evaluation of remotely sensed evapotranspiration over the ceop

Journal of the Meteorological Society of Japan, Vol. 85A, pp. 439--459, 2007 439

Evaluation of Remotely Sensed Evapotranspiration Over the

CEOP EOP-1 Reference Sites

H. SU, E.F. WOOD

Department of Civil and Environmental Engineering, Princeton University, Princeton, NJ 08544, USA

M.F. MCCABE

Los Alamos National Laboratory, PO Box 1663, MS-D436, Los Alamos, NM 87545, USA

and

Z. SU

International Institute for Geo-Information Science and Earth Observation (ITC), Hengelosestraat 99,P.O. Box 6, 7500 AA Enschede, The Netherlands

(Manuscript received 18 February 2006, in final form 26 October 2006)

Abstract

In this paper, the Coordinated Enhanced Observing Period (CEOP) data during an Enhanced Observ-ing Period (EOP-1) is used to assess the Surface Energy Balance System (SEBS) model. The purpose ofthis study is to evaluate the adaptability of SEBS to different climatic zones and land cover classifica-tions at two different scales. The SEBS model was examined at the field (tower) scale based primarilyon in-situ observations from CEOP sites. To examine a broader scale application, remotely sensed landsurface temperature (LST) from the MODIS sensor and surface meteorology from the Global Land DataAssimilation System (GLDAS) were used for the required forcing datasets. Comparisons at tower scaleshow that the model predictions of the energy fluxes agree reasonably well with the observations. Theroot mean square error (RMSE) of the ET prediction based on MODIS Land Surface Temperature(LST) plus CEOP meteorological observations is about 61 W m�2 at a grassland site (Cabauw) and a nee-dle leaf forest site (BERMS). The RMSE of ET predication at a corn site (Bondville) is 96 W m�2 and thecorresponding percentage error is 28.9%. When GLDAS forcing was used instead of the CEOP tower ob-servations, the RMSEs of ET prediction at Cabauw, BERMS and Bondville are increased to 82, 84 and140 W m�2 respectively. The negative bias of surface downward radiative forcing from GLDAS contrib-uted much to the larger deviation of the ET prediction when compared to tower based values. The inno-vative aspects of our study in this paper are: a) No similar work on evaluating remote sensing based ETmodel under a diverse climate and land cover condition has been done before; b) ET modeling was as-sessed in different scales ranging from site scale to GLDAS grid cell; c) The framework of estimatingthe spatial distribution of ET combining satellite data and available ground meteorology is tested.

1. Introduction

Evapotranspiration (ET) is the combinationof water evaporated from the surface and tran-spired by plants and has been the focus of

Corresponding author: Hongbo Su, Department ofCivil and Environmental Engineering, PrincetonUniversity, Princeton, NJ 08544, USA.E-mail: [email protected]( 2007, Meteorological Society of Japan

intensive research for many decades (e.g., Pen-man 1948; Priestley and Taylor 1972; Monteith1981). ET is an important process which linksthe different components of the energy balanceand water cycle for the atmospheric and landsurface, and much research is focused on quan-tifying this process (Famiglietti and Wood 1991;Braun et al. 2001). Conventional measurementtechniques such as the Bowen ratio and eddycovariance approaches provide relatively accu-rate means to estimate components of energybalance, but are essentially point scale predic-tions. Such local scale measurements are in-adequate in studying the land surface energybalance at regional or global scales, due inpart to the spatial heterogeneity of the landsurface parameters and the dynamic nature ofthe evaporative process (Famiglietti and Wood1995; Hu and Islam 1998).

Remote sensing techniques can make it pos-sible to estimate the surface energy compo-nents at a regional to global scale (Bastiaans-sen et al. 1998; Kustas and Norman 1996; Suand Menenti 1999). There remain however, sig-nificant issues related to the appropriate vali-dation of regionally or globally distributedmodel predictions, particularly as the scalingbehavior of the hydrological fluxes require care-ful consideration (Wood 1995; McCabe et al.2005) and the choice of representative valida-tion sites is critical for thorough model evalua-tion (Adolphs 1999; Plana 2002).

The Surface Energy Balance System (SEBS)proposed by Su (2002) is a model which integra-tes satellite data and available surface meteo-rological information to estimate atmosphericturbulent fluxes and the evaporative fractionat the land surface. It has been evaluated andapplied from the tower to regional scale (Su2002; Jia et al. 2003; Su et al. 2005) and foundto accurately reproduce the measured fluxes.However to date, validation studies have beenpredominantly undertaken as local scale evalu-ations over specific hydroclimatic and land sur-face conditions.

Both observations and theoretical simulationshow that the climate condition and vegetationcover type have a major control on surface en-ergy flux patterns (Lafleur and Rouse 1995;Bridgham et al. 1999; MacKay et al. 2002). Atthe same time, changes in these energy fluxpatterns could have a complex effect on the cli-

mate variability (Delworth and Manabe 1993)and may ‘‘amplify or reduce the effects of poten-tial climatic change’’ (Eugster et al. 2000). Un-doubtedly, improved estimation of ET at largerscales is essential to improve our understand-ing of the global climate and its variability,both in the spatial and temporal domain(Miller et al. 1995).

SEBS requires evaluation against diversedata to assess its potential for routine globalscale evapotranspiration estimation. Specifi-cally, the SEBS model needs to be evaluated ina variety of climate zones and land covers,thereby assessing the adaptability of the SEBSmodel to climate and land cover variability. Tofulfill this goal, a validation dataset that satis-fies these criteria is required. The CoordinatedEnhanced Observing Period (CEOP) formsan element of the World Climate ResearchProgram (WCRP) (Koike 2002). Initiated aspart of the Global Energy and Water CycleExperiment (GEWEX), the CEOP dataset isideal for the requirements of the SEBS valida-tion, providing continuous high quality in-situmeasurements at various locations across theglobe.

In our study, eight CEOP stations are used toassess the SEBS model, specifically encompass-ing a variety of hydroclimatic conditions. Sincein-situ observations of the surface broadbandemissivity, leaf area index (LAI) and vegetationfraction (VF) are not routinely available fromthe CEOP dataset, these land surface parame-ters are derived from MODIS land products.An inter-comparison of energy fluxes fromSEBS predictions and in-situ eddy correlationbased observations are examined at daily and10-day scale for each station.

Spatially representative in-situ data is oftenonly available at a limited number of sites. Op-erational remotely sensed data and meteorolog-ical data from the Global Land Data Assimila-tion System (GLDAS) have been incorporatedinto the SEBS model to derive the evapo-transpiration, to alleviate model dependenceon ground measurements. In the satellite basedanalysis, MODIS LST is coupled with meteorol-ogy from CEOP and GLDAS (Rodell et al. 2004)to formulate two different forcing datasets forSEBS, offering a thorough examination of theimpact on ET prediction from varying scales ofmeteorological forcing data.

440 Journal of the Meteorological Society of Japan Vol. 85A

2. Model description

The Surface Energy Balance System (SEBS)model was proposed by Su (2002) to estimatethe surface energy balance and evaporativefraction using remotely sensed data in combi-nation with surface meteorological data. Cur-rently, a MODIS based terrestrial ET algo-rithm is under development mainly using theSEBS model. The required data variables bySEBS are listed in Table 1.

The net radiation, ground heat flux and evap-orative fraction are estimated in separate mod-ules in SEBS, although the different modulesshare some common variables, such as thevegetation parameters. A brief overview of themethods used to estimate the net radiation,soil heat flux and the partitioning of availableenergy into sensible and latent heat fluxes inSEBS is presented below, but the reader is re-

ferred to (Su 2002) for a more thorough expla-nation.

Net radiation is estimated using radiativeenergy balance at the land surface:

Rn ¼ ð1 � aÞRswd þ eRlwd � esT4s ; ð1Þ

where a is the broadband albedo in the visibleand near infrared band, e is the broadbandemissivity in the thermal infrared band, Rswd

is the incident solar radiation, Rlwd is the down-ward longwave radiation, Ts is the surface tem-perature and s is the Stephan-Boltzman con-stant.

In satellite based applications or at large re-gional scales, ground heat flux measurementsare generally unavailable. To account for thislack, an empirical parameterization of theground heat flux based on net radiation andthe vegetation fraction is used to estimate the

Table 1. SEBS model data requirements (The observations in italic are used for validation pur-poses).

Data Type Variables Possible Sources

Air temperature

Surface MeteorologyPressure In situ Observation, or Land Data Assimilation

DatabaseWind

Humidity

Incident Shortwave Radiation

Radiative Energy FluxOutgoing Shortwave Radiation

In situ Observation, or Satellite RetrievalDownward Longwave Radiation

Outgoing Longwave Radiation

Land Surface Temperature

Emissivity

Albedo

Land Surface Variables Vegetation Height In situ Observation, or Satellite Retrieval

Vegetation Fraction

Leaf Area Index

Vegetation Type

Ground Heat Flux

Surface Energy FluxSensible Heat Flux

In situ ObservationLatent Heat Flux

Net Radiation

February 2007 H. SU et al. 441

total ground heat flux for the pixel. If theground heat flux ratio is defined as:

G ¼ G0/Rn ð2Þ

Then, the ground heat flux can be parameter-ized following (Monteith 1973) and (Kustasand Daughtry 1990):

G0 ¼ Rn � ðGvfv þ Gsð1 � fvÞÞ ð3Þ

where Gv and Gs are the ground heat flux ratiosfor full vegetation covered area and for bare soilrespectively, fv is the vegetation fraction. Itshould be noted that Gs and Gv are dependenton vegetation types and climate conditions.

The partitioning of latent heat (LE) and sen-sible heat (H) from the available energy isachieved in part by solving the following set ofequations which are based on the similaritytheory of the atmospheric surface layer (Brut-saert 1982, 1999):

u ¼ u�k

�ln

z � d0

z0m

� ��Cm

z � d0

L

� �

þCmz0m

L

� ��

y0 � ya ¼ H

ku�rCp

�ln

z � d0

z0h

� �ð4Þ

�Chz � d0

L

� �þCh

z0h

L

� ��;

L ¼ � rCpu3�yv

kgH

where u is the wind speed, u� is the friction ve-locity, r is the air density, Cp is the specific heatof air at constant pressure, k is the von Kar-man’s constant, d0 is the zero plane displace-ment, z is the height above the surface, z0m

and z0h are the roughness height1 for momen-tum and heat transfer respectively, y0 and ya

are the potential temperatures at surface andat height z respectively, Ch and Cm are thestability correction functions for sensible heatand momentum transfer respectively, L is theObukhov length, g is the acceleration due to

gravity and yv is the virtual temperature nearthe surface.

3. Data and site description

In attempting the inter-comparison of re-sults, a number of different data sets were col-lated based on different sources of operationalmeteorology determined either from the CEOPor from the GLDAS data. These forcing data-sets are compiled and listed in Table 2, withboth the data sources and their scales de-scribed. The surface meteorological variablesinclude wind speed, humidity, pressure, airtemperature, downward shortwave and long-wave radiation. The meteorological data in Da-taset I was hourly observed from CEOP sites,while the Forcing Dataset II (CEOPþMODIS)and III (GLDASþMODIS) are daily instanta-neous, to match with the overpass time ofMODIS. The three different forcing datasetsrepresent a variety of scales, ranging in thespatial domain from the tower scale to theMODIS 1 km and from instantaneous to hourlyin the time domain. Basically, the error in themodel predictions is composed of two parts.One is caused by the scale disparity of forc-ing data; the other is due to the scale depen-dence of the model itself. We agree that it ishard to separately isolate the scale dependenceof the model predictions from the influences ofthe model and the measurements. It is chal-lenging in the study of ET, since spatial repre-sentative of the measurement of ET in naturalconditions ranges from tens of meters to kilo-meters, depending on both the heterogeneity ofthe site characteristics and the limitation of theinstruments. To separate the influences fromthe model analysis and measurements, one pos-sible method may be the numerical simulation,where the heterogeneity of the forcing is morecontrollable. In our current study, the ‘‘impacton ET prediction’’ is actually the comprehen-sive influences from the combination of modelanalysis and measurements.

The hourly in-situ measurements of the sur-face energy fluxes from CEOP sites are em-ployed to compare with the model predictionsfrom Forcing Dataset I. To compare with thesurface heat flux estimation from the other twoForcing Datasets, the measurements of the sur-face energy fluxes from CEOP sites are interpo-lated linearly, just like the processing of the

1 The initial values of the roughness lengths anddisplacement height are from a look-up tablewhich is adapted from land surface models(http://ldas.gsfc.nasa.gov/LDAS8th/MAPPED.VEG/web.veg.table.html). Then the roughness lengthsfor momentum and heat transfer are adjustedduring the iterative solution of Eq. (4).


corresponding meteorological data. Since mea-surements of the ground heat flux ðG0Þ are notdirectly available at most of the CEOP sites2,the prediction of the ground heat flux by SEBSis not assessed in this study.

In the following sections, further details onactual site characteristics and the processingof MODIS data is discussed.

3.1 In-situ data from CEOP and sitecharacteristics

The Enhanced Observing Period 2001 (EOP-1) provides hourly observation of surface mete-orological variables and energy fluxes from July1st through September 30th in 2001. Elevenbasic variables are required by SEBS from theCEOP dataset. These variables fall into twocategories: surface meteorological and radiativedata (i.e., air temperature, pressure, humidity,wind speed, downward and upward longwaveradiation, downward and upward shortwaveradiation) and surface energy flux data (i.e.,net radiation, sensible and latent heat fluxes)which are used to validate the model predic-tions. There are only eight stations (6 sites)from CEOP EOP-1 that meet the data require-ments of the SEBS model3, with site names andcharacteristics (including country, location, land

classification and climate type) listed in Table3. The six sites are distributed across five coun-tries and three continents. The dominant landcover types are listed in the last column inTable 3.

The climate types of the selected CEOP sites(see) with their names are listed in the 4th col-umn in Table 3. The six sites fall into threebroad climate types and four different vegeta-tion covers. Brief descriptions of each site aregiven below, with further details availablefrom http://www.eol.ucar.edu/projects/ceop/dm/.

a Cabauw (grassland)The Cabauw tower site is located in the cen-

tral region of the Netherlands. The surround-ings are flat and consist of meadows which areused for grazing and for the production of hay(Beljaars and Bosveld 1997). The grass at the

Table 2. Forcing datasets for SEBS.

Data Source Variables Spatial Resolution Temporal Resolution

CEOP Surface meteorology Tower scale Hourly

Forcing Dataset IEmissivity 1 km Instantaneous

MODIS LAI 1 km 8-day

Land Cover 1 km Yearly

CEOP Surface meteorology Tower scale Instantaneous, interpolated

LST/Emissivity 1 km Instantaneous

Forcing Dataset II MODIS LAI 1 km 8-day


Albedo 1 km 16-day

GLDAS Surface meteorology 14� Instantaneous, interpolated

LST/Emissivity 1 km Instantaneous

Forcing Dataset III MODIS LAI 1 km 8-day


Albedo 1 km 16-day

2 Soil heat flux measurements at 5 cm and 10 cmare available at two of the eight chosen stations,but require correction to the ground surface to beused as G0. Five of the eight stations have no ob-servations of soil heat flux. Only one station pro-vides sufficient ground heat flux measurements.

3 Southern Great Plains (SGP) sites in CEOP EOP-1 are excluded from this study, but are the focusof ongoing investigations.


measuring field is kept at a height of approxi-mately 8 cm by frequent mowing. The vegeta-tion cover at Cabauw is close to 100% all year.Caubauw is one of the sites in the Baltic SeaExperiment (BALTEX) and has been demon-strated to be a useful case study for mid-latitude homogeneous grassland within theProject for Intercomparison of Land-SurfaceParameterization Schemes (PILPS) framework(Chen et al. 1997; Henderson-Sellers et al.1995).

b Lindenberg (mixed forest and grassland)Lindenberg is located in the east of Germany

and also forms a member site of BALTEX. Theclimate type of Lindenberg is the same as Cab-auw, with heterogeneous land use dominantedby a mixture of forest (43%) and agriculturalfarmland (45%) with a number of small andmedium-sized lakes (7%) (Beyrich and Adam2004). The Lindenberg data for CEOP EOP-1include the near surface measurements carriedout at GM Falkenberg, which represent onlythe farmland (low vegetation, grassland) partof the area.

c Bondville (cropland)Bondville is located in central Illinois, USA

and is both a member site of the GAPP pro-gram (GEWEX America Prediction Project)

and a member of the AmeriFlux network. Theclimate of Bondville is temperate continentaland the vegetation type in the summer of 2001was predominantly corn. Although the MODISland classification shows the land cover atBondville is homogeneous, the MODIS pixelover this site is actually a mixture of corn andsoybean since there is a companion site atBondville which is located 400 m north of themajor site and is planted with opposite crop incorn/soybean rotation4. It is the only represen-tative agricultural site from the six CEOP sitesused here.

d Rondonia and Manaus (tropical forest)Both Rondonia and Manaus are in Brazil and

are member sites of the Large Scale Biosphere-Atmosphere Experiment in Amazonia (LBA).The vegetation covers for the two sites are pri-marily forest and rain forest (Andreae et al.2002). They are of tropical humid climate witha dry season in winter. The frequent cloudcover makes it extremely difficult to measurethe reflectance in the visible and near infraredbands from satellites, which in turn leads tomore uncertainties in LAI estimation. Both

Table 3. Characteristics of the reference sites.

Site Name Country Lat/Lon (�) KoppenClimate*Dominant Land Cover

(DLC)

Cabauw The Netherlands (51.97, 4.93) C Grassland

Lindenberg Germany (52.17, 14.12) C Grassland

Bondville USA (40.01, �88.29) D Cropland (Corn)

Rondonia Brazil (�10.01, �61.93) A Rain forest

Manaus Brazil (�2.61, �60.21) A Rain forest

BERMS(Old_Aspen)

Canada (53.63, �106.20) DForest(Old_Aspen)

BERMS(Old_Jack_Pine)

Canada (53.92, �104.69) DForest(Old_Jack_Pine)

BERMS(Old_Black_Spruce)

Canada (53.99, �105.12) DForest(Old_Black_Spruce)

*A: tropical; B: dry; C: warm temperate rainy climates and mild winters; D: cold forest climates and severewinters; E: polar; H: highland.

4 From AmeriFlux: http://public.ornl.gov/ameriflux /Site_Info/siteInfo.cfm?KEYID=us.bondville.02.


sites have a rather high LAI in the whole 3months, which ranges from 5 to 7.

e BERMS (cold forest)The study area of the Boreal Ecosystem

Research and Monitoring Sites (BERMS) inCEOP consist of three individual stations withvegetations covers of Old Aspen, Old Jack Pineand Old Black Spruce respectively. Old Aspenis deciduous while the other two are needleleaf forest, which inhabit the cold mid-latitudeclimate typical of BERMS. All the in-situ towerflux measurements from each of the three sta-tions during CEOP EOP-1 were used in ourstudy.

3.2 Surface meteorological data from GLDASThe surface meteorological data in Forcing

Dataset III were extracted from the GLDASforcing dataset over each CEOP site. The sur-face forcing fields from GLDAS is an integra-tion of observational fields and outputs fromthe atmospheric data assimilation system(ADAS) component of a weather forecast andanalysis system (Rodell et al. 2004). Six basicvariables are required by SEBS from theGLDAS forcing dataset, including the air tem-perature, humidity, pressure, wind speed,downward longwave and incident shortwaveradiation. The surface meteorological and radi-ative variables from GLDAS were interpolatedin the time domain to match with the instanta-neous MODIS LST. GLDAS forcing, in combi-nation with the necessary MODIS data, areused to form Forcing Dataset III which is moreoperationally available than the Dataset I andII.

3.3 Data from MODISTo estimate the terrestrial ET globally, oper-

ational remotely sensed data need to be uti-lized, specifically LAI, vegetation fraction, sur-face broadband emissivity and Albedo whichare derived from MODIS data. These MODISbased remote sensing measurements can pro-vide SEBS the required land physical parame-ters: LAI (Myneni et al. 2002), albedo (Schaafet al. 2002), vegetation index (Huete et al.2002), land classification (Friedl et al. 2002)and land surface temperature and emissivity(Wan et al. 2004). In Forcing Datasets II andIII, MODIS data, such as MODIS LST and al-

bedo, are incorporated to form a more opera-tional forcing data for SEBS.

a LAI, vegetation fraction and land coverLAI describes an important structural prop-

erty of a plant canopy, defined as the one sidedgreen leaf area per unit ground area (Chen andBlack 1992). The MOD15 Leaf Area Index(LAI) is a 1 km global data product updated us-ing an 8-day window (Myneni et al. 2002). Col-lection 4 of MODIS LAI product has been eval-uated by (Cohen et al. 2003) and (Wang et al.2004). Some abrupt changes of the MODIS de-rived LAI were found over the CEOP referencesites during the summer season of 2001, espe-cially for the forest sites in this study. Whenvegetation densities are high, there is less con-fidence in the output value of the assigned LAIsince the reflectance belongs to the ‘‘saturationdomain’’ and the low values of LAI, such as inAmazon forest, may also be caused by frequentcloud contamination (Myneni et al. 2002). Tominimize the instability of the MODIS LAItime series extracted from a single pixel, anaveraged LAI over a 3 by 3 pixel box centeredon each CEOP reference site is used as a surro-gate. The dynamics of MODIS derived LAIfrom a 3 by 3 pixel box corresponding to eachof the 8 CEOP stations in summer of 2001 areshown in Fig. 1. The LAI in Lindenberg, whichis a mixure of grassland and forest, is found tobe much lower than that in the grassland Cab-auw. The low LAI estimates from MODIS atLindenberg may be caused by the misclassifica-tion of the vegetation cover or the algorithm ofMODIS LAI itself. One real point that needs at-tention is the temporal /spatial consistency ofthe MODIS datasets. The abrupt change ofLAI, which can be identified in Fig. 1, can notbe explained by the growing of the plants ex-cept the old aspen station. Even forest fire can’tgive a reasonable explanation. It might becaused by the MODIS LAI data product itself.Clearly, the sacrifice of spatial resolution toachieve a smoothed time series of LAI may beproblematic if the land cover of the site foot-print is highly heterogeneous. To maintain con-sistency with the hourly observation data fromCEOP, the 8-day LAI values are linearly inter-polated to the daily scale. Vegetation fraction iscomputed based on LAI with an assumption ofa particular leaf angle distribution for a given


vegetation type which can be obtained fromMODIS land cover data product.

b AlbedoThe MODIS based albedo product MOD43

(Schaaf et al. 2002) is used in Forcing DatasetsII and III. There are two different albedo pro-vided by MODIS: black-sky and white-sky al-bedo. Ideally, these require column water vaporand aerosol concentration to force the atmo-spheric radiative transfer model which parti-tions the direct and diffuse incident solar en-ergy to determine the overall surface albedo.In this current study, the black-sky and white-sky albedo are averaged to derive the overallsurface albedo, since the parameters requiredto do the atmospheric radiative correction arenot readily available.

The simplification of the albedo estimation invegetated areas is not expected to introducesignificant error in flux estimation. Betts andBall (1997), undertaking a study of surface Al-bedo in Boreal forest, formulate the relation be-tween daily averaged albedo and the diffuseflux ratio of insolation:

a ¼ A þ B � D#

S#

!ð5Þ

where a is the daily averaged surface albedo, Aand B are two coefficients derived by regres-sion, D# is the daily diffuse incident solar en-ergy and S# is the daily total incident solar en-ergy. The coefficient A varies with differentvegetation types. B is found to be in the rangeof �0.009 to �0.018 for grass, aspen, black

spruce and jack pine when there is no snowcover (Betts and Ball 1997). The low value ofcoefficient B implies that the diffuse flux ratioof insolation does not play a significant role inthe overall surface albedo. Absolute value of co-efficient B is less than 0.018, which means thatthe error of the estimated overall albedo iswithin 0.009 if the direct and diffuse solar radi-ation ratio is assumed to be 1 :1. The assump-tion of equal direct and diffuse solar radiationis equivalent to our simple averaging method.

c Broadband emissivity and LSTRight now, there is no measurement of the

broadband emissivity at site scale from theCEOP reference sites. The only possible way toobtain the emissivity is to get it from the satel-lite data product, such as MODIS data, al-though it is at a larger spatial scale. The broad-band emissivity is derived from the dailynarrow band emissivity of MODIS bands 29,31 and 32 (Wan and Li 1997; Wan et al. 2004).The formula of the emissivity conversion (pers.comm. Dr. Z. Wan 2004) is as follows:

e ¼ 0:2031 � e29 � 0:0648 � e31 þ 0:8602 � e32 ð6Þ

where e is the broadband emissivity and e29, e31

and e32 are the narrow band emissivities forbands 29, 31 and 32 respectively. The broad-band emissivity is interpolated linearly to fillin the gaps in Forcing Dataset I when theMODIS derived broadband emissivity is absentin order to match up with the CEOP dataset attimes other than MODIS overpass.

For Forcing Dataset I, in-situ surface temper-ature observations are only available at Bond-

Fig. 1. LAI dynamics derived from MODIS at CEOP stations.


ville and Rondonia. For locations where thesurface temperature ðTsÞ is not observed di-rectly, it is estimated indirectly from the up-ward longwave radiation ðRlwuÞ and downwardlongwave ðRlwdÞ, using a correction for thebroadband emissivity:

Ts ¼Rlwu � ð1 � eÞ � Rlwd

es

� �1/4

ð7Þ

which is derived from the longwave radiationbalance equation.

The land surface temperature in Forcing Da-tasets II and III is obtained directly from the1 km instantaneous MODIS (onboard TERRA)LST product.

4. Results

Usually, remote sensing based ET model isevaluated in a very short time period; typically,the experiments are within days or weeks. Inour study, 8 stations which are of differentland cover and climate zones are used to evalu-ate our ET model in the 3 month period. Wedon’t worry about the number of observationsin our study. In addition, rRMSE (relativeRoot Mean Square Error) is also employed toassess the performance of the ET modeling. Be-cause of the large uncertainty in ET modelingor even in the ET measurement on ground, itis hard to set a standard to judge whether ornot a model is good. Instead, the usual way isto have a look at the bias and RMSE to see ifit is comparable with those for ET measure-ment. Experiment showed that, two collocatedET measurement equipments may give a devia-tion of about 10–15% on ET (Lloyd et al. 1997).

Evapotranspiration predictions were obtainedfor each of the three different forcing datasetslisted in Table 2. In this section, the ET estima-tion at both the tower and satellite scales arepresented and analyzed. SEBS predicted ETwhich was based mainly on CEOP observationsis examined at daily and 10-day time scales.The MODIS based instantaneous ET estima-tion from Forcing Datasets II and III is alsoevaluated with the CEOP tower flux measure-ments.

4.1 Tower scale estimationSEBS model produced hourly predictions of

daytime surface energy fluxes for the eightCEOP stations described above. Intercompari-

son of the remaining heat fluxes is undertakenat daily and 10-day time scales.

a Daily comparisonDaily averages of the energy fluxes for each

of the eight stations are presented in Fig. 2.Since the SEBS model does not predict the sur-face energy fluxes at night, fluxes are only cal-culated during the daytime, specifically from5:00 a.m. through 6:00 p.m. in local time ateach site. The time period in daytime is chosento ensure that all valid data can be used in theanalysis. In high latitude region in north hemi-sphere, there is a much longer daytime in sum-mer. To filter out periods of inclement weather(cloud affected or rainy days), the daily averageis computed only for those days when both theobservations and the model predictions areavailable for more than 4 hours per day. Thenumbers of available days in the three monthsand the bias of SEBS predicted daily heatfluxes are computed for each station and la-beled in each panel in Fig. 2.

The RMSE of the daily average of the net ra-diation estimation is within 20 W m�2 for alleight stations, since the four components of ra-diation flux measurements (Rswd, Rswu, Rlwd

and Rlwu) are used directly in this applicationof SEBS to derive net radiation. The largestbias of daily ET estimation is found in BERMS(Old Aspen) with a negative bias of 28 W m�2;the largest sensible heat bias is at Rondoniawhich presents a positive bias of 29 W m�2.Both are forest sites. No systematic bias of ETestimates is found at the 8 different sites.

Statistics of the heat flux comparison be-tween SEBS predictions and CEOP observa-tions on daily scale are presented in Table 4.Specifically the root mean square (RMSE) errorand the relative RMSE (rRMSE) of the cor-responding SEBS predication, the mean of theobserved and SEBS derived daily daytimeaveraged surface heat fluxes (H and LE), andthe daytime averaged evaporative fraction fromCEOP observation and SEBS estimates arelisted in Table 4 for all eight stations. The biasof model predictions on H and LE for each sitecan be found in the corresponding panel in Fig.2. The rRMSE is chosen here as one of the cri-teria together with root mean squared error(RMSE), to evaluate the accuracy of model pre-diction and it is defined following Conte et al.


(1986) as:

rRMSE ¼ RMSE

MeanðObservationÞ ð8Þ

The performance of SEBS in different landcover and climatic classifications is assessed

mainly based on the two criteria mentionedabove. The minimum RMSE of daily daytimeaveraged LE estimation is 23 W m�2 at Linden-berg, which is a grassland site. The correspond-ing rRMSE of daily LE estimation at this site is18.9%. For the other grass site, Cabauw, the

Fig. 2. Comparison of daily (5 H–18 H in local time) averaged energy fluxes.


Ta

ble

4.

Sta

tist

ics

ofth

ed

ayti

me

Hea

tF

luxes

[Wm

�2]

an

dE

va

por

ati

ve

Fra

ctio

n(E

F)

at

the

da

ily

sca

le(5

H–

18

Hin

loca

lti

me)

.

Sit

eN

am

e

RM

SE

ofS

EB

SL

E[W

m�

2]

rRM

SE

ofS

EB

SL

E

Mea

nof

CE

OP

LE

[Wm

�2]

Mea

nof

SE

BS

LE

[Wm

�2]

RM

SE

ofS

EB

SH

[Wm

�2]

rRM

SE

ofS

EB

SH

Mea

nof

CE

OP

H[W

m�

2]

Mea

nof

SE

BS

H[W

m�

2]

Mea

nE

Ffr

omC

EO

PM

ean

EF

from

SE

BS

No.

ofD

ays

Ca

ba

uw

30

20

.6%

14

61

44

27

58

.4%

47

59

0.7

56

0.7

08

45

Lin

den

ber

g2

21

8.9

%1

19

10

22

66

2.4

%4

16

10

.74

10

.62

57

0

Bon

dvil

le7

42

4.7

%2

99

29

35

72

9.0

%1

98

21

00

.60

20

.58

22

3

Ron

don

ia4

01

7.9

%2

26

23

53

54

1.4

%8

51

14

0.7

25

0.6

72

66

Ma

na

us

50

15

.9%

31

83

44

28

27

.6%

10

29

50

.75

80

.78

73

1

BE

RM

SO

ldA

spen

48

31

.9%

15

11

22

32

26

.2%

12

41

16

0.5

48

0.5

76

84

BE

RM

SO

ldJ

ack

Pin

e2

92

8.8

%1

01

10

64

52

0.5

%2

20

19

10

.31

50

.27

87

6

BE

RM

SO

ldB

lack

Sp

ruce

56

42

.1%

13

41

18

53

22

.6%

23

72

51

0.3

62

0.4

02

62


rRMSE of daily LE estimation is 20.6%, whichshows no significant difference with that of Lin-denberg. It demonstrates that SEBS workspretty well to predict the ET over the grass-land. The largest RMSE of daily LE predictionfrom SEBS is found at Bondville (corn site)with a value as high as 74 W m�2, the rRMSEat Bondville, however, is only 24.7%. Consider-ing that it has a much higher daily averagedET in its rapid growing period in summer andthe site heterogeneity which will be discussedlater in Section 5, rRMSE of 24.7% at the cropsite is a reasonable model prediction. The leastaccurate reproduction of daily LE relative to in-situ measurements is found at the Old BlackSpruce station of BERMS, with a value of42.1%. For the other two stations at BERMS(Old Aspen and Old Jack Pine), the rRMSE ofthe daily LE are 31.9% and 28.8% respectively.One of the explanations for the high rRMSE ofthe daily LE estimation over the BERMS re-gion is that the mean daily LE of the CEOP ob-servation is quite low in the cold region, in therange of 102 to 151 W m�2, for the Boreal for-est. Considering the rRMSE of the daily aver-aged LE for each site, the accuracy of the LEestimation for tropical forest is the highest,then the grasslands, followed by the corn site.The LE estimation for the Boreal forest in thehigh latitude region has the largest uncertaintyfrom the perspective of rRMSE.

The rRMSE of the daily sensible heat (H) es-timation are generally higher than those ofdaily LE prediction except for the threeBERMS stations. One reason may be that themean values of H are relatively larger atBERMS, considering that the three BERMSstations have the lowest mean evaporative frac-tions (EF) among all the stations. The mean EFin Table 4 is defined as:

MeanðEFÞ ¼P92

i¼1 LEiP92i¼1ðLEi þ HiÞ

ð9Þ

LEi and Hi are the daytime averaged surfaceheat fluxes for each day during EOP-1 and theith data is not used in the calculation whenthe data is not available. The comparison ofthe mean EF shows that estimates from SEBSagree very well with those from observationsand exhibit no significant bias. Both the modelpredicted and observed mean evaporative frac-

tions at the two tropical sites are above 0.65during the summer of 2001, which is reason-able considering the climatic conditions in theAmazon, where water supply is sufficient forsustained forest growth. Cabauw and Linden-berg also have very high mean EF in summer.For the Cabauw grassed site, the soil is re-ported to be at field capacity all year long andthe evapotranspiration is seldom limited bythe water supply (Chen et al. 1997). The largestbias of the mean EF estimation is 0.111 (nega-tive bias) at Lindenberg. Generally, the predic-tion of mean EF from SEBS illustrates a closeagreement with that from observations.

From Fig. 2 and Table 4, it is shown thatboth the climate and the vegetation type havean important control over the daytime surfaceheat flux patterns. It is true that similar con-clusion has been made by others before (Lafleurand Rouse 1995; Bridgham et al. 1999), basi-cally at a longer time scales, such as seasonalor annual. The finding here is still meaningfulsince it is confirmed again by a model (SEBS)which is driven by different input data (combi-nation of surface meteorology and satellitedata) to predict the terrestrial ET. The assess-ment of the model’s performance in reproducingthe different ET patterns under different cli-mate and vegetation types is essential to ex-pand the ET predictions from site scale to thescales of satellite data. The SEBS model is ob-served to reproduce the patterns of the dailysurface heat flux under diverse climate condi-tions and vegetation types. While estimates ofthe sensible and latent heat fluxes in SEBS forthe Boreal forests are not as accurately repro-duced, such vegetated stands represent someof the most difficult surface types over which tomodel ET, due to their strong coupling with theatmosphere (Margolis and Ryan 1997) and un-certainties in retrieving required variables.However, for the most part (6 out of the 8 sites),estimates of the daytime ET agree within50 W m�2 or 20%.

b Ten day averaged comparisonThe satellite based modeling of ET is differ-

ent with those based on mass balance in landsurface models, hydrological models or ecolog-ical models. Instead, it is based on instanta-neous energy balance. There is a gap betweenthe time scale of ET prediction based on satel-


lite and the requirement of its related applica-tion, for example in draught monitoring, agri-cultural yield estimation. The applicationsoften need ET estimation in different levels oftime scale, ranging from hourly to monthly.The satellite based instantaneous ET predic-tions need to be expanded in temporal domain.A 10-day global ET prediction based on satel-lite observations is part of the planned ETdata product, which can bridge the temporalgap between remotely sensed observations andthe needs of hydrological, climatological andagricultural studies. Thus, testing the ETmodel at different time scales where in situmeasurement is available is very helpful to en-hance the later application of the ET data.

Ten day cloud free daytime averaged sensibleand latent fluxes from SEBS predictions andthe corresponding CEOP measurements areshown in the bar plots of Fig. 3. Surface heatflux at 10-day time scale was examined sinceremotely sensed data from MODIS, especiallyland surface temperature, is unlikely to beavailable over a spatially continuous extent atthe daily time scale. Daytime is defined from5:00 a.m. through 6:00 p.m. in local time foreach site. Results are not presented when ei-ther the CEOP observations or the model pre-dictions are unavailable during the 10-dayperiod. Similar to Fig. 2, net radiation compari-son between the SEBS model prediction andthe CEOP observation are not shown, since thedeviation of 10-day cloud free averaged net ra-diation predictions is below 20 W m�2. Figure3 presents the bar plots comparing heat fluxesfrom SEBS and CEOP observations. The twocurves in each panel show the 10-day cloudfree daytime averaged evaporative fractions(EF) derived from the SEBS model outputsand the CEOP observation.

The variations of the evaporative fraction at10-day scale can also be examined from Fig. 3.Cabauw, Lindenberg, Rondonia and Manausshow relatively flat curves of evaporative frac-tion, while the evaporative fraction curves showmore temporal variability at the other 4 sta-tions. At the Bondville station and 3 BERMSsites (Old Aspen, Old Jack Pine and Old BlackSpruce), the peak EF is found towards mid-lateJuly, with EF tending to decrease after thistime. The low temporal variability of evapo-transpiration at stations (Cabauw, Lindenberg,

Rondonia and Manaus) may be due to the suffi-cient water supply. The deviation of the pre-dicted and observed LE in July at Lindenbergis much greater than the other two months.The underestimation of LE in SEBS will be dis-cussed further in the Section 5.

The EF patterns at the Old Aspen forest(panel f ) are dissimilar to those at the otherBERMS sites (panel g,h) in that the Aspen for-est exhibits the largest intra-seasonal variationof the three BERMS stations, with the evapora-tive fraction 0.2 units larger than the othersduring July, despite all located in the Borealforest climate zone. One possible reason is thatproperties of the species of the cold forest aredifferent, for example, different vegetationheight corresponding to different roughnesslength in the momentum and turbulent heattransfer. The ET prediction from SEBS verifiesagain that deciduous species represents ahigher water flux than the coniferous speciesin Boreal forest, which was revealed by pre-vious Boreas study (Margolis and Ryan 1997).Another reason is that the Old Aspen forest inBERMS consistently has the highest LAI inthe 3 months among the three forest stations(refer to Fig. 1). In July and August, the LAI atOld Aspen forest is nearly 3 times that at OldJack Pine forest and 1.5 times the Old BlackSpruce forest. In September, the LAI at allthree forest stations begins to decrease, butthe LAI at Old Aspen remains significantlylarger than that of the other two stations untilthe end of September. LAI, together with vege-tation fraction, have a direct impact on the par-titioning of available energy into sensible andlatent heat fluxes in the SEBS model and areamongst several of the required model parame-ters which can be provided from remote sensingdata.

Table 5 shows the statistics of the heat fluxesand evaporative fraction intercomparison at the10-day time scale. The RMSE and model bias ofthe 10-day averaged heat fluxes from SEBShave been improved significantly from those ofthe daily averaged heat fluxes for all 8 CEOPstations. The 10-day averaged model predictedheat fluxes from SEBS have the largest devia-tion at Bondville (a corn site) with a RMSE of47 W m�2 for LE and 43 W m�2 for H (see Ta-ble 5). The rRMSE of LE and H at Bondvilleare 14.0% and 25.4% respectively, a result of


Fig. 3. Comparison of 10-day cloud-free daytime averaged Heat Fluxes (a: Cabauw, b: Lindenberg, c:Bondville, d: Rondonia, e: Manaus, f: BERMS (Old_Aspen), g: BERMS (Old_Jack_Pine), h: BERMS(Old_Black_Spruce)).


relatively high values of surface fluxes at thesesites (see panel c in Fig. 3). RMSEs of LE arebelow 40 W m�2 at all remaining stations,while for H, the RMSEs are below 32 W m�2,demonstrating that the predictions of heatfluxes from SEBS are in good agreements withthe observations at 10-day time scale. The meanevaporative fractions at 10-day time scale inTable 5 show no significant difference withthose in Table 4, which is at daily scale. Theslight difference is due to the fact that somedaily data points may have been discarded dur-ing the aggregation process.

The 10-day cloud free daytime averaged en-ergy fluxes and evaporative fraction deter-mined by SEBS are in closer agreement withobservations at all eight CEOP stations thanwhat estimated at the daily scale. The latitudi-nal and intra-seasonal variation of the heatfluxes caused by vegetation types and climateconditions can be reproduced very well bySEBS. For example, the forest in Amazon dem-onstrates different patterns of evapotranspira-tion than the Boreal forest. Different types oftrees in the same Boreal area show distinct EFpatterns in both the SEBS prediction and theCEOP observation.

It has to be noted that the improvement of

the RMSEs for ET predictions when the timescale varies from daily to 10-day is actually in-duced by two processes. One is the statisticaltemporal averaging (mathematical process) andthe other is the changing course of the daily ET(physical process). The two processes are actu-ally coupled together and are hard to be sepa-rated, since the time series of the daily ET isauto-correlated. The detail analysis of the tem-poral scaling behavior of ET needs to use landsurface or hydrological models which can pre-dict ET continuously in time domain. Or thesite observations of ET can be used, althoughthe analysis will be limited to a certainlocation/point in this case. The temporal scal-ing behavior is an interesting topic and willnot be further discussed in this paper.

4.2 Satellite based estimation of ETMODIS based ET estimation were deter-

mined using both CEOP and GLDAS data (seeTable 2). The surface meteorological and radia-tive data in both CEOPþMODIS (Dataset II)and GLDASþMODIS (Dataset III) were inter-polated to correspond with the overpass timeof MODIS LST data. Since the 1 km MODISLST was used, the MODIS based ET estima-tion represents a spatial scale of 1 km. The

Table 5. Statistics of the daytime Heat Fluxes and Evaporative Fraction (EF) at the 10-day scale.

Site Name

RMSE ofSEBS LE[W m�2]

Mean ofCEOP LE[W m�2]

Mean ofSEBS LE[W m�2]

RMSE ofSEBS H[W m�2]

Mean ofCEOP H[W m�2]

Mean ofSEBS H[W m�2]

MeanEF fromCEOP

MeanEF from

SEBS

No.of

Data

Cabauw 12 138 138 8 43 48 0.760 0.739 9

Lindenberg 18 116 100 23 42 61 0.734 0.620 9

Bondville 47 325 331 43 175 186 0.650 0.639 8

Rondonia 36 217 236 27 88 114 0.709 0.674 9

Manaus 35 321 353 13 110 102 0.743 0.774 7

BERMS(OldAspen)

35 149 177 23 125 132 0.542 0.573 9

BERMS(Old JackPine)

16 99 93 31 215 243 0.315 0.277 9

BERMS(Old BlackSpruce)

33 136 157 30 231 213 0.371 0.425 9


three sites with the most available and validestimates of ET are chosen to present the com-parison between 1 km SEBS outputs and towerscale CEOP observations. It is instructive torealize that in practice, surface temperatureavailability is strongly linked to geographic lo-cation and climate type. For example, at Man-aus, only 26 MODIS LST data were availablein 92 days. When considering the quality con-trol of MODIS LST and emissivity, the avail-able data become even less. Scatter plots ofcomparisons of MODIS based retrievals andin-situ flux measurements are shown in Fig. 4.The number of available model predictions andthe RMSE are included in each panel. Addi-tional statistical analysis of the comparisons,including average, bias, mean absolute devia-tion (MAD), can be found in Table 5. The re-sults at Old Jack Pine station was used to rep-resent the BERM cold forest since the largestnumber of valid outputs was found in this sta-tion. The three sites represent quite differentland classifications, including grassland, crop-land and cold forest respectively. Abnormal la-tent heat flux predictions, such as those resultsout of the range from zero to the net radiation,are discarded. More predictions from the SEBSmodel have been obtained from CEOPþMODISdataset than from GLDASþMODIS dataset, al-

though they share the same remotely sensedland surface parameters.

At all the three sites, the accuracy of the ETpredictions from CEOPþMODIS dataset is bet-ter than those from GLDASþMODIS dataset.Specifically, the results from CEOPþMODISdataset have smaller bias, MAD and RMSE,which is perhaps expected, considering GLDASmeteorological forcing represents a coarserspatial and temporal resolution (1/4� and 3-hourly). The RMSE of ET estimation at Cab-auw and BERMS based on CEOPþMODISdataset is around 61 W m�2; the largest RMSEof ET estimation from CEOPþMODIS datasetis found at Bondville, which reaches 96 W m�2.The RMSE of ET estimation is increased by20@30 W m�2 at Cabauw and BERMS whenCEOP forcing is replaced by GLDAS forcing inDataset III. For Bondville, the RMSE of ETestimation from GLDASþMODIS is nearly45 W m�2 larger than that of CEOPþMODIS.The absolute values of ET bias are also in-creased by the incorporation of GLDAS forcing.The bias of ET estimation at Cabauw, Bond-ville and BERMS from CEOPþMODIS datasetis below 21 W m�2, which is reasonably small.The bias of ET estimation at the three sitesbased on GLDASþMODIS dataset are all nega-tive, within which the values of ET bias at

Fig. 4. Comparison of MODIS based ET and CEOP observation (Unit: W m�2)


Bondville and Cabauw are quit large (morethan 70 W m�2). The potential reasons for theconstant negative bias of ET estimation fromthe GLDASþMODIS dataset and the largerbias of ET estimation, especially at site Bond-ville, will be discussed in detail in the followingsection.

5. Discussion

There are many factors that can affect the ac-curacy of model predicted energy fluxes. In thissection, we will firstly discuss the uncertaintiesin the prediction and the validation of energyfluxes which may originate from remote sens-ing data, scaling issues, model parameteriza-tion and ground observation. Then based onthe analysis of the results, the future work andimprovement of the SEBS model is outlined.

With reference to remote sensing data, someof the required model parameters such as LAI,vegetation fraction and broadband emissivityare derived from MODIS data. The evaporativefraction for a corn field may increase from 0.50to 0.90 within three weeks in the rapid growingseason according to data from the SMEX02field study (Su et al. 2005). LAI from MODIS isretrieved every eight days, which may be toocoarse to capture stressed, senescent or ex-tended growing periods of some vegetated sur-faces such as Bondville. The LAI dynamics dur-ing EOP-1 shown in Fig. 1 demonstrate thetemporal variation of the MODIS LAI duringthe course of the period studied here. Consider-ing all stations are vegetated areas, the emis-sivity values are reasonable in the growingseason. However, it should be noted that thesuggested broadband emissivity conversion al-gorithm has not been validated.

As observed in panel (b) of Fig. 3, there was adisparity in the observed and modeled evapora-tive fractions during July at Lindenberg. Al-though both sites are classified as grassland,the LAI at Cabauw is almost twice that of Lin-denberg from mid-July to mid-August, whilethe LAIs are most closely matched during Sep-tember and late August. Such a difference maygo someway towards explaining observed differ-ences, being the product of an underestimationof the MODIS LAI at Lindenberg. The landcover classification from MODIS data showsthat the 3 � 3 km2 area at Lindenberg consistssolely of grassland. While this may be true at

1 km scale of MODIS, the recent report byBeyrich and Adam (2004), reveals that theland cover of the footprint of Lindenberg site isactually not homogeneous, but rather is a mix-ture of grassland and forest. It is not expectedthat sub-pixel heterogeneity is able to be de-tected by the 1 km MODIS land cover data, orindeed the surface temperature information(McCabe et al. 2005). When the footprint of thesite exhibits heterogeneous characteristic, re-mote sensing data with higher spatial resolu-tion may be more helpful.

Another source of uncertainty is a result ofscaling, related to the incorporation of the re-motely sensed data. Scaling has been recog-nized as an important issue in both remotesensing and land surface models, but remainslargely unresolved (Famiglietti and Wood 1995;McCabe et al. 2005). The MODIS data used inthis study has a spatial resolution of 1 km (foraggregated LAI, 3 km) and GLDAS forcing is0.25 degree (@ 25 km). The spatial representa-tiveness of tower measurements is not as easyto determine, even though CEOP stations areestablished in relatively homogeneous areas.Bondville site was actually a mixture of cornand soybean in the 1 km MODIS pixel, sincethe two stations at Bondville, which are 400 mapart, were planted with opposite crops (cornand soybean) in 2001. It is very common tofind that corn exhibits an ET which is100 W m�2 larger than soybean in their grow-ing season (Su et al. 2005), thus the heteroge-neity of landcover in a MODIS pixel may ac-count for the relatively larger deviation of ETestimation at Bondville from all three forcingdatasets. Remote sensing data with finer reso-lution, such as Landsat data, has proved to beable to eliminate part of the uncertaintiescaused by the spatial heterogeneity (Su et al.2005; McCabe et al. 2005). Additionally, the un-certainty of the heat flux measurements shouldnot be neglected. For example, during the Hy-drologic Atmospheric Pilot Experiment (HA-PEX) in the Sahel, the hourly summed fluxwas observed to have a deviation of 20% for la-tent heat fluxes from eddy correlation measure-ments (Lloyd et al. 1997). From their analyses,ET observations taken under ideal site condi-tions are likely to achieve accuracies in therange 10% to 20%.

From Table 5, it is shown that ET estimation


based on GLDASþMODIS has a significantnegative bias at Bondville and Cabauw, com-pared with the observations from CEOP. Thecomparison of incoming shortwave and long-wave forcing at the three sites (Cabauw, Bond-ville and BERMS) between GLDAS forcing andCEOP observation is shown in Table 6. The sta-tistics of the net radiation estimated based onGLDAS forcing is also listed in this table. Thedata is counted only when GLDAS data give avalid prediction of net radiation. The numberof available net radiation estimates fromGLDAS is larger or equal to that of the ET esti-mation which is shown in Table 6 since the par-tition of available energy requires stricter con-ditions than net radiation. The bias of netradiation estimation from GLDAS forcing atCabauw, Bondville and BERMS are �79, �127and �54 W m�2 respectively. The largest biasis found at Bondville, which corresponds tothe largest deviation of ET estimation usingGLDASþMODIS. Because ET estimation isoften less reliable than the estimation of net

radiation, it is imperative that an accurate netradiation estimate be available. Further inves-tigation of the incoming shortwave and long-wave radiation in Table 7 reveals that theincoming shortwave and longwave have a nega-tive bias larger than 66 W m�2 over all threesites. The negative bias of incoming shortwavecan explain the negative bias of net radiation.Among them, Bondville has the largest nega-tive bias for incoming shortwave, which is aslarge as 160 W m�2, more than twice of thebias at the other sites. The poorer estimationof ET at Bondville and Cabauw is causedmainly by the large error in the GLDAS down-ward radiative forcing. The 1

4 degree meteoro-logical data in GLDAS forcing are eitherderived by model outputs or from the interpola-tion of observations by a sparse network of sta-tions, which means their accuracy may not beas high as that of the CEOP observations, espe-cially for the downward radiative forcing. Theadvantage of the GLDAS forcing data is that ithas a global coverage and it is operationally

Table 6. Statistics of the MODIS based ET estimation [W m�2].

Forcing Datasets Site Name No. of Data Mean OBS Mean SEBS Bias RMSE

Cabauw 18 180 175 �5 61

IIBondville 27 332 353 20 96


60 121 131 10 60

Cabauw 16 190 142 �48 81

IIIBondville 16 401 330 �71 140


27 133 113 �20 84

NLDASþMODIS Bondville 15 358 344 �13 121

Table 7. Statistics of GLDAS based surface radiative forcing and the net radiation estimation[W m�2].

Incoming Shortwave Downward Longwave Net Radiation

Site Name Mean OBS BiasNo. ofData Mean OBS Bias No. of Data Mean OBS Bias

No. ofData

Cabauw 660 �78 30 377 �34 30 456 �79 30

Bondville 875 �160 24 372 �12 24 628 �127 24

BERMS 723 �66 36 332 �9 36 542 �54 36


available. The bias of ET estimation fromGLDAS and MODIS at Cabaw and BERMSare �48 and �20 W m�2, and their RMSE are82 and 84 W m�2. When the problem of nega-tive bias of the incoming radiative forcing inGLDAS is solved or offset, the accuracy of ETestimation based on the GLDASþMODIS forc-ing data (Dataset III) is expected to be im-proved, considering the large values of the biasin the incoming shortwave radiation data. Fur-ther investigation was done at Bondville, whichis the only site in USA. North America LandData Assimilation System (NLDAS) meteoro-logical and radiative forcing was introduced toreplace the GLDAS forcing. The bias of theradiative forcing from NLDAS at Bondvilleis within 20 W m�2. The statistics of the cor-responding ET estimation based on theNLDASþMODIS at Bondville is listed in thelast row of Table 6. The bias of ET dropped dra-matically to �13 W m�2. The RMSE based onthe new dataset was about 121 W m�2 whichis about 20 W m�2 lower than that fromGLDASþMODIS. The ET estimation is accept-able at this site, considering a relative RMSE of33.7% was obtained at Bondville with a hetero-geneous vegetation cover.

6. Conclusion

The adaptability of the SEBS model to cli-mate and land cover variability has been as-sessed for the first time at both tower scaleand 1 km scale based on CEOP EOP-1, GLDASforcing and MODIS data. The in situ measure-ments over the CEOP reference sites provide abaseline dataset for the evaluation of ET mod-eling at multi-scales. The tower scale ET esti-mation at daily and 10-day time scales showsthat the SEBS model can predict the daytimeenergy fluxes with accuracies comparable tothe eddy correlation measurements for theeight CEOP stations. These results are particu-larly pleasing given the wide variety of climateconditions (tropical to cold mid-latitude) andvegetation types (grassland, corn, rain forestand Boreal forest). The model is also evaluatedusing more operational forcing datasets, basedon MODIS and GLDAS forcing data. The un-certainties in ET estimation has been discussedand identified. It was revealed that the radia-tive forcing in GLDAS is critical to using thisdata source for ET prediction. For many sites,

satellite land cover observations with finer res-olution than MODIS is needed to capture thespatial heterogeneity, especially in areas, suchas Bondville, which are cropped. Promising re-sults have been obtained based on the CEOPand MODIS based forcing dataset, which pro-vides a baseline evaluation before the SEBSmodel is used to derive global land surfaceevapotranspiration product from operationalforcing data. A more reliable surface radiativeforcing with a global coverage is needed to esti-mate the global terrestrial evapotranspirationoperationally, together with currently availableGLDAS surface meteorology and MODIS basedsurface variables.

Acknowledgements

The authors would like to express the appre-ciation to the teams of the CEOP referencesites: Cabauw/BALTEX, Lindenberg/BALTEX,Bondville/GAPP, Manaus/LBA, Rondonia/LBAand BERMS/MAGS in the EOP-1. The anal-yses would not be possible without the contri-butions and efforts of the National Center forAtmospheric Research (NCAR) in elaboratingthe CEOP integrated database. Research wassupported by funding from NASA grantNNG04GQ32G: A Terrestrial EvaporationProduct Using MODIS Data.

References

Adolphs, U., 1999: Representativity analysis and sta-tistical modeling of snow and ice thicknessdata sets from the southern polar PacificOcean. Journal of Geophysical Research-Oceans, 104(C6), 13615–13625.

Andreae, M.O., P. Artaxo, C. Brandao, F.E. Carswell,P. Ciccioli, A.L. da Costa, A.D. Culf, J.L. Es-teves, J.H.C. Gash, J. Grace, P. Kabat, J. Lelie-veld, Y. Malhi, A.O. Manzi, F.X. Meixner, A.D.Nobre, C. Nobre, M.D.L.P. Ruivo, M.A. Silva-Dias, P. Stefani, R. Valentini, J. von Jouanne,and M.J. Waterloo, 2002: Biogeochemical cy-cling of carbon, water, energy, trace gases, andaerosols in Amazonia: The LBA-EUSTACH ex-periments. Journal of Geophysical Research-Atmospheres, 107, No. D20.

Bastiaanssen, W.G.M., M. Menenti, R.A. Feddes, andA.A.M. Holtslag, 1998: A remote sensing sur-face energy balance algorithm for land(SEBAL)—1. Formulation. Journal of Hydrol-ogy, 213(1–4), 198–212.

Beljaars, A.C.M. and F.C. Bosveld, 1997: Cabauwdata for the validation of land surface parame-


terization schemes. Journal of Climate 10(6),1172–1193.

Betts, A.K. and J.H. Ball, 1997: Albedo over the bo-real forest. Journal of Geophysical Research-Atmospheres, 102(D24), 28901–28909.

Beyrich, F. and W. Adam, 2004: A note on the use ofCEOP reference site data for comparison withthe output of global models: The Lindenbergexample. CEOP Newsletter, 6, 6–7.

Braun, P., B. Maurer, G. Muller, P. Gross, G. Heine-mann, and C. Simmer, 2001: An integratedapproach for the determination of regionalevapotranspiration using mesoscale modelling,remote sensing and boundary layer measure-ments. Meteorology and Atmospheric Physics,76(1–2), 83–105.

Bridgham, S.D., J. Pastor, K. Updegraff, T.J. Mal-terer, K. Johnson, C. Harth, and J.Q. Chen,1999: Ecosystem control over temperature andenergy flux in northern peatlands. EcologicalApplications 9(4), 1345–1358.

Brutsaert, W., 1982: Evaporation into the atmo-sphere. D. Reidel, 299 pp.

Brutsaert, W., 1999: Aspects of bulk atmosphericboundary layer similarity under free-convective conditions. Rev. Geophy., 37, 439–451.

Chen, J.M. and T.A. Black, 1992: Defining Leaf-AreaIndex for Non-Flat Leaves. Plant Cell and En-vironment, 15(4), 421–429.

Chen, J.M., 1999: Spatial scaling of a remotelysensed surface parameter by contexture. Re-mote Sensing of Environment, 69(1), 30–42.

Chen, T.H., A. Henderson-Sellers, P.C.D. Milly, A.J.Pitman, A.C.M. Beljaars, et al. 1997: Cabauwexperimental results from the project for inter-comparison of land-surface parameteriza-tion schemes. Journal of Climate 10(6), 1194–1215.

Cohen, W.B., T.K. Maiersperger, Z.Q. Yang, S.T.Gower, D.P. Turner, W.D. Ritts, M. Berter-retche, and S.W. Running, 2003: Comparisonsof land cover and LAI estimates derived fromETM plus and MODIS for four sites in NorthAmerica: a quality assessment of 2000/2001provisional MODIS products. Remote Sensingof Environment, 88(3), 233–255.

Conte, S.D., H.E. Dunsmore, and V.Y. Shen, 1986:Software Engineering Metrics and Models.Menlo Park, Calif.: Benjamin-Cummings,173–174.

Delworth, T. and S. Manabe, 1993: Climate Variabil-ity and Land-Surface Processes. Advances inWater Resources 16(1), 3–20.

Eugster, W., W.R. Rouse, R.A. Pielke, J.P. McFad-den, D.D. Baldocchi, et al., 2000: Land-atmosphere energy exchange in Arctic tundra

and boreal forest: available data and feedbacksto climate. Global Change Biology, 6, 84–115.

Famiglietti, J.S. and E.F. Wood, 1991: Evapotranspi-ration and Runoff from Large Land Areas—Land Surface Hydrology for AtmosphericGeneral-Circulation Models. Surveys in Geo-physics, 12(1–3), 179–204.

Famiglietti, J.S. and E.F. Wood, 1995: Effects of Spa-tial Variability and Scale on Areally AveragedEvapotranspiration. Water Resources Research,31(3), 699–712.

Friedl, M.A., D.K. McIver, J.C.F. Hodges, X.Y.Zhang, D. Muchoney, et al., 2002: Global landcover mapping from MODIS: algorithms andearly results. Remote Sensing of Environment,83(1–2), 287–302.

Henderson-Sellers, A., A.J. Pitman, P.K. Love, P.Irannejad, and T.H. Chen, 1995: The Projectfor Intercomparison of Land-Surface Parame-terization Schemes (Pilps)—Phase-2 andPhase-3. Bulletin of the American Meteorolog-ical Society, 76(4), 489–503.

Hu, Z.L. and S. Islam, 1998: Effects of subgrid-scaleheterogeneity of soil wetness and temperatureon grid-scale evaporation and its parametriza-tion. International Journal of Climatology,18(1), 49–63.

Huete, A., K. Didan, T. Miura, E.P. Rodriguez, X.Gao, and L.G. Ferreira, 2002: Overview of theradiometric and biophysical performance ofthe MODIS vegetation indices. Remote Sensingof Environment, 83(1–2), 195–213.

Jia, L., Z.B. Su, B. van den, Hurk, M. Menenti, A.Moene, et al., 2003: Estimation of sensibleheat flux using the Surface Energy BalanceSystem (SEBS) and ATSR measurements.Physics and Chemistry of the Earth, 28, 75–88.

Koike, T., 2002: CEOP Starts—A step for predict-ability improvement of the water cycle andwater resources. CEOP Newsletter, 1, 1.

Kustas, W.P. and C.S.T. Daughtry, 1990: Estimationof the Soil Heat-Flux Net-Radiation Ratio fromSpectral Data. Agricultural and Forest Meteo-rology, 49(3): 205–223.

Kustas, W.P. and J.M. Norman, 1996: Use of remotesensing for evapotranspiration monitoring overland surfaces. Hydrol. Sci. J., 41, 495–516.

Lafleur, P.M. and W.R. Rouse, 1995: Energy Parti-tioning at Treeline Forest and Tundra Sitesand Its Sensitivity to Climate-Change.Atmosphere-Ocean 33(1), 121–133.

Lloyd, C.R., P. Bessemoulin, F.D. Cropley, A.D. Culf,A.J. Dolman, J. Elbers, B. Heusinkveld, J.B.Moncrieff, B. Monteny, and A. Verhoef, 1997:A comparison of surface fluxes at the HAPEX-Sahel fallow bush sites. Journal of Hydrology,189, 400–425.


Margolis, H.A. and M.G. Ryan, 1997: A physiologicalbasis for biosphere-atmosphere interactions inthe boreal forest: an overview. Tree Physiology,17(8–9): 491–499.

McCabe, M.F., E.F. Wood, and H. Su, 2005: Theinfluence of scale on multi-sensor estimationof evapotranspiration (submitted to RemoteSensing of Environment).

MacKay, D.S., D.E. Ahl, B.E. Ewers, S.T. Gower,S.N. Burrows, S. Samanta, and K.J. Davis,2002: Effects of aggregated classifications offorest composition on estimates of evapotrans-piration in a northern Wisconsin forest. GlobalChange Biology, 8(12), 1253–1265.

Miller, D., J. Washburne, and E. Wood, 1995: EOSWorkshop on land-surface evaporation andtranspiration, The Earth Obs., 7(4), 52–56.

Monteith, J.L., 1981: Evaporation and Surface-Temperature. Quarterly Journal of the RoyalMeteorological Society, 107(451), 1–27.

Myneni, R.B., S. Hoffman, Y. Knyazikhin, J.L. Priv-ette, J. Glassy, et al., 2002: Global products ofvegetation leaf area and fraction absorbedPAR from year one of MODIS data. RemoteSensing of Environment 83(1–2), 214–231.

Penman, H.L., 1948: Natural evaporation from openwater, bare soil and grass. Proc. Roy. Soc., A,193, 120–146.

Plana, V., A. Ruiz, M.C. Ruiz-Sanchez, J.A. Franco,and J.M. Abrisqueta, 2002: Spatial representa-tivity of the possible sites for measuring thewater balance of apricot trees. AgriculturalWater Management, 57(2), 145–153.

Priestley, C.H.B. and R.J. Taylor, 1972: On the as-sessment of surface heat flux and evaporationusing large-scale parameters. Mon. WeatherRev., 100, 81–92.

Rodell, M., P.R. Houser, U. Jambor, et al., 2004: Theglobal land data assimilation system. Bulletin

of the American Meteorological Society, 85(3):381–394.

Schaaf, C.B., F. Gao, A.H. Strahler, W. Lucht, X.W.Li, T. Tsang, et al., 2002: First operationalBRDF, albedo nadir reflectance products fromMODIS. Remote Sensing of Environment,83(1–2), 135–148.

Su, H., M.F. McCabe, E.F. Wood, Z. Su, and J.H.Prueger, 2005: Modeling evapotranspirationduring SMACEX: Comparing two approachesfor local- and regional-scale prediction, J. Hy-drometeorol, 6, 910–922.

Su, Z. and M. Menenti, (Eds.), 1999: Mesoscale cli-mate hydrology: the contribution of the new ob-serving systems. Report USP-2, 99-05, Publica-tions of the National Remote Sensing Board(BCRS). 141 pp.

Su, Z., 2002: The Surface Energy Balance System(SEBS) for estimation of turbulent heat fluxes.Hydrology and Earth System Sciences 6(1), 85–99.

Wan, Z. and Z.-L. Li, 1997: A physics-based algo-rithm for retrieving land-surface emissivityand temperature from EOS/MODIS data, IEEETrans. Geosci. Remote Sens., 35(4), 980–996.

Wan, Z., Y. Zhang, Q. Zhang, and Z.L. Li, 2004: Qual-ity assessment and validation of the MODISglobal land surface temperature. InternationalJournal of Remote Sensing 25(1), 261–274.

Wang, Y.J., C.E. Woodcock, W. Buermann, P. Sten-berg, P. Voipio, H. Smolander, T. Hame, Y.H.Tian, J.N. Hu, Y. Knyazikhin, and R.B. My-neni, 2004: Evaluation of the MODIS LAI algo-rithm at a coniferous forest site in Finland. Re-mote Sensing of Environment, 91(1), 114–127.

Wood, E.F., 1995: Scaling Behavior of HydrologicalFluxes and Variables—Empirical-Studies Us-ing a Hydrological Model and Remote-SensingData. Hydrological Processes, 9(3–4), 331–346.


evaluation of remotely sensed evapotranspiration over the ceop

Documents