesof mean surface flux conditions can be obtained by av-eraging fluxes sampled at multiple locations within amodel grid cell (e.g., Doran et al. 1998; Gao et al. 1998)or along an aircraft transect (Desjardins et al. 1997;Oechel et al. 1998; Mahrt et al. 2001; Kustas et al.2001b; Chen et al. 2003); however, such measurementstypically require large investments of time and resourcesand therefore cannot be widely implemented. Alterna-tively, methods capable of disaggregating regional fluxestimates down to the subkilometer scale provide ameans for quantitative validation against observationsmade at individual towers.

Large-scale remote sensing projects are also con-fronted with the temporal/spatial coverage trade-off in-herent in the current suite of earth-observing satellites.While low-spatial-resolution sensors like the Geosta-tionary Operational Environmental Satellite (GOES)can provide good temporal coverage (15 min), high-spatial-resolution instruments, as on the Land RemoteSensing Satellite (Landsat), tend to have more limitedtemporal coverage (biweekly) and yield little informa-tion about diurnal variability in land surface behavior.In addition, the thermal infrared (TIR) bands, which canprovide important information about surface moisturestatus, are typically instrumented at significantly lowerspatial resolution than are the visible–near-infrared (vis–NIR) bands, which can be used to estimate vegetation-cover amount (see Table 1). For many applications, syn-ergistic use of both high-temporal-resolution and high-spatial-resolution datasets in multiple wave bands canbe advantageous.

Methods have been developed for combining high-resolution vis–NIR and TIR remote sensing data to de-duce the surface energy balance, each with strengthsand weaknesses in terms of routine application. Somerely on the availability of contemporaneous in situ mea-

surements, primarily near-surface meteorological con-ditions such as air temperature, wind speed, and hu-midity, and are therefore difficult to implement opera-tionally over large regions (Gardner et al. 1992; Choud-hury et al. 1994; Moran et al. 1994, 1996). Othersrequire that a sufficient range in vegetation cover andsurface temperature (Gillies and Carlson 1995; Jiangand Islam 2001) or hydrologic (dry–wet) conditions(Bastiaansen et al. 1998a,b) be present within the sceneso that end points in the temperature–vegetation-coverrelationship can be defined.

The nested-scale flux-modeling system introduced byNorman et al. (2003) uses high-resolution thermal, vis-ible, and near-infrared imagery to disaggregate 5-kmflux estimates from the regional Atmosphere–Land Ex-change Inverse (ALEXI; Anderson et al. 1997) modeldown to the scale of a surface flux-tower footprint (101–102 m) without the need for any locally made mea-surements or restrictions regarding in-scene variability.This procedure [Disaggregated ALEXI (DisALEXI)]was applied to remote sensing data collected by aircraftat 24-m resolution over a 4-day dry down during the1997 Southern Great Plains field experiment (SGP97).Footprint-weighted latent heat flux estimates agreedwith eddy covariance measurements made at four fluxtowers to within 12% and captured temporal effects ofthe dry-down event (Norman et al. 2003).

As an ancillary enhancement to the DisALEXI tech-nique, Kustas et al. (2003) describe an algorithm thatuses the physical relationship between surface temper-ature and vegetation cover to sharpen TIR remote sens-ing imagery to higher resolutions associated with vis–NIR bands (Table 1). While the interband resolutiondifference is only a factor of 2 for Landsat 7 (30 versus60 m), the discrepancy for the Moderate Resolution Im-aging Spectroradiometer (MODIS), which provides bet-ter temporal coverage, is significant (250 m versus 1km). Townshend and Justice (1988) argue that spatialresolutions of 250–500 m are required to effectivelymonitor disturbances in surface energy balance due toland-use and land-cover changes; thus, methodologiesfor improving important remote sensing inputs to suchresolutions will be of value. Kustas et al. (2003) applied


FIG. 2. Schematic diagram representing the coupled ALEXI–DisALEXI modeling scheme, highlighting fluxes of sensible heat (H) fromthe soil and canopy (subscripts ‘‘c’’ and ‘‘s’’) along gradients in temperature (T ) and regulated by transport resistances Ra (aerodynamic),Rx (bulk leaf boundary layer), and Rs (soil surface boundary layer). DisALEXI uses the air temperature predicted by ALEXI at 50 m AGL(Ta) to disaggregate 5-km ALEXI fluxes, given vegetation cover [ f (f)] and directional surface radiometric temperature [TRAD(f)] informationderived from high-resolution remote sensing imagery. See Norman et al. (2003) for further details.

this sharpening algorithm (called DisTrad) to aircraft-acquired thermal imagery from SGP97, artificially de-graded in resolution to simulate various satellite-bandcombinations listed in Table 1. DisTrad reconstructedhigh-spatial-resolution structure found in the native-res-olution thermal imagery with good accuracy.

In this paper, the DisALEXI and DisTrad techniqueswill be applied to satellite data collected with the En-hanced Thematic Mapper Plus (ETM1) instrument onLandsat 7. Four Landsat scenes from 2000–01 are ex-amined, each containing eddy covariance towers asso-ciated with the Oklahoma Mesonet (Brock et al. 1995).In addition to using satellite (as opposed to aircraft) data,this experiment covers a wider range of climatic andphenological conditions than those considered by Nor-man et al. (2003) and Kustas et al. (2003), who focusedon 4 consecutive days in a small area around El Reno,Oklahoma. Disaggregated fluxes are compared toground-based eddy covariance measurements, and spa-tiotemporal patterns in the distributed flux predictionsare evaluated for reasonability. This study also examinesthe benefits of the DisTrad thermal-sharpening algo-rithm in terms of elucidating small-scale structure inmodeled surface flux distributions and improving modelbehavior. The paper begins by briefly reviewing the AL-EXI, DisALEXI, and DisTrad models.

2. Model descriptionsThe ALEXI/DisALEXI modeling framework, dia-

grammed schematically in Fig. 2, is intended for diverse,

routine applications and therefore attempts to balancethe competing demands of generality and simplicity.The models have been designed to accommodate vary-ing surface conditions while remaining computationallyinexpensive and requiring only a tractable array of sur-face parameters that can be estimated remotely. To im-prove robustness, they are built where possible arounddifferential (as opposed to absolute) radiometric signals(Kustas et al. 2001a).

a. The two-source model

At the core of this modeling system is a two-source(plant and substrate), land surface representation cou-pling conditions inside the canopy to fluxes from thesubstrate (typically soil), plants, and atmosphere. Thetwo-source model (TSM; Norman et al. 1995b; Kustasand Norman 1999a,b, 2000) partitions the compositedirectional radiometric temperature [TRAD(f)] of a het-erogeneous scene into soil and canopy contributions (Ts

and Tc) given an estimate of fractional vegetation coverapparent at thermal view angle f [ f (f)]:

4 4 4T (f) 5 f (f)T 1 [1 2 f (f)]T .RAD c s (1)

Equation (1) is solved along with a system of surfaceenergy balance equations:

RN 5 H 1 LE 1 G,

RN 5 H 1 LE 1 G,s s s

RN 5 H 1 LE , (2)c c c


where RN is net radiation; H, LE, and G are fluxes ofsensible, latent, and soil heat conduction, respectively;and the subscripts ‘‘c’’ and ‘‘s’’ represent fluxes fromthe soil and canopy components of the scene. Net ra-diation is estimated from modeled or measured down-welling shortwave and longwave radiation (upwellingcomponents are modeled by the TSM), while G is re-lated to the net radiation just above the soil surface(RNs), following Norman et al. (2000). Given specifi-cation of upper boundary conditions in air temperatureand transport resistances based on wind speed, vege-tation cover, and surface roughness, Hc and Hs are com-puted assuming the series resistance network in Fig. 2.A modified Priestley–Taylor (1972) relationship pro-vides an initial estimate of canopy evapotranspiration(LEc), and the soil evaporation rate (LEs) is computedas a residual. If LEs is negative during the day (con-densation onto the soil), this is taken as a sign of mois-ture stress, and LEc is throttled back.

The two-source representation is a major improve-ment over previous single-layer thermal models that re-quired site-specific adjustments to compensate for dif-ferences in aerodynamic coupling between the soil, can-opy, and atmosphere (Kustas et al. 1989; Hall et al.1992; Stewart et al. 1994; Kubota and Sugita 1994).The TSM also provides a means for accommodating thedependence of apparent surface temperature on viewangle [Eq. (1)], caused by the variable obscuration ofthe underlying bare soil when a canopy is viewed offnadir (Vining and Blad 1992). This feature is critical toremote sensing applications, where a scene may beviewed from many different angles.

Whereas lower boundary conditions are supplied bythermal remote sensing data, the TSM requires speci-fication of temperature above the canopy and is partic-ularly sensitive to biases in this input with respect tothe TIR reference (Zhan et al. 1996; Anderson et al.1997; Kustas and Norman 1997). Just a 1-K error inthe assumed surface-to-air temperature difference cantranslate into errors in predicted sensible heating of upto 100 W m22 (Norman et al. 1995a). Because shelter-level atmospheric properties can be strongly coupled tolocal surface fluxes, these upper boundary conditionsgenerally cannot be interpolated with adequate accuracyfrom synoptic weather network observations, with a typ-ical spacing of 100 km. For regional-scale flux-mappingapplications, the TSM has been coupled with a simplemodel of atmospheric boundary layer (ABL) develop-ment (McNaughton and Spriggs 1986) so that near-sur-face air temperatures are simulated and consistent withmodeled fluxes.

b. ALEXI

The Atmosphere–Land Exchange Inverse model is acoupled TSM–ABL model (Anderson et al. 1997; Me-cikalski et al. 1999). The lower boundary conditions forALEXI are provided by TIR observations taken at two

times during the morning, separated by about 4 h, froma geostationary platform such as GOES. The ABL mod-el relates the modeled rise in air temperature above thecanopy during this interval and the resulting growth ofthe ABL to the time-integrated influx of sensible heatingfrom the surface. This results in a prediction of spatiallydistributed fluxes at the time of the second surface tem-perature observation, about an hour before local noon.

Input data required for regional application of theALEXI model are listed in Table 2 and reviewed byMecikalski et al. (1999). Use of time-differential TIRmeasurements reduces model sensitivity to errors in ab-solute temperature due to sensor calibration and at-mospheric and surface emissivity corrections. Impor-tantly, the air temperature in the surface layer is notdefined as a boundary condition—it is evaluated by themodel at the TSM–ABL interface and responds to feed-back from both the surface fluxes and the atmosphericprofile. Sensitivity studies with ALEXI have shown thatthe modeled air temperature adjusts to instrumental bi-ases in the TIR data, tending to preserve the true surface-to-air temperature gradient (Anderson et al. 1997).

Using local-scale TIR measurements made withground-based infrared thermometers, which sample anarea on the order of 10 m 3 10 m, ALEXI flux pre-dictions agree to within 20% with tower measurementsmade over a variety of land-cover types including corn,soybean, tallgrass prairie, sparse desert shrubs, andrangeland (Anderson et al. 1997, 2003). In practice,however, ALEXI is more suitably applied to satellite-based thermal data acquired at the 5–10-km scale—thescale at which organized land surface behavior becomeseffective in influencing mean atmospheric conditionsand driving boundary layer growth. To compare pre-dictions at this scale with ground-based flux measure-ments, the regional-scale model fluxes need to be spa-tially disaggregated.

c. DisALEXI

The DisALEXI algorithm (Norman et al. 2003) is atwo-step process (Fig. 2): first, ALEXI is executed at aresolution of 5 km to estimate the air temperature at theinterface between the land surface and ABL submodels[;50 m above ground level (AGL)]; then, the TSM isapplied to high-resolution surface temperature and veg-etation-cover data (from aircraft or satellite), holdingthe air temperature at 50 m constant at the ALEXI-derived value. To ensure the ALEXI-derived verticaltemperature gradient is preserved in the disaggregationstage, the high-resolution radiometric temperature fieldis corrected for biases (due to differences in sensor band-width, calibration, and/or atmospheric corrections) toyield an average consistent with the 5-km temperaturedata used in ALEXI.

The resulting high-resolution flux predictions canthen be reaggregated and compared directly with towerobservations, assuming some weighting function de-


scribing the surface source footprint contributing to theflux detected at the height and location of the sensor.Current techniques for estimating flux footprint distri-butions fall into two major categories: analytical solu-tions to the diffusion equation and Lagrangian models,which numerically simulate air parcel trajectories fromsource to detector. Analytical models (e.g., Schuepp etal. 1990; Horst and Weil 1992; Schmid 1994) use var-ious simplifying approximations regarding turbulenceparameterization and source variability but are relativelyeasy to implement. The more rigorous Lagrangian tech-niques (e.g., Leclerc and Thurtell 1990; Finn et al. 1996)require detailed specifications of the turbulence field,which may not be available in many applications.

The DisALEXI approach, like other ‘‘tile’’ or ‘‘mo-saic’’ land surface models, assumes that horizontal flux-es are small in comparison with vertical fluxes and thatconditions at 50 m AGL [approximating the atmospheric‘‘blending height’’ (Wieringa 1986; Mason 1988) aremore or less uniform over scales of 5 km. Both as-sumptions may be violated in strongly heterogeneouslandscapes, where advection and small-scale surface–atmospheric coupling can become significant. The im-pacts of these assumptions under varying states of het-erogeneity, and possible operational correction tech-niques, are being investigated in large-eddy simulationswith an embedded TSM land surface representation (Al-bertson et al. 2001; Kustas and Albertson 2003).

d. Thermal sharpening

The ALEXI/DisALEXI models use remote sensinginformation from both the TIR (surface temperature)and vis/NIR (vegetation cover) wave bands. Model out-put resolution is typically limited by the resolution ofthe thermal sensor (;102–103 m), which is often lowerthan would be desirable for local analyses and effectiveland surface change detection (;250 m; Townshend andJustice 1988). On current satellites, vis–NIR sensors op-erate at 1–6 times higher resolution than thermal sensors(Table 1) and more closely approach this optimal res-olution. The DisTrad algorithm developed by Kustas etal. (2003) sharpens the resolution of thermal band datato that of vis/NIR bands, taking advantage of the func-tional relationship between surface temperature andvegetation indices (VIs), such as the Normalized Dif-ference Vegetation Index (NDVI; Rouse et al. 1973),observed within localized areas.

The DisTrad technique is based on fitting a second-order polynomial between radiometric temperature(TRAD the dependent variable) and the VI, aggregated tothe coarser thermal resolution through linear averaging(VIlow, the independent variable):

2T (VI ) 5 a 1 bVI 1 cVI .RAD low low low (3)

Here, the ‘‘hat’’ symbol indicates a temperature valuepredicted by the resulting regression equation. Coarse-scale pixels used in this regression are selected from the

scene such that each shows small subpixel heterogeneity(in terms of its coefficient of variation in VI), yet col-lectively represent a wide dynamic range in VI (seeKustas et al. 2003). Pixels with standing water are notconsidered.

The least squares relationship developed at the coarsescale [Eq. (3)] is then applied to the full-resolution VIdata to predict temperature at that finer scale:

ˆ ˆ ˆT 5 T (VI ) 2 DT ,RAD,high RAD high RAD,low (4)

where

ˆ ˆDT 5 T 2 T (VI )RAD,low RAD,low RAD low (5)

is a residual correction ensuring the average temperaturein each coarse-scale pixel area matches the unsharpenedtemperature, TRAD,low. The restoration of these residuals,DTRAD,low, to the sharpened image is critical because itpreserves spatial patterns in surface temperature that aredue to moisture (rather than cover) variability and there-fore are not accounted for by Eq. (3). The result is thathigh-spatial-frequency structure is created within eachcoarse-scale pixel such that it is consistent with overalltemperature versus VI behavior exhibited within thescene, yet with the constraint that reaggregation willrecover the original low-resolution thermal image.

3. Data

In this study, the DisALEXI and DisTrad algorithmswere applied to Landsat imagery collected over the stateof Oklahoma, which hosts an extensive network ofweather and flux-monitoring stations known as theOklahoma Mesonet (Brock et al. 1995; Shafer et al.2000).

The mesonet became operational in 1994 and consistsof 115 automated stations that measure a set of coreparameters including air temperature and relative hu-midity at 1.5 m, wind speed and direction at 10 m,atmospheric pressure, incoming solar radiation, rainfall,and bare and vegetated soil temperatures. In 1999, aspart of the Oklahoma Atmosphere Surface-layer Instru-mentation System (OASIS) Project (Brotzge et al.1999), 90 mesonet stations were augmented with in-struments for measuring surface heat fluxes using a gra-dient-profile technique (Paulson 1970). For comparison,eddy covariance instrumentation was additionally in-stalled at 10 of these OASIS sites, designated as ‘‘su-persites.’’ The analyses here focus on seven supersitesthat were located within Landsat 7 scenes acquired on4 exceptionally clear days during 2000–01 (Fig. 1; Table3).

a. Model input

1) METEOROLOGICAL INPUTS

Ancillary surface and atmospheric data required bythe modeling system include an estimate of the wind


TABLE 2. Sources of input data used in this study for the ALEXI and DisALEXI models.

Input data Purpose Source (ALEXI) Source (DisALEXI)

Thermal IR Surface temperature GOES (5 km) Landsat (30 m)*Vegetation-cover fraction Temperature partitioning AVHRR NDVI (1 km) Landsat NDVI (30 m)Land-cover type (with cover

fraction)Surface roughness, displace-

ment height, and radiomet-ric properties

AVHRR (1 km) Perennial ground cover as-sumed

Downwelling shortwave andlongwave radiation

Net radiation GOES (20 km) GOES (20 km)

Wind speed Transport resistances ASOS/AWOS** analysis ASOS/AWOS** analysisABL temperature and mixing ra-

tio profilesABL submodel (ALEXI), at-

mospheric correctionsRadiosonde network Radiosonde network

* Sharpened from native resolution of 60 m using DisTrad.** Data obtained from the national Automated Surface Observing System (ASOS)/Automated Weather Observing System (AWOS)

network.

TABLE 3. OASIS supersites used in the ALEXI/DisALEXI analyses.

Date observed Site Station ID Lat (8N) Lon (8W) Primary land use

29 May 2000

12 Aug 2000

10 Jun 2001

12 Jul 2001

AlvaBessieIdabelStiglerMarenaNormanGrandfieldNorman

ALV2BESSIDABSTIGMARENORMGRA2NORM

36.7135.4033.8335.2736.0635.0634.2435.06

98.7199.0694.8895.1897.2197.4898.7497.48

Agriculture/pasturePasturePasturePasturePastureScrubAirport/scrubScrub

speed field at 50 m AGL and an early-morning atmo-spheric profile of temperature and mixing ratio (to 5–8-km altitude) at each 5-km grid cell in the ALEXImodeling domain (see Table 2). These input fields arecurrently created with the analysis component of a me-soscale model (in initialization mode) using standardobservations from the synoptic weather and radiosondenetworks. The temperature profile is used in the ABLsubmodel component of ALEXI; temperature and mix-ing ratio profiles are used to atmospherically correct theremotely sensed TIR imagery. Note that model sensi-tivity to errors in these meteorological inputs is rela-tively small (Anderson et al. 1997); therefore, local mea-surements are not required.

2) COARSE-RESOLUTION REMOTE SENSING DATA

The coarse-resolution surface temperature observa-tions used in ALEXI were obtained with the GOES-8imager instrument within the 10.2–11.2-mm (band 4)window. Thermal imager data are available every 15min at an average nadir-viewing angle of approximately408 and a nominal spatial resolution of 5 km at thelocation of the Oklahoma study area. Atmospheric cor-rections were performed using the methods describedby French et al. (2003), and a vegetation-cover-depen-dent correction for surface emissivity was applied as inMecikalski et al. (1999).

Downwelling solar and longwave radiation were es-timated at each pixel in the ALEXI grid from hourly

GOES data at 20-km resolution (Diak et al. 1996, 2000).Using a biweekly composited NDVI product (Eiden-shink 1992) generated with the Advanced Very HighResolution Radiometer (AVHRR), estimates of vege-tation-cover amount were derived with the expressionsuggested by Choudhury et al. (1994). These cover es-timates were used in conjunction with a land surfaceclassification created for the conterminous United States(USGS 1995) to assign relevant surface parameters suchas surface roughness, displacement height, and radio-metric properties (see Mecikalski et al. 1999 for moredetails).

3) HIGH-SPATIAL-RESOLUTION REMOTE SENSING

DATA

High-resolution surface temperature and vegetation-cover maps were derived from Landsat7 ETM1 ther-mal and vis/NIR imagery. From Landsat archives forthe years 2000–01, four predominantly clear sceneswere selected, each containing two OASIS supersites(Table 3). A 5 km 3 5 km area coincident with theALEXI grid cell containing the supersites was thencarved from these Landsat images, resulting in eightsubscenes composed of 167 3 167 pixels of dimension30 m.

The TIR data were extracted from ETM1 band 6imagery, which are acquired at 60-m resolution. Eachthermal subscene was atmospherically corrected withthe method of French et al. (2003) using temperature


and mixing ratio profiles from the corresponding ALEXI5-km grid cell, and cover-dependent emissivity correc-tions were applied. Solar and longwave radiation esti-mates were also extracted from the ALEXI input fieldsand were held constant across each subscene.

NDVI was computed using 30-m resolution imageryfrom ETM1 bands 3 (visible) and 4 (NIR) and wasused to map vegetation-cover fraction following Choud-hury et al. (1994). To assign surface roughness char-acteristics, a land classification of ‘‘perennial groundcover’’ (class 7 in Mecikalski et al. 1999) was useduniformly across each subscene. This class was chosento best represent the grassy land cover typical at OASIStower locations, but will cause surface roughness to beunderestimated in forested areas, resulting in locally de-pressed sensible heating rates. In future studies, clas-sification maps should ideally be generated from high-resolution multispectral data.

b. Validation data

Eddy covariance (EC) flux measurements were col-lected at the seven OASIS supersites listed in Table 3.The EC instrumentation at these sites is described indetail by Brotzge et al. (1999), Brotzge (2000), andBasara and Crawford (2002). Sensible and latent heatfluxes are estimated with eddy covariance analysis andstandard corrections using data from a 3D sonic ane-mometer and Krypton hygrometer mounted on an in-strument tower at 4.5 m AGL. Additionally, net radi-ation is monitored with a four-component net radiometermounted at 2 m AGL, while the soil heat flux is com-puted with the combination approach of Tanner (1960)using measurements from heat flux plates installed at5-cm depth near the base of the tower, with heat storagein the surface soil layer estimated with two platinumresistance detectors at 0- and 5-cm depths. At each site,the instrument set is enclosed within a fenced area todeter vandalism and damage by grazing animals.

Errors in flux comparisons due to the difference be-tween modeled and measured flux reference heights (50and 4.5 m AGL, respectively) are expected to be smallunder the conditions studied here. Under clear noontimeskies, and with sensible heating rates greater than 100W m22, the boundary layer depths are likely on theorder of 1 km. Studies conducted in similar conditionsand landscapes in Oklahoma concerning the use of 2-m-AGL versus mixed-layer meteorological measure-ments in the TSM indicate that the resulting differencesin flux predictions were small (Kustas et al. 1999). Fur-thermore, MacPherson et al. (1999) found that aircraftflux observations at 25–35 m during SGP97 agreed wellwith 2-m flux-tower data when segmented by land use.

As is often the case with eddy covariance datasets(Kizer and Elliott 1991; Fritschen et al. 1992; Stannardet al. 1994; Lloyd et al. 1997; Twine et al. 2000; Wilsonet al. 2002; Brotzge and Crawford 2003), the sum ofthe turbulent sensible and latent heat fluxes, H and LE,

measured at the supersites was consistently less than theavailable energy, RN 2 G. Closure in the surface energybalance represented in these measurements ranged from95% to ,40%, as quantified by the ratio of (H 1 LE)/(RN 2 G). To maintain comparability with the models(which enforce closure), the flux measurements werecorrected by modifying H and LE such that, over houraveraging intervals, they summed to the available en-ergy (RN 2 G) yet retained the observed Bowen ratio(Twine et al. 2000). Fluxes from the BESS supersiteconsistently showed closure of less than 40%; therefore,measurements of LE and H from this site are not usedin any quantitative flux comparisons.

Local meteorological datasets were also assembledfor each supersite, although these data are used for in-terpretation purposes only. Time traces of precipitationand saturation deficit measured at the supersites areshown in Fig. 3 for a period of 2 weeks prior to eachstudy day. Observations of wind speed and direction areused in the footprint analyses presented below.

4. Methods

a. Flux disaggregation

For each day of Landsat imagery, the ALEXI modelwas executed at 5-km spatial resolution over the entireOklahoma modeling domain using a GOES-derived sur-face temperature rise signal measured from an hour pastlocal sunrise to the time of the Landsat overpass (tLS),typically 4 h later. Output included modeled fields ofthe four major flux components and air temperature at50 m AGL, each field coincident in time with the Land-sat imagery. Modeled air temperatures were extractedfrom each ALEXI grid cell containing a target OASISsupersite and used as an upper boundary for disaggre-gating the 5-km fluxes associated with that cell (Fig.2).

The atmospherically corrected Landsat TIR data usedin the disaggregation were normalized with respect tothe GOES temperature data to preserve the surface-to-air temperature gradient predicted by the ALEXI model.To minimize effects of errors in registration between thecoarse-scale and finescale imagery, this normalizationwas determined over a 35 km 3 35 km area (7 3 7ALEXI cells) surrounding each mesonet station, yield-ing an average correction of 10.98C. The bias-corrected60-m thermal data were then interpolated to a 30-m gridcoincident with the Landsat vegetation-cover data. Forcomparison, three different interpolation schemes wereemployed: a bilinear interpolation and two variations ofthe thermal-sharpening technique, DisTrad, as describedbelow.

DisALEXI was applied to the resulting 30-m tem-perature and vegetation-cover fields for each Landsatsubscene. As with ALEXI, the DisALEXI output fluxfields represent a snapshot of instantaneous surface con-ditions at the time tLS. Jackson et al. (1983) suggested


FIG. 3. Daily precipitation (bars) and hourly atmospheric saturation deficit (lines) measured atthe seven OASIS flux stations supplying validation data for the current study, shown over a 2-week window prior to each analysis date.

a procedure for extrapolating daily integrated surfacefluxes from single-time estimates by assuming that thesurface evaporative fraction (EF; the ratio of latent heatto the available energy) remains approximately constantthroughout the day. Daily ET can then be related directlyto the diurnal available energy curve, which can be pre-dicted using only remote sensing and standard weatherdata. This is often a good approximation but can neglecteffects due to afternoon changes in atmospheric con-ditions and/or surface moisture/temperature stress (Sug-ita and Brutsaert 1991; Brutsaert and Sugita 1992; Hallet al. 1992; Kustas et al. 1993).

For comparison with mesonet EC measurements, 30-m disaggregated fields of net radiation, and sensible,latent, and soil heat flux were reaggregated using theone-dimensional analytical footprint model of Schueppet al. (1990, 1992), applied as a weighted average alonga transect (;30 m wide) at the angle of the mean localwind direction (averaged over the interval tLS 6 0.5 h).This weighting function uses estimates of local surface

roughness and friction velocity, which were computedas weighted averages from model output along the sametransect. While this is a relatively simplistic represen-tation of complex, time-dependent transport phenome-na, it does capture some signature of the azimuthal var-iability in surface conditions.

b. Thermal sharpening

The two-source model embedded in ALEXI andDisALEXI has difficulty converging given incongruoussurface temperature and vegetation-cover data. Simpletechniques for interpolating thermal data to the higherresolution of the cover information can result in incom-patible inputs in localized areas, especially along narrowdiscontinuities such as roads and riverbeds. This leadsto fringes of nonconvergent grid cells; for example,where the high radiometric temperature of an unresolvedblack asphalt road spills over into the full green coverin the roadside ditch. In such circumstances, model it-


eration using Eq. (1) can produce diverging estimatesof soil temperature and sensible heating rates. TheDisTrad thermal sharpening algorithm has the potentialof improving the performance of the disaggregation pro-cess by interpolating structure into the thermal mapsthat is functionally consistent with the high-resolutioncover information. The alternative is settling for thelower native resolution of the thermal sensor.

The sharpening tests presented here used NDVI asthe vegetation index forming the independent variablein Eq. (3). Two methods of treating the residual termDTRAD,low in Eq. (4) were examined. Again, these resid-uals are critical to maintaining fidelity with the observedthermal data at their coarse native scale. One methodis to restore the residuals in an unadulterated form, asimplemented by Kustas et al. (2003); this will be re-ferred to as the ‘‘standard sharpening method’’ (SS).This assures perfect correspondence with the low-res-olution image but can give a boxy structure to the re-sultant sharpened image. The restored 60-m residualsoften mask high-resolution structure reconstructedthrough the sharpening process.

A close examination showed that this boxiness couldbe removed, with little damage to coarse-scale fidelity,if the residuals are smoothed before being added backinto the sharpened image. An optimal balance betweenfidelity and visual information content was achievedwhen the residual map was convolved with a Gaussiankernel with full width half maximum equal to the rawthermal resolution (60 m in this case). This will bereferred to as the ‘‘convolved sharpening method’’ (CS).

Figure 4 demonstrates the utility of these thermal-sharpening techniques as applied to the Landsat sub-scene containing the IDAB mesonet site. Here, the raw60-m Landsat temperature data have been interpolatedto a 30-m grid using the standard and convolved-resid-ual sharpening algorithms. For comparison, a simplebilinear interpolation (BL) from 60- to 30-m resolutionis also shown. The convolved sharpening method is ableto pull out significant detail in comparison with the othermethods, aiding visual interpretation of features in theimage. On the 60-m scale, however, similarity is main-tained between the four thermal images.

5. Results and discussion

Disaggregated fields of latent heating surrounding theOASIS supersites are shown in Fig. 5. These maps dem-onstrate the influence of various thermal interpolationtechniques (bilinear interpolation and standard and con-volved sharpening) on model flux distributions. In thefollowing subsections, the accuracy and utility of thedisaggregation process and associated thermal-sharp-ening techniques are evaluated quantitatively in com-parison with ground-based tower-flux data and quali-tatively in terms of enhanced information content thatemerges at high resolution where flux patterns can beidentified with recognizable surface phenomena.

a. Model–tower-flux comparisons

Figure 6a compares instantaneous fluxes predicted bythe ALEXI model at the 5-km scale with measurementsmade at the mesonet stations listed in Table 3. In com-parison, Fig. 6b shows the improvement in agreementobtained when the 5-km fluxes are disaggregated downto the 30-m scale and integrated over the tower-fluxfootprint. Statistical measures of model performance inpredicting the observed flux components RN, H, LE,and G are listed in Table 4, including the coefficient ofefficiency (E) proposed by Nash and Sutcliffe (1970)as a performance metric preferable to the coefficient ofdetermination (R2), which can indicate perfect agree-ment even in the presence of systematic model biases.The coefficient of efficiency is particularly useful incomparing the accuracy of predictions from multiplemodels with respect to a given set of observed data.With this index, the relative utility of various thermal-sharpening algorithms can be assessed in terms ofDisALEXI flux prediction accuracy.

The benefit of using a disaggregation technique tovalidate kilometer-scale flux predictions made over het-erogeneous landscapes is readily apparent in Fig. 6 andTable 4. While the overall bias in the ALEXI 5-kmpredictions is small in comparison with ground-basedobservations, the scatter is relatively large (18%) andwill be strongly dependent on the exact location of theobserving station within each ALEXI grid cell (see Fig.5). The standard deviation in disaggregated latent heat-ing estimates over a given 5 km 3 5 km scene rangesfrom 40 to 80 W m22; comparable to the root-mean-square difference (rmsd) of 90 W m22 between ALEXIpredictions and measured ET fluxes in Fig. 6a. To di-rectly validate ALEXI predictions with tower data, mul-tiple stations would need to be deployed within eachcell, carefully positioned to capture a representativerange in surface conditions. Disaggregation reduces therelative model error by a factor of 2 (to 9%) and yieldsa higher coefficient of efficiency in comparison withtower-based flux measurements.

The statistics in Table 4 indicate, however, that thespecific technique used to interpolate the thermal Land-sat data from 60 to 30 m did not significantly affectmodel accuracy. A simple bilinear interpolation per-formed as well as the standard and convolved thermal-sharpening algorithms over this range in resolution. Thisis consistent with the findings of Kustas et al. (2003),who note that flux predictions from DisALEXI weregreatly improved using data sharpened to the 200-mscale, but little was gained by further sharpening to the30-m scale. In that experiment, the dominant scale oftemperature variations was defined by the typical sizeof the fields within the study scene (;200 m); thus,sharpening to finer scales did not reap significant benefit.Similar scales of heterogeneity are evident in the scenesstudied here.

The supersite showing the greatest discrepancy with


FIG. 4. Demonstration of the DisTrad thermal-sharpening algorithm, as applied to a 5 km 3 5 km scenefrom 8 Aug 2000 containing the IDAB mesonet station. (top) The 30-m NDVI data used in the sharpeningprocess, along with the 60-m residuals to the sharpening polynomial. Raw thermal data at 60-m resolutionwere resampled to 30-m resolution using (middle) bilinear interpolation and (bottom) the standard and con-volved DisTrad sharpening techniques. The white rectangle highlights residual features discussed in the text.


FIG. 5a. Disaggregated latent heat fields at 30-m resolution, demonstrating the impact of different inter-polation procedures applied to thermal input data: bilinear interpolation, standard sharpening, and convolvedsharpening. Scenes contain (a) the ALV2 and BESS sites on 29 May 2000. (b) the IDAB and STIG sites on8 Aug 2000, (c) the MARE and NORM sites on 10 Jun 2001, and (d) the GRA2 and NORM sites on 12Jul 2001. OASIS site locations are indicated with a white circle. Nonconvergent pixels appear as isolated orblocks of black cells.


FIG. 5b.


FIG. 5c.


FIG. 5d.


FIG. 6. Comparison between OASIS flux-tower observations with (a) modeled flux predictionsfrom ALEXI at 5-km resolution and (b) disaggregated fluxes from DisALEXI at 30-m resolution,reaggregated over the tower-flux footprint. The energy budget closure has been enforced amongthe observed fluxes over hourly intervals using in situ measurements of the soil heat flux. Graysquares indicate fluxes at site ALV2.

TABLE 4. Quantitative measures of model performance* in estimating instantaneous and daytime-integrated surface fluxes of RN, H, LE,and G considered collectively.

Model run N O MBE Rmsd a b R2 E % Error

Instantaneous W m22 W m22 W m22 W m22

ALEXIDisALEXI—BLDisALEXI—SSDisALEXI—CS

30303030

294.3294.3294.3294.3

8.25.86.26.6

63.534.534.234.6

29.521.721.621.4

0.930.950.950.95

0.900.970.970.97

0.900.970.970.97

17.59.09.09.2

Daytime MJ m22 MJ m22 MJ m22 MJ m22

DisALEXI—CS 28 8.99 0.13 1.42 1.04 0.90 0.96 0.95 12.2

* Here N is the number of observations, is the mean observed flux, rmsd is the root-mean-square difference between the modeled (P)Oand observed (O) quantities, MBE is the mean bias error ( 2 ), a and b are the intercept and slope of the linear regression of P on O,¯ ¯P OR2 is the coefficient of determination, E is the coefficient of effciency, and the percent error is defined as the mean absolute differencebetween P and O divided by the mean observed flux. Thermal inputs to DisALEXI were interpolated from 60- to 30-m resolution usingBL, SS, and CS techniques.

model predictions is ALV2, where instantaneous latentheating is underestimated by 20% (;90 W m22). Thissite lies in pasture due north of a fenced field with lowvegetation cover on the day in question (see Fig. 5a).High surface temperatures suggest the soil surface inthis field had dried significantly since the heavy rainfall3 days prior (Fig. 3). At the time of observation, windsat the site were from the south-southwest, causing theanalytical footprint transect to intersect this field andlowering the footprint-integrated model latent heating.Proximity to the sharp discontinuity between pastureand field makes this site particularly sensitive to errorsin image navigation and footprint analysis.

Consistent overestimation of G by both the ALEXIand DisALEXI models, evident in Fig. 6, may be duein part to subgrid-scale anomalies in vegetation covercoincident with the supersites. Most of the mesonet sitesconsidered in this study are located in pasture, screenedfrom animal intrusion by fencing. At some sites, the

disparity in green cover fraction inside and outside ofthe site fencing is quite significant. The soil heat fluxplates are buried within the protected area and thereforeare susceptible to this type of small-scale cover bias,while the eddy covariance sensors integrate fluxes em-anating from inside and outside the fencing. The up-welling radiation sensors on the tower-mounted net ra-diometer will similarly sample only a small area (;10m across) within the enclosure. In this study, however,the mean bias in modeled RN (8 W m22) is small com-pared to that for G (33 W m22). This is consistent withfindings by Brotzge and Crawford (2003), who observedlarge differences (factor of 2) between diurnal soil heatcurves measured at two flux stations spaced 100 m apartat the Foraker, Oklahoma, OASIS supersite (not studiedhere), but only slight differences in net radiation, sug-gesting that G may be more sensitive than RN to spatialvariations in vegetation cover, soil type, and moisturecontent. Because midday net radiation is heavily driven


FIG. 7. Comparison between OASIS flux-tower observations withdisaggregated fluxes from DisALEXI at 30-m resolution, reaggre-gated over the tower-flux footprint. The energy budget closure hasbeen enforced among the observed fluxes over hourly intervals usingthe modeled soil heat flux. Gray squares indicate fluxes at site ALV2.

TABLE 5. Quantitative measures of model performance* in estimating instantaneous and daytime-integrated surface fluxes of LE and Hconsidered collectively (RN and G are excluded from these comparisons). ‘‘Closure’’ indicates whether the measured fluxes were closedusing the observed G (GO) or the model predicted G (GP).

Model run Closure N O MBE Rmsd a b R2 E % Error

Instantaneous W m22 W m22 W m22 W m22

DisALEXI—BL

DisALEXI—SS

DisALEXI—CS

GO

GP

GO

GP

GO

GP

141414141414

245.5229.8245.5229.8245.5229.8

211.04.6

210.84.9

210.65.4

35.628.734.828.135.229.1

46.239.243.336.540.733.8

0.770.850.780.860.790.88

0.870.880.870.880.870.88

0.840.880.850.880.850.87

10.29.7

10.09.3

10.29.9

Daytime MJ m22 MJ m22 MJ m22 MJ m22

DisALEXI—CS GO

GP

1212

7.947.46

20.420.06

1.741.08

2.622.45

0.620.68

0.850.87

0.770.83

15.714.2

* N, , rmsd, MBE, a, b, R2, E, and % error are defined as in Table 4.O

by the shortwave component, particularly under clear-sky conditions, and because the albedo of the soil andvegetation is similar, this is not unreasonable.

Given the fact that the soil heat flux plates and eddycovariance flux systems are sampling different surfacefootprints, one might ask whether it is reasonable to usepoint-scale measurements of G to close the observedenergy budget (Lloyd et al. 1997; Brotzge et al. 1999;Baldocchi 2003). A remote sensing model may in factpredict an average soil heat flux that is more represen-tative of the source area contributing to the EC fluxesthan do one or two measurements made in the vicinityof the tower base. Figure 7 shows a comparison of pre-dicted and observed fluxes, where the observed energy

budget has been closed using the modeled G integratedover the EC flux footprint, instead of the observed G;statistical measures of model performance in predictingthe flux components H and LE (the components alteredin enforcing closure) are listed in Table 5. With thismodification the rmsd in predicted H and LE decreased,with the most significant improvement occurring at theALV2 site where the error in latent heating was reducedto 10% (35 W m22).

Figure 8a compares flux measurements integratedover the daytime hours (when solar radiation is nonzero)with time-integrated fluxes extrapolated from instanta-neous flux predictions assuming a constant daytimeevaporative fraction. A statistical description of thesecomparisons is provided in Table 4. In general, theagreement is good (12% relative error), but slightlypoorer than for the instantaneous flux determinations.This may be due in part to diurnal changes in the EF,but also to the fact that biases in eddy covariance mea-surements tend to amplify when integrated over dailyor annual time scales (Moncrieff et al. 1996). As withinstantaneous fluxes, agreement in daytime totals im-proves when the modeled soil heat flux is used to closethe observed energy budget (Fig. 8b; Table 5).

b. Spatial and temporal flux variations

While thermal sharpening did not significantly im-prove statistical comparisons between modeled andmeasured fluxes, in all cases it did improve model con-vergence frequency and visual information content inthe output flux fields (Fig. 5; nonconvergent pixels ap-pear as isolated or blocks of black cells). Improvementin model convergence is most evident in scenes withstrong, narrow discontinuities, such as along the riv-erbeds in the BESS and ALV2 scenes (Fig. 5a) and theurban street systems in the scenes containing the NORMsite (Figs. 5c,d). For the MARE site (Fig. 5c), usingbilinearly interpolated thermal inputs resulted in manynonconvergent pixels fringing the high ET river bot-toms, where abrupt changes in vegetation-cover fraction


FIG. 8. Comparison between OASIS flux-tower observations, integrated over the daytime hours,with daily integrated fluxes from DisALEXI at 30-m resolution, reaggregated over the tower-fluxfootprint. The energy budget closure has been enforced among the observed fluxes over hourlyintervals using (a) in situ measurements of the soil heat flux and (b) the modeled soil heat flux.Gray squares indicate fluxes at site ALV2.

FIG. 9. Disaggregated fields of latent heating at 30-m resolution over scenes containing the NORM OASIS site on10 Jun and 12 Jul 2001 (thermal inputs sharpened using the CS technique). An aerial photograph of the scene at 16-m resolution taken on 20 Feb 1995 is included for comparison.

are detected at 30-m resolution. Nonconvergence wasreduced in employing the standard thermal-sharpeningalgorithm and was almost eliminated with the convolvedsharpening technique.

The enhancement of visual information content at-tained through thermal sharpening is most pronouncedin the vicinity of the NORM supersite, which is locatedwithin the city limits of Norman, Oklahoma (Fig. 9). Inthe 2-week period prior to 10 June 2001, 89 mm ofcumulative precipitation was recorded at the NORMweather station, as compared to 6 mm during the 2weeks before 12 July. The impact of decreased ante-cedent rainfall on latent heating for 12 July is clear in

Fig. 9—only in riparian areas or grounds receiving reg-ular irrigation (golf course, lawns) are high ET ratessustained and in fact enhanced over 10 June rates be-cause of increased atmospheric demand (Fig. 3). Withthe CS thermal inputs, higher ET from individual fair-ways in the Westwood Park Golf Course (just south ofthe airport) can be detected; these fairways would bemore heavily irrigated than adjacent areas in the course.Low ET rates are predicted for both days in industri-alized areas around the Sooner Fashion Mall (locatedin the southwest corner, near the cloverleaf interchange)and eastward along the Main Street business corridor.

The residuals to the sharpening polynomial [DTRAD,low


in Eq. (5)] also reveal interesting information regardingspatial anomalies in the scene-average surface temper-ature–vegetation-cover relationship. Clouds and waterbodies will show up as anomalies (both negative—ob-served surface temperatures are colder than is typicalfor a given cover fraction). Variations in soil moistureand vegetative stress may also be revealed (positive re-siduals for soil moisture deficiency and stressed vege-tation; negative for wet soils). In the IDAB scene, forexample, a backward-L-shaped field with low vegeta-tion cover shows strong negative residuals compared toan adjacent field of similarly low cover (highlighted inFig. 4). These residuals appear indicative of high soilmoisture content, perhaps due to recent irrigation; notethat DisALEXI has predicted higher latent heating inthis field (Fig. 5b). The extent to which this particularfield stands out in the residual map suggests that thismay be a valuable tool in identifying soil moistureanomalies over large regions.

6. Conclusions

A procedure (DisALEXI) for disaggregating large-scale surface flux predictions from the ALEXI modeldown to much finer spatial scales has been applied tomultiscale and multiwavelength remote sensing datacollected with the GOES/AVHRR (5 km) and Landsat(30–60 m) satellites. In this modeling system, TIR andvis/NIR wave bands provide information on surfacetemperature and vegetation-cover status, while kilo-meter-scale imagery is used to provide atmosphericboundary conditions for meter-scale surface energy bal-ance evaluations. This nested-mapping technique is par-ticularly well suited for operational applications becauseit does not need local surface meteorological measure-ments, nor does it require that a wide range in surfaceconditions be present within the modeling scene, bothof which are common limitations among current ther-mal-mapping approaches.

Disaggregation provides a means for indirectly val-idating regional-scale flux predictions by breaking themdown to scales consistent with the surface source foot-print associated with flux-tower observations. When re-aggregated using an analytical footprint weighting func-tion, modeled fluxes from DisALEXI agreed with mea-surements made at seven stations in the OASIS fluxnetwork in Oklahoma to within 10% on an instantaneousbasis and 15% for daytime totals. In comparison, 5-kminstantaneous flux predictions from the regional-scaleALEXI model exhibited a 17% relative error with re-spect to ground observations, which was strongly de-pendent on the exact locations of the measurement siteswithin the model grid cells. Clearly, direct comparisonwith individual ground-based tower measurements is notan optimal means for assessing the accuracy of regional-scale flux models, especially over complex landscapes.

The utility of an algorithm (DisTrad) for sharpeningthermal satellite imagery to the higher resolutions typ-

ically associated with vis–NIR bands was also evaluatedin the context of disaggregation results. This techniquehas potential value in that high-resolution thermal data(,1 km) are not currently available with daily coverageand will become even less accessible as the thermal bandis being eliminated from future Landsat platforms. Un-like DisALEXI, DisTrad does require a certain amountof variability be present within the scene to develop adecent regression between temperature and vegetationindex. Sharpening thermal data inputs to DisALEXIfrom 60 to 30 m using DisTrad did not significantlyimprove agreement between modeled and measuredfluxes in comparison to inputs processed with a simplebilinear interpolation. This is not surprising, as the typ-ical scale of surface heterogeneity in the vicinity of theflux stations was on the field scale, ;200–500 m. Great-er benefit in terms of observational agreement is to beexpected in applying DisTrad to thermal data fromMODIS, where the potential for resolution enhancementis 1 km (TIR) to 250 m (vis/NIR); testing of this hy-pothesis is underway. Nevertheless, the DisTrad sharp-ening of Landsat thermal data in the current study didresult in superior model output in terms of visual in-formation content and model convergence rate, partic-ularly when the residuals to the sharpening functionwere nominally smoothed through convolution.

There is significant motivation for developing robustand operational nested flux-mapping capabilities. Fluxmaps at 101–102-m pixel resolution will be useful inprecision agriculture, drought and yield forecasting,groundwater modeling, large-eddy simulation, and indetecting functional changes in natural and managedecosystems in response to climatic and anthropogenicstressors. The scalability in assessment, from local tolandscape to regional scales, that nested modeling pro-vides is beneficial in linking cause with effect, whichmay be occurring on very different spatial scales. Fur-thermore, methods for upscaling local ground-based ob-servations into a regional context will be critical to thesuccess of ongoing continental- and global-scale fluxmeasurement programs.

Acknowledgments. Funding for this research was pro-vided primarily by NASA Grant NAG13-99008 and inpart by USDA Cooperative Agreement 58-1265-1-043.Suggestions and comments made by three anonymousreviewers greatly improved the clarity and overall pre-sentation of the paper.

REFERENCES

Albertson, J. D., W. P. Kustas, and T. M. Scanlon, 2001: Large-eddysimulation over heterogeneous terrain with remotely sensed landsurface conditions. Water Resour. Res., 37, 1939–1953.

Anderson, M. C., J. M. Norman, G. R. Diak, W. P. Kustas, and J. R.Mecikalski, 1997: A two-source time-integrated model for es-timating surface fluxes using thermal infrared remote sensing.Remote Sens. Environ., 60, 195–216.

——, W. P. Kustas, and J. M. Norman, 2003: Upscaling and down-


scaling—A regional view of the soil–plant–atmosphere contin-uum. Agron. J., 95, 1408–1423.

Baldocchi, D. D., 2003: Assessing the eddy covariance technique forevaluating carbon dioxide exchange rates of ecosystems: Past,present and future. Global Change Biol., 9, 479–492.

——, and Coauthors, 2001: FLUXNET: A new tool to study thetemporal and spatial variability of ecosystem-scale carbon di-oxide, water vapor, and energy flux densities. Bull. Amer. Me-teor. Soc., 82, 2415–2434.

Basara, J. B., and K. C. Crawford, 2002: Linear relationships betweenroot-zone soil moisture and atmospheric processes in the plan-etary boundary layer. J. Geophys. Res., 107, 4274, doi:10.1029/2001JD000633.

Bastiaansen, W. G. M., M. Menenti, R. A. Feddes, and A. A. M.Holtslag, 1998a: A remote sensing surface energy balance al-gorithm for land (SEBAL); 1. Formulation. J. Hydrol., 212–213,198–212.

——, H. Pelgrum, J. Wang, Y. Ma, J. F. Moreno, G. J. Roerink, andT. van der Wal, 1998b: A remote sensing surface energy balancealgorithm for land (SEBAL); 2. Validation. J. Hydrol., 212–213,213–229.

Brock, F. V., K. C. Crawford, R. L. Elliott, G. W. Cuperus, S. J.Stadler, H. L. Johnson, and M. D. Eilts, 1995: The OklahomaMesonet: A technical overview. J. Atmos. Oceanic Technol., 12,5–19.

Brotzge, J. A., 2000: Closure of the surface energy budget. Ph.D.dissertation, University of Oklahoma, 208 pp.

——, and K. C. Crawford, 2003: Examination of the surface energybudget: A comparison of eddy correlation and Bowen ratio mea-surement systems. J. Hydrometeor., 4, 160–178.

——, S. Richardson, K. Crawford, T. Horst, F. Brock, K. Humes, Z.Sorbjan, and R. Elliott, 1999: The Oklahoma Atmospheric Sur-face-layer Instrumentation System (OASIS) Project. Preprints,13th Symp. on Boundary Layers and Turbulence, Dallas, TX,Amer. Meteor. Soc., 612–615.

Brutsaert, W., and M. Sugita, 1992: Application of self-preservationin the diurnal evolution of the surface energy budget to determinedaily evaporation. J. Geophys. Res., 97, 18 377–18 382.

Chen, F., D. N. Yates, H. Nagai, M. A. LeMone, K. Ikeda, and R.L. Grossman, 2003: Land surface heterogeneity in the Coop-erative Atmosphere Surface Exchange Study (CASES-97). PartI: Comparing modeled surface flux maps with surface-flux towerand aircraft measurements. J. Hydrometeor., 4, 196–218.

Choudhury, B. J., N. U. Ahmed, S. B. Idso, R. J. Reginato, and C.S. T. Daughtry, 1994: Relations between evaporation coefficientsand vegetation indices studied by model simulations. RemoteSens. Environ., 50, 1–17.

Desjardins, R. L., and Coauthors, 1997: Scaling up flux measurementsfor the boreal forest using aircraft-tower combinations. J. Geo-phys. Res., 102 (D24), 29 125–29 133.

Diak, G. R., W. L. Bland, and J. R. Mecikalski, 1996: A note on firstestimates of surface insolation from GOES-8 visible satellitedata. Agric. For. Meteor., 82, 219–226.

——, ——, ——, and M. C. Anderson, 2000: Satellite-based esti-mates of longwave radiation for agricultural applications. Agric.For. Meteor., 103, 349–355.

Doran, J. C., J. M. Hubbe, J. C. Liljigren, W. J. Shaw, G. J. Collatz,D. R. Cook, and R. L. Hart, 1998: A technique for determiningthe spatial and temporal distributions of surface fluxes of heatand moisture over the Southern Great Plains Cloud and RadiationTestbed. J. Geophys. Res., 103, 6109–6121.

Eidenshink, J. C., 1992: The 1990 conterminous U.S. AVHRR dataset. Photogramm. Eng. Remote Sens., 58, 809–813.

Fan, S., M. Gloor, J. Mahlman, S. Pacala, J. Sarmiento, T. Takahashi,and P. Tans, 1998: A large terrestrial carbon sink in North Amer-ica implied by atmospheric and oceanic carbon dioxide data andmodels. Science, 282, 442–446.

Finn, D., B. Lamb, M. Y. Leclerc, and T. W. Horst, 1996: Experimentalevaluation of analytical and Lagrangian surface-layer flux foot-print models. Bound.-Layer Meteor., 80, 282–308.

French, A. N., J. M. Norman, and M. C. Anderson, 2003: Simplifiedcorrection of GOES thermal infrared observations. Remote Sens.Environ., 87, 326–333.

Fritschen, L. J., P. Qian, E. T. Kanemasu, D. Nie, E. A. Smith, J. B.Stewart, S. B. Verma, and M. L. Wesely, 1992: Comparisons ofsurface flux measurement systems used in FIFE 1989. J. Geo-phys. Res., 97, 18 697–18 713.

Gao, W., R. L. Coulter, B. M. Lesht, J. Qiu, and M. L. Wesely, 1998:Estimating clear-sky regional surface fluxes in the southern GreatPlains atmospheric radiation measurement site with ground mea-surements and satellite observations. J. Appl. Meteor., 37, 5–22.

Gardner, B. R., D. C. Nielsen, and C. C. Shock, 1992: Infrared ther-mometry and the crop water stress index: II Sampling proceduresand interpretation. J. Prod. Agric., 5, 466–475.

Gillies, R. R., and T. N. Carlson, 1995: Thermal remote-sensing ofsurface soil-water content with partial vegetation cover for in-corporation into climate models. J. Appl. Meteor., 34, 745–756.

Hall, F. G., K. F. Huemmrich, S. J. Goetz, P. J. Sellers, and J. E.Nickerson, 1992: Satellite remote sensing of surface energy bal-ance: Success, failures and unresolved issues in FIFE. J. Geo-phys. Res., 97, 19 061–19 089.

Horst, T. W., and J. C. Weil, 1992: Footprint estimation for scalarflux measurements in the atmospheric surface layer. Bound.-Layer Meteor., 59, 279–296.

Jackson, R. D., J. L. Hatfield, R. J. Reginato, S. B. Idso, and P. J. J.Pinter, 1983: Estimation of daily evapotranspiration from onetime-of-day measurements. Agric. Water Manage., 7, 351–362.

Jiang, L., and S. Islam, 2001: Estimation of surface evaporation mapover Southern Great Plains using remote sensing data. WaterResour. Res., 37, 329–340.

Kizer, M. A., and R. L. Elliott, 1991: Eddy correlation systems formeasuring evapotranspiration. J. Atmos. Sci., 47, 1721–1741.

Kubota, A., and M. Sugita, 1994: Radiometrically determined skintemperature and scalar roughness to estimate surface heat flux,I. Parameterization of radiometric scalar roughness. Bound.-Lay-er Meteor., 69, 394–416.

Kustas, W. P., and J. M. Norman, 1997: A two-source approach forestimating turbulent fluxes using multiple angle thermal infraredobservations. Water Resour. Res., 33, 1495–1508.

——, and ——, 1999a: Evaluation of soil and vegetation heat fluxpredictions using a simple two-source model with radiometrictemperatures for partial canopy cover. Agric. For. Meteor., 94,13–25.

——, and ——, 1999b: Reply to comments about the basic equationsof dual-source vegetation–atmosphere transfer models. Agric.For. Meteor., 94, 275–278.

——, and ——, 2000: A two-source energy balance approach usingdirectional radiometric temperature observations for sparse can-opy covered surfaces. Agron. J., 92, 847–854.

——, and J. D. Albertson, 2003: Effects of surface temperature con-trast on land–atmosphere exchange: A case study from Monsoon90. Water Resour. Res., 39, 1159, doi:10.1029/2001WR001226.

——, B. J. Choudhury, M. S. Moran, R. D. Reginato, R. D. Jackson,L. W. Gay, and H. L. Weaver, 1989: Determination of sensibleheat flux over sparse canopy using thermal infrared data. Agric.For. Meteor., 44, 197–216.

——, T. J. Schmugge, K. S. Humes, T. J. Jackson, R. Parry, M. A.Weltz, and M. S. Moran, 1993: Relationships between evapo-rative fraction and remotely sensed vegetation index and micro-wave brightness temperature for semiarid rangelands. J. Appl.Meteor., 32, 1781–1790.

——, J. H. Prueger, K. S. Humes, and P. J. Starks, 1999: Estimationof surface heat fluxes at field scale using surface layer versusmixed layer atmospheric variables with radiometric temperatureobservations. J. Appl. Meteor., 38, 224–238.

——, G. R. Diak, and J. M. Norman, 2001a: Time difference methodsfor monitoring regional scale heat fluxes with remote sensing.Land Surface Hydrology, Meteorology, and Climate: Observa-tions and Modeling, V. Lakshmi, J. Albertson, and J. Schaake,


Eds., Water Science and Application Series, Vol. 3, Amer. Geo-phys. Union, 15–29.

——, T. J. Jackson, A. N. French, and J. I. MacPherson, 2001b:Verification of patch- and regional-scale energy balance esti-mates derived from microwave and optical remote sensing duringSGP97. J. Hydrometeor., 2, 254–273.

——, J. M. Norman, M. C. Anderson, and A. N. French, 2003: Es-timating subpixel surface temperatures and energy fluxes fromthe vegetation index–radiometric temperature relationship. Re-mote Sens. Environ., 85, 429–440.

Leclerc, M. Y., and G. W. Thurtell, 1990: Footprint prediction ofscalar fluxes using a Markovian analysis. Bound.-Layer Meteor.,52, 247–258.

Lloyd, C. R., and Coauthors, 1997: A comparison of surface fluxesat the HAPEX-Sahel fallow bush sites. J. Hydrol., 188–189,400–425.

MacPherson, J. I., R. Dobosy, S. B. Verma, W. P. Kustas, J. H. Prueger,and A. Williams, 1999: Intercomparisons between flux aircraftand towers in SGP97. Preprints, 14th Conf. on Hydrology, Dal-las, TX, Amer. Meteor. Soc., 125–128.

Mahrt, L., D. Vickers, and J. Sun, 2001: Spatial variations of surfacemoisture flux from aircraft data. Adv. Water Resour., 24, 1133–1142.

Mason, P. J., 1988: The formation of areally-averaged roughnesslengths. Quart. J. Roy. Meteor. Soc., 114, 399–420.

McNaughton, K. G., and T. W. Spriggs, 1986: A mixed-layer modelfor regional evaporation. Bound.-Layer Meteor., 74, 262–288.

Mecikalski, J. M., G. R. Diak, M. C. Anderson, and J. M. Norman,1999: Estimating fluxes on continental scales using remotelysensed data in an atmosphere–land exchange model. J. Appl.Meteor., 38, 1352–1369.

Moncrieff, J. B., Y. Mahli, and R. Leuning, 1996: The propagationof errors in long term measurements of land atmosphere fluxesof carbon and water. Global Change Biol., 2, 231–240.

Moran, M. S., T. R. Clarke, Y. Inoue, and A. Vidal, 1994: Estimatingcrop water deficit using the relation between surface-air tem-perature and spectral vegetation index. Remote Sens. Environ.,49, 246–263.

——, A. F. Rahman, J. C. Washburne, D. C. Goodrich, M. A. Weltz,and W. P. Kustas, 1996: Combining the Penman–Monteith equa-tion with measurements of surface temperature and reflectanceto estimate evaporation rates from semiarid grassland. Agric.For. Meteor., 80, 87–109.

Nash, L. E., and J. V. Sutcliffe, 1970: River flow forecasting throughconceptual models—Part 1: A discussion of principles. J. Hy-drol., 10, 282–290.

Norman, J. M., M. Divakarla, and N. S. Goel, 1995a: Algorithms forextracting information from remote thermal-IR observations ofthe earth’s surface. Remote Sens. Environ., 51, 157–168.

——, W. P. Kustas, and K. S. Humes, 1995b: A two-source approachfor estimating soil and vegetation energy fluxes from observa-tions of directional radiometric surface temperature. Agric. For.Meteor., 77, 263–293.

——, ——, J. H. Prueger, and G. R. Diak, 2000: Surface flux esti-mation using radiometric temperature: A dual temperature dif-ference method to minimize measurement error. Water Resour.Res., 36, 2263–2274.

——, and Coauthors, 2003: Remote sensing of surface energy fluxesat 101-m pixel resolutions. Water Resour. Res., 39, 1221, doi:10.1029/2002WR001775.

Oechel, W. C., G. L. Vourlitis, S. Brooks, T. L. Crawford, and E.Dumas, 1998: Intercomparison among chamber, tower, and air-craft net CO2 and energy fluxes measured during the ArcticSystem Science Land–Atmosphere–Ice Interactions (ARCSS-LAII) flux study. J. Geophys. Res., 103 (D22), 28 993–29 003.

Paulson, C., 1970: The mathematical representation of wind speed

and temperature profiles in the unstable atmospheric surface lay-er. J. Appl. Meteor., 9, 857–861.

Priestley, C. H. B., and R. J. Taylor, 1972: On the assessment ofsurface heat flux and evaporation using large-scale parameters.Mon. Wea. Rev., 100, 81–92.

Rouse, J. W., R. H. Haas, J. A. Schell, and D. W. Deering, 1973:Monitoring vegetation systems in the Great Plains with ETRS.Proc. Third ETRS Symp., Washington, DC, NASA, NASASP353, 309–317.

Running, S. W., D. D. Baldocchi, D. P. Turner, S. T. Gower, P. S.Bakwin, and K. A. Hibbard, 1999: A global terrestrial monitoringnetwork integrating tower fluxes, flask sampling, ecosystemmodeling and EOS satellite data. Remote Sens. Environ., 70,108–127.

Schmid, H. P., 1994: Source areas for scalars and scalar fluxes.Bound.-Layer Meteor., 67, 293–318.

Schuepp, P. H., M. Y. Leclerc, J. I. MacPherson, and R. L. Desjardins,1990: Footprint prediction of scalar fluxes from analytical so-lutions of the diffusion equation. Bound.-Layer Meteor., 50,355–373.

——, J. I. MacPherson, and R. L. Desjardins, 1992: Adjustment offootprint correction for airborne flux mapping over the FIFE site.J. Geophys. Res., 97, 18 455–18 466.

Shafer, M. A., C. A. Fiebrich, D. S. Arndt, S. E. Fredrickson, and T.W. Hughs, 2000: Quality assurance procedures in the Oklahomamesonetwork. J. Atmos. Oceanic Technol., 17, 474–494.

Stannard, D. K., J. H. Blanford, W. P. Kustas, W. D. Nichols, S. A.Amer, T. J. Schmugge, and M. A. Weltz, 1994: Interpretation ofsurface flux measurements in heterogeneous terrain during theMonsoon ’90 experiment. Water Resour. Res., 30, 1227–1239.

Stewart, J. B., W. P. Kustas, K. S. Humes, W. D. Nichols, M. S.Moran, and A. A. R. DeBruin, 1994: Sensible heat flux–radio-metric surface temperature relationship for eight semiarid areas.J. Appl. Meteor., 33, 1110–1117.

Sugita, M., and W. Brutsaert, 1991: Daily evaporation over a regionfrom lower boundary layer profiles measured with radiosondes.Water Resour. Res., 27, 747–752.

Tanner, C. B., 1960: Energy balance approach to evapotranspirationfrom crops. Soil Sci. Soc. Amer. Proc., 24, 1–9.

Townshend, J. G. R., and C. O. Justice, 1988: Selecting the spatialresolution of satellite sensors required for global monitoring ofland transformations. Int. J. Remote Sens., 9, 187–236.

Twine, T. E., and Coauthors, 2000: Correcting eddy-covariance fluxunderestimates over a grassland. Agric. For. Meteor., 103, 279–300.

U.S. Geological Survey (USGS), 1995: Prototype 1990 ConterminousU.S. Land Cover Characteristics Data Set CD-ROM. U.S. Geo-logical Survey, Sioux Falls, SD.

Valentini, R., Ed., 2003: Fluxes of Carbon, Water and Energy ofEuropean Forests. Springer-Verlag, 260 pp.

Vining, R. C., and B. L. Blad, 1992: Estimation of sensible heat fluxfrom remotely sensed canopy temperatures. J. Geophys. Res.,97 (D17), 18 951–18 954.

Wieringa, J., 1986: Roughness-dependent geographical interpolationof the surface wind speed averages. Quart. J. Roy. Meteor. Soc.,112, 867–889.

Wilson, K., and Coauthors, 2002: Energy balance closure at FLUX-NET sites. Agric. For. Meteor., 113, 223–243.

Wofsy, S. C., and Coauthors, 2002: The North American CarbonProgram (NACP): A Report of the NACP Committee of the U.S.Carbon Cycle Science Steering Group. U.S. Global Change Re-search Program, Washington, DC, 56 pp.

Zhan, X., W. P. Kustas, and K. S. Humes, 1996: An intercomparisonstudy on models of sensible heat flux over partial canopy surfaceswith remotely sensed surface temperatures. Remote Sens. En-viron., 58, 242–256.

A Multiscale Remote Sensing Model for Disaggregating Regional Fluxes to Micrometeorological Scales

Documents