New Scheme for Validating Remote-Sensing Land Surface

Preparing to load PDF file. please wait...

0 of 0
New Scheme for Validating Remote-Sensing Land Surface

Transcript Of New Scheme for Validating Remote-Sensing Land Surface

remote sensing
New Scheme for Validating Remote-Sensing Land Surface Temperature Products with Station Observations
Wenping Yu 1,*, Mingguo Ma 1,* ID , Zhaoliang Li 2, Junlei Tan 3 and Adan Wu 3 1 Chongqing Engineering Research Center for Remote Sensing Big Data Application, School of Geographical Sciences, Southwest University, No. 2 Tiansheng Road, Beibei District, Chongqing 400715, China 2 ICube, Uds, CNRS (UMR7357), 300 Bld Sebastien-Brant, CS10413, 67412 Illkirch, France; [email protected] 3 Heihe Remote Sensing Experimental Research Station, Northwest Institute of Eco-Environment and Resources, Chinese Academy of Sciences, 320 Donggang West Road, Lanzhou 730000, China; [email protected] (J.T.); [email protected] (A.W.) * Correspondence: [email protected] (W.Y.); [email protected] (M.M.)
Received: 27 September 2017; Accepted: 20 November 2017; Published: 24 November 2017
Abstract: Continuous land-surface temperature (LST) observations from ground-based stations are an important reference dataset for validating remote-sensing LST products. However, a lack of evaluations of the representativeness of station observations limits the reliability of validation results. In this study, a new practical validation scheme is presented for validating remote-sensing LST products that includes a key step: assessing the spatial representativeness of ground-based LST measurements. Three indicators, namely, the dominant land-cover type (DLCT), relative bias (RB), and average structure scale (ASS), are established to quantify the representative levels of station observations based on the land-cover type (LCT) and LST reference maps with high spatial resolution. We validated MODIS LSTs using station observations from the Heihe River Basin (HRB) in China. The spatial representative evaluation steps show that the representativeness of observations greatly differs among stations and varies with different vegetation growth and other factors. Large differences in the validation results occur when using different representative level observations, which indicates a large potential for large error during the traditional T-based validation scheme. Comparisons show that the new validation scheme greatly improves the reliability of LST product validation through high-level representative observations.
Keywords: spatial representativeness; heterogeneity; validation; land-surface temperature products (LSTs); observations; HiWATER; remote sensing

1. Introduction
Land-surface temperature (LST) is an important parameter related to the surface energy and water balance at local and global scales and has principal significance for applications such as monitoring the climate, hydrological cycle, and vegetation [1]. Satellite remote sensing provides a repetitive synoptic view in short intervals of the global land surface and is a vital tool for monitoring the LST of the Earth. With the development of remote-sensing technology, many LST products have been provided by different groups based on retrieval from different satellite data [2–5]. The first long-term global sensing LST dataset, the NOAA/NASA Pathfinder AVHRR Land dataset (PAL) [2], was released in 1994. The second generation AVHRR Land Pathfinder Π (PALΠ) was a refinement product from the PAL released in 2000 [3]. Sun and Pinker estimated LST products from a Geostationary Operational Environmental Satellite (GEOS) in 2003 [4]. The LSTs from Spinning Enhanced Visible and Infrared Imager (SEVIRI) onboard the Meteosat Second Generation (MSG) was retrieved in 2008 [6]. As a part of NASA Earth Observing System (EOS) project, MODIS LSTs have played an important role in recent studies, especially in regional studies, because of the suitable temporal and spatial resolution,

Remote Sens. 2017, 9, 1210; doi:10.3390/rs9121210

Remote Sens. 2017, 9, 1210

2 of 24

acceptable accuracy, and accessibility of these LSTs. Therefore, MODIS daily LST/LSE products with 1-km spatial resolution are validated in this study.
Remotely sensed LSTs must be appropriately and precisely evaluated to ensure effective application [7]. Mainly two types of methods exist for validating LST products retrieved from thermal-infrared satellite data: temperature-based methods (T-based) and radiance-based methods (R-based) [8–14]. The main advantage of R-based methods is that they work during both the daytime and nighttime because in situ LST observations are not required, and finding validation sites with small spatial variations in land-surface emissivity is relatively easy [14]. However, the atmospheric and water vapor profiles at validation sites from radiosonde balloons that are synchronously launched with the satellite are a necessary dataset, which limits the implementation of this method for long-term and large-region validation. Therefore, T-based methods remain common, and ground-based measurements are still the primary source of datasets to directly validate remotely sensed LSTs. However, we cannot perform a direct comparison with a pixel grid, especially for a coarse-resolution product over heterogeneous areas, because of the spatial heterogeneity and different scales between ground-based observations and remotely sensed LST pixels.
A generally accepted method is a systematic site-to-network method, which deeply develops an in situ sampling strategy and upscaling approaches to acquire the truth at the pixel scale over a heterogeneous surface based on multiscale, multi-platform and multi-source observations [15]. This approach employs both field measurements from nodes of a wireless sensor network (WSN) and high-resolution remote-sensing data from synchronous high-resolution satellites or airborne sensors to establish a site-specific relationship and generate high-resolution LST reference maps over the validation area [15]. These LST reference maps are then treated as benchmarks to obtain multiscale validation datasets by upscaling methods [16–18]. However, only a few high-resolution LST reference maps can be synchronously obtained for a given region because of cost limitations and the revisiting cycles of satellites with high resolution, which is the greatest challenge towards the global validation of LST products, especially in terms of temporal consistency in product validation.
In contrast to more complicated R-based, site-to-network methods with limited LST reference maps, simple T-based methods are directly based on existing global, long-term ground LST measurements and are an important supplement for validation. Simple T-based methods have been widely used to validate remotely sensed LST products at homogeneous stations [8,13,19]. When directly validating LST products with spatial resolutions above hundreds or even thousands of meters by ground-based measurements, the error from the scale mismatch changes with the land-cover type (LCT) and the proportions of mixtures in pixel grids reduce the reliability of the validation and hinder the application of ground-based LST measurements during the validation of remotely sensed LST products. Coll et al. have pointed out that during the day, LST can vary by 10 K or more over a few meters in a heterogeneous surface [9]. Ground-based LST measurements from two types LST observation instruments with different field of view (FOV) were selected to discuss the scale mismatch implications for validation of remote sensing LST products in the study by Yu et al., and the validation results show that there is an extra 26.9% in the error >3 K range caused by the 41.5 FOV difference [20].Therefore, we must assess the spatial representativeness of station observations at a given spatial resolution to reliably validate remotely sensed LSTs. Recently, several methods have been used to assess the spatial representativeness of different land-surface parameters, such as the leaf area index [21], surface solar radiation [22], bidirectional reflectance distribution function (BRDF)/albedo [23], air temperature [24] and air quality [25], which are observed by ground stations. These methods are based on two factors: the point-to-area consistency and the spatial heterogeneity [21]. The point-to-area consistency indicator can be calculated through two methods. The first involves directly comparing the footprint of the ground-station observations to the corresponding product pixels [26] or the average value of the corresponding area [27]. In the second approach, the observational representativeness is determined by the average difference between a given station and its neighboring stations based on multi-temporal observations from multiple stations [9,28]. The semi-variance is usually selected to

Remote Sens. 2017, 9, 1210

3 of 24

describe the spatial representativeness by analyzing the spatial heterogeneity around the stations [29].
A first-order statistical algorithm is an important spatial heterogeneity indicator, for example, using
window-size analysis to assess the spatial variation of the landscape around a given station [30].
Spatial representativeness assessments have been widely implemented to validate satellite-albedo,
evapotranRsempoitreaStenios. n20,17a,n9,d12L10AI products [21,23,26,31,32]. However, few representativeness3aosf 2s4essments
exist for station LST observations, which increases the uncertainty of the validation of LST products, particularhlyetefroorgsenimeitpylearTo-ubnadsethde ismtaptiolenms [e2n9]t.aAtiofnirss,t-aorndderhsitnadtisetricsatlhaelgaopritphlmicaistioann iomfpsotrattainotnspoabtisaelrvations.
heterogeneity indicator, for example, using window-size analysis to assess the spatial variation of the
This lpanadpsecrappereasroeunntsd aa ngeivwenmsteatthioond[o3l0o].gSypfaotiralvraelpidreasetinntagtiLveSnTesps raosdseusscmtsetnhtsathafovecubseeesn ownidqeulyantifying the spatiaiml prelepmreensteendtatotivaelnideastes soafteslltiatet-iaolbnedoob, seevravpaottriaonnsspirtaotioimn, parnodvLeAtIhperoadcuccutsra[2c1y,23a,n26d,31r,e3l2i]a. bility of LST prodHuocwt evvaerl,idfeawtiroepnr.eseTnthaetivteenremss a“sssepssamtieanltsreexpisrtefsoer nsttaatitoinveLnSTesosb”serrveafteiorsns,two hmicheainscureraesmesethnets of the degree touwncheritcahingtyroofutnhedv-balaidsaetdionobofsLeSrTvaptrioodnusctcs,apnarrteicsuolalvrlye ftohresismuprlreoTu-bnadseidngimLpSleTmebnytaetixontes,nadnidng to the
hinders the application of station observations.
satellite footpTrhiinstp.apTerhpisresveanltisdaanteiwonmetethcohdnoilqoguyefoarsvsaelisdsaetsingthLeSTsppraotdiuacltscthhaatrafocctuesreissotincqsuoanf titfhyeingLST, and the seasotnhaelspreatpiarlerseepnretsaetnivtaetinveensessschofasntagteiosn wobistehrivnatiaonssttaotiismtpicraolvefrthaemaeccwuroarcky.anAd rselcihabeimlityeotfhLaStTis based on spatiapl rroedpurcet sveanlidtaatiiovne.nTehsesteirnmdi“csaptaotirasl riesprperseensteantitveednesisn” trhefiesrsptaopmeera,saunredmetnhtesnoftthhe dgergardeeintog criteria are outlinwehdicihngdroeutnadil-.baAseldl othbseersvtaattioionns scafnrormesoltvheethHeesiuhreroWunadtinegrsLhSeTdbAy lelxietedndTinegletmo ethtreysaEtexllpiteerimental Research f(oHotipWrinAt.TTEhRis)v[a3li3d]ataiornetescehlneicqtueedasfsoerssaesptphleysipnagtiatlhcehavraacltiedriasttiicosnofstthreatLeSgTy, .anTd hthee ssetuasdoynalarea and
representativeness changes within a statistical framework. A scheme that is based on spatial
data-procreespsriensegntpatriovceneedsus irnedaicraetoirns tisropdreusecnetdediinn tSheisctpiaopner3, a. nIdnthSeenctthioe ngra4d,inthgecrriteepriraeasreenotuattliinveednienss of the given statdioetnaiol. bAsellrvthaetiostnastioisnsasfsroemssetdhetoHveaihleidaWteateMrsOheDd ISAlVlie5ddTaeilyemLeStrTy pErxopderuimctesn(taMl OReDse/aMrchYD11A1). The repres(HeniWtaAtiTvEeRn)e[s3s3]aasrseesseslemctedntfoarnadppLlySinTgpthreodvaulicdtavtiaonlidstarattieogny.aTrheeasltsuodyanaraelayaznedddatnad-prdoicsecsusisnsged in this Section. Fpinroacleldy,utrhe earceoinnctrloudsuiocends ianreSescutimonm3a. rIinzeSdec.tion 4, the representativeness of the given station
observations is assessed to validate MODIS V5 daily LST products (MOD/MYD11A1). The
2. Methodreoplroegseyntativeness assessment and LST product validation are also analyzed and discussed in this
Section. Finally, the conclusions are summarized.
2.1. New V2a. MlideathtioodnoSlocghyeme
The L2.S1T. NieswaVlaalniddat-isoun rSfcahceme eparameter with great spatial and temporal heterogeneity, which creates many challenges for “point-to-pixel” comparisons. Local changes in the surface temperature within and between TdhifefeLrSeTntiseacolasnyds-tseumrfascienptraoradmuecteerscwailthe mgreisamt saptacthialerarnodrste.mMpoorraelohveeter,rotgheenseeityp,awttheircnhs change
creates many challenges for “point-to-pixel” comparisons. Local changes in the surface temperature
seasonallywaitnhdinaarnedpbaerttwiceuenladrlifyfedreifnfticeucolstytsoteimdesnintitfryodduucerisncgalepmeriisomdastcohferrarporisd.lMyocrheaonvegri,nthgessue rpfaatcteerncsonditions. Thereforec,h“apnogeinset”asmonealalysuanredmarenptasrtaicluolnarelyadrieffincuoltt stouifdfeicniteifnytdtuorinvgapliedriaotdessoaftrealpliidtely-dchearnivgiendg sLuSrfTacreetrievals, especiallycornedmitiootnes-.sTehnesreinfogreL, “SpToinptr”omdeuacsutsrewmeitnhtsmaloondeearraetneoat nsudffilcoiewnt rtoesvoalliudtaitoensa(teilllliutes-tdrearitveedd iLnSTFigure 1). This temproetrraielvmalsis, mespaetccihallcyanrembeotseo-slevnesidngbyLSiTncprreoadsuicntsgwthithe ombosdeerrvataetiaonnd-laocwqureissoitluiotinonfr(eilqluusternatceyd oinf stations.
Figure 1). This temporal mismatch can be solved by increasing the observation-acquisition frequency
The schemofesttahtiaotnsis. Tdheevscehloempeedthahteirsedetvoevloapleiddhaetererteomvaolitdea-tseernemsiontge-sLenSsTinpgrLoSdTupcrotsdu(sctese(sFeeigFuigruere1)1)attempts to solve sapttaetmiapltsmtoissmolvaetcshpaetifaflemctissmdautcrhinegffevctasldiduaritnigonv.aliTdahteionk.eTyheinketyhiins tshcihs escmheemiesistotoaasssseesssstthhee spatial representasptiavteianl eresps,rewsehnitcahtiviesnbesass,ewdhoicnh risembaosetde-osnenresminogted-saetnasiwngitdhathaiwghithsphiagthiaslpraetisaol lruestiooluntitohnatthatre closely related toatrhe ecloLsSelTy.reInladteidcatotothres LaSrTe. pInrdoicpaotosresdarteoprqoupaosnetdiftyo qthuaenstipfyatthiaelspreatpiarlerseepnretsaetnivtaetinveensse,ssa, nandd then the grading crthiteenritaheagrreaddiensgigcrniteedriabaerfeodreessigenleedctbinefgoraepspelreoctpinrgiaatpepgroropuriantde -gbroausendd-bmaseeadsumreeamsuerenmtsenfotsrfovralidation.

Figure 1. FNigeuwres1c.hNeemwescfhoermleanfodr lsaundrfsaucrefatceemtepmepreartauturree (LSSTT))vvalaidliadtiaotniobnasebdasoendthoenastshesesmasesnetsosfmloecanlt of local spatial repspreatsiealnrteaptrievseennteastisve(snietses l(esivteelle)v. eIln). Itnhethsecshcheemmee,, LLCCTTisisthtehaebabrbebvriaetvioinatoifolnanodf-cloavnedr -tycpoev,earndtype, and NDVI is thNeDaVbI bisrtehveiaabtbiorenvioaftionnoormf naolrimzealdizdedifdfeifrfeernecnecevveeggeettaattioionninidnedxe. x.

Remote Sens. 2017, 9, 1210

4 of 24

2.2. Indicators for Assessing Spatial Representativeness
Three indicators are proposed to describe the spatial characteristics for a specific parameter, mainly including the consistency from points to pixels and the spatial heterogeneity within pixels. These indicators are calculated on high-resolution images, which are much easier to access, thus simplifying the process. The LST is a direct parameter for assessing the representativeness of a given station’s LST observations. However, obtaining temporal high-spatial-resolution LST matches for LST products with lower resolution is difficult. Therefore, high-resolution land-cover type (LCT) and the normalized difference vegetation index (NDVI), which can indicate the surface conditions and their changes, are chosen as additional supporting parameters to evaluate ground-based LST observations in addition to LST data with a high spatial resolution.
If the LCTs observed by stations do not match the dominant LCTs in the pixels, these station LST observations cannot represent the LST of the LCTs in the pixels. Thus, the dominant LCT (DLCT), which is given by the percentage of the observed LCT throughout the pixel’s area, is defined as

DLCT = M(s) × 100 (1) N(s)
where s is the product pixel, M(s) is the area with the LCT that is observed by the given station, and N(s) is the total area of the LCTs in the LST product’s pixel grid. When using a high-resolution LCT map, the DLCT can also be described as the percentage of fine pixel numbers covered by station-observed LTC to total LCT pixel numbers in the LST product’s pixel range. A high DLCT indicates that the LCT observed by a station is consistent with that in the LST product’s pixel and low heterogeneity within the product pixel because the mixing rate of LCTs in the pixel is not large.
We developed a relative bias (RB) indicator to assess how close a ground-based LST measurement is to the value of the corresponding pixel area. According to the high-spatial-resolution LST reference images, the relative bias is used to describe the difference between the LST value T(s) at a station and the average LST value T(s) in the product pixel’s area. If we consider the comparability between different ranges of LST values, the RB is defined as

RB = |T(s) − T(s)| × 100 (2) T(s)
where s is the product pixel and RB depends on both its resolution and the resolution of the reference LST image. This indicator can quantify the certainty of the ground-based measurements to the product pixel area LST values. A smaller RB indicates more spatially representative observations for the corresponding pixel at the specific spatial resolution of s.
The two above indicators mainly measure the point-to-area value consistency, so the heterogeneity of the spatial distribution of the LST within a pixel, which is correlated with the vegetation growth, must be seasonally quantified. In terms of the spatial autocorrelation of LST parameters, semivariogram is one of the most commonly used and efficient geostatistical analysis tools for quantitatively evaluating spatial variations. Semi-variance and related geostatistical kriging were developed from mining research during the late 1950s and have been widely used after a publication by Journel and Huijbregts in 1978 [34,35]. These geostatistical techniques have been used in many scientific projects, such as in describing the distribution and density of plants and animals [36,37] and in determining the spatial scales of variation and sampling strategies in remote sensing [38–40]. In regionalized variable theory, the semi-variance measures the dissimilarity of a spatial variable observed at different locations. The semi-variance is calculated by the average squared difference between observations Z(xi) and Z(xj), which are separated by distance h, as described below:

r(h) = 1

∑ (Z(xi) − Z(xj))2


2N(h) ||xi−xj||=h

Remote Sens. 2017, 9, 1210

5 of 24

where 2N(h) is the number of observation pairs, which are separated by a distance h, or lag, as intervals to calculate the semi-variance. In this study, the variogram estimator r(h) is computed on discretized point values from high-spatial resolution LST pixels, and then a variogram model is established as a parametric functional approximation based on these semi-variance values. Several theoretical variogram model types exist, including linear models, spherical models, exponential models, and Gaussian models. Among these models, spherical models are the most widely used variogram models for their strong fitting and generalization capabilities and are recommended for assessing the spatial representativeness of observations [23,32]. The isotropic spherical variogram that is used to estimate the variogram is as follows:

rsph(h) =  c0 + c × 1.5 × ha − 0.5 × ha 3 f or 0 ≤ h ≤ a (4)

c0 + c f or h > a

In Equation (4), it is obvious that the rsph(h) generally increases from a nonzero value to a relatively stable constant value with h. When h = 0, the rsph(h) value is a nonzero value c0, namely, rsph(0) = c0, and c0 is the nugget of the variogram. The stable constant value is the sill (c0 + c) of the variogram. When rsph(h) reaches the sill, the value of the variable h is a, namely, the range of the variogram. The key parameters range (a), nugget (c0), and sill (c0 + c) can be obtained from fitting Equation (4). a is the maximal distance between the two correlated points and indicates the average structural scale (ASS) of the given area. The nugget c0 indicates the level of Z(xi)’s randomness, which may be caused by internal variations in Z(xi) over a smaller distance h than the sampling distance or may be derived from the sampling error. The sill (c0 + c) represents the largest extent of the regionalized variation. Li and Reynolds [41] introduced the proportion of structural variation, which is based on subtracting the variogram nugget (c0) from the sill (c) and then dividing by the sill, to discuss the definition and quantification of ecological heterogeneity. However, the heterogeneity
of LST parameters considerably varies over time when compared to that of ecological parameters,
and reference high-resolution LST maps may not be completely synchronous with the validated LST
products. Therefore, the ASS is introduced based on the range of the variogram model, which reflects
the average structural scale of the given area and the size of the homogeneous area. A large ASS value
indicates that the station observations represent a large homogeneous area.

3. Data Instruction and Preparation

3.1. Ground-Based LST Measurements
In this study, we selected the Heihe River Basin (HRB), which is the second largest inland valley in China’s arid regions, to evaluate MODIS LSTs with 1-km resolution. The HRB is located in the northern arid region within 97.1◦E–102.0◦E and 37.7◦N–42.7◦N. Glaciers, frozen soils, alpine meadows, forests, irrigated crops, riparian ecosystems, deserts, and gobi are distributed from upstream to downstream regions (see Figure 2). The HRB was selected as an experimental watershed to reveal the processes and mechanisms of the ecohydrological system in an inland river basin. Allied telemetry experiments such as the Heihe Basin Field Experiment (HEIFE) [42] and Watershed Allied Telemetry Experimental Research (WATER) [43] have been conducted in the HRB, and the HiWATER [33] project has been ongoing since 2012. The stations that collect watershed hydrological observations cover a wider range than those in previous studies and provide a large amount of ecohydrological data for evaluation. Eighteen atmospheric stations from HiWATER are scattered around the HRB region. Longwave-radiation data for eighteen stations from 2013 to 2014 were selected to obtain ground-based LSTs to evaluate the MODIS LSTs. The locations of the stations are shown in Figure 2. The information for the stations is listed in Table 1, and environmental photos of these sites are shown in Figure 3. The ARC, ARS, ARY, DSL, JYL, HZS, HCG, and EBZ stations are located in the upstream area; the DMZ, GBZ, HZZ, SDZ, and SSW stations are located in the midstream area; and the downstream area

Remote Sens. 2017, 9, 1210

6 of 24

contains the SDQ, HJL, HYZ, NTZ, and LTZ stations. The LCTs of these stations are all typical types in

theRtemhorteeeSesnisg. n20i1fi7c, a9,n1t2l1y0 different areas.

6 of 24

FiFgiugruere2.2.SStutuddyyaarreeaaaanndd lloocations off tthheevvaalliiddaatitoionnstsatatitoionns.s.

TTabablele11. .InInffoorrmmaattiioonn ffor the staattiioonnssiinntthhisissstutuddyy. .

SSttaattiioonn NNaammee A’rou superstation (ARC) AA’’rroouu ssuupnnerystsaltoipone s(AtaRtiCon) (AA’rRoSu)sunny slope Ast’artoiounsh(AadReS)station (ARY)
DA’arsohuaslohnagdestsatatitoionn(D(ASRLY) )

LLoonnggiittuuddee//◦°EE 100.46 100.46
100.52 100.52 100.42

LLataittiutudde/e◦/N°N 38.05
38.09 38.09

AAltlittiutudde/em/m 3033 3033
3559 3559 3538

HHeeiigghhtt//mm 5 5
6 6 6

FFoooottpprriinntt11/m/m 37.32 37.32
44.78 44.78 44.78

LLaannddssccaappeess Alpine meadow Alpine meadow
Alpine grassland Alpine grassland Alpine grassland





6 6





6 6

Huangcaogou station (HCG) 100.73




EH’ubaonsgtaztainogns(iEsBtaZti)on (HZS)




6 6

DHaumanagncasuogpoeurstsatatitoionn(D(HMCZG)) 110000..7337




4444.7.788 4444.7.788 44.78 4444.7.788 8494.5.778

SAwlpaimnep mmeeaaddooww AClrpoipnleanmdea(wdohweat) Alpine grassland CArloppinlaengdra(wsshlaenadt) AClrpoipnleangdra(smslaanizde)

GE’obboissttaattiioonn((EGBBZZ) )






AGlopbini De gesraesrtsland

HDaumazahnasizuipdeersstearttiostnat(iDoMn Z)











6 6


SHhueanzshhaaiwzioddeseesretrsttsattaiotinon(HZZ) 100.32





(WSSetWla)nd station (SDZ)









44.78 44.78

GWoebtilaDnedsert Desert steppe Desert Wetland

PShoepnuslhuaswfooredsetssetrattion (HYZ) 101.12



Cstraotipolnan(SdSsWta)tion (NTZ)




BPaorpruelnu-slafnodressttasttiaotnio(nLT(HZY) Z) 110011..1123







Barren-land station (LTZ)






6 6








Populus forest DCeroseprltand
PBoapreulsuosilforest Euphrates poplar ColriovpelaannddTamarix Bmairxeedsofiol rest







EEuphrates poppllaarr oolive and TTaammaarriixx mmiixxeedd ffoorreesstt

1 “Footprint” refers to the diameter of the footprint; “Height” indicates the installationEhuepihgrahtte.s poplar

Mixed forest station (HJL)






olive and Tamarix

mixed forest
Eighte1e“nFomoteptreinotr”orleofgeriscatol tthoewdeiarms eatreer olof tchaetefodotipnritnht;e“HHeRigBhtr”eignidoinca,tceos nthseisintisntagllaotfiotnhhreeeighstu. perstations and fifteen ordinary stations. Pyrgeometers are deployed at these 10-m to 35-m high meteorological

staEtiiognhtteoewnemrsetotemoreoalsougriecalol ntogwwearvsearraedlioatciaotned(siene tFhigeuHreR3BarnedgiToanb,lceo1n)s. iAsttinlegasotftwthorepeysrugpeoemrsetatetirosns andarfie fpteoesnitioorndeidnaornyastsaintigolnest.oPwyerrg: eoonme efatecirnsgaruepdweaprldoyanedd athtethoetsheer10fa-mcintgod3o5w-mnwhaigrdh.mTheteefoireoldlo-ogfi-cal
view (FOV) of the upward-facing pyrgeometer is nearly 180°, while that of the downward-facing staptiyorngetoomweetresrtios m15e0a°s. uTrheelroefnogrwe,athvee eraffdeciatitvioend(iasemeeFteigr uorfeth3eaFnOdVTaobf ltehe1)p. yArgt eleoamstettewrsoopnyarg1e0o-mmettoers are3p5-omsittioowneedr iosnapapsrionxgilme atotewlyer2:.5o–n2e4fmacwinigthuapw6-mardavaenrdagtehemootuhnetrinfagcihnegigdhot,wannwd athrde .dTiahme efiteelrds-ooff-tvhieew

ground-observation footprints are shown in Table 1. The pyrgeometers are sensitive to the spectral

range from 4.5 to 42 µm in the longwave band. All the instruments at each station were calibrated

Remote Sens. 2017, 9, 1210

7 of 24

(FOV) of the upward-facing pyrgeometer is nearly 180◦, while that of the downward-facing pyrgeometer is 150◦. Therefore, the effective diameter of the FOV of the pyrgeometers on a 10-m to

35-m tower is approximately 2.5–24 m with a 6-m average mounting height, and the diameters of the

ground-observation footprints are shown in Table 1. The pyrgeometers are sensitive to the spectral

range from 4.5 to 42 µm in the longwave band. All the instruments at each station were calibrated

before and after field deployment. Field-routing exams were implemented once per month. Assurance

and quality control provided the best possible data for the level-2 daily data. All these data and related

information can be found at the HiWATER website [44]. All the ground-based measured data from the eighteen RHemioWte SAenTs.E20R17,s9i,t1e2s10were 10-min averaged values. The longwave radiation data w7 oef r2e4 selected

accordingbetfoorteheanfideladftevriefwielidngdetipmloeymfoenr tt.hFeieMldO-roDutIiSngLSeTxa. mTshewLerSeTimisprleelmaetendtedtoolnacnedp-serurmfaocnethl.ongwave radiationAascscuorarndcienagntdoqtuhaelitSytecofanntr-oBl oplrtozvmidaednnthleabwes[t1p,4os5s]i:ble data for the level-2 daily data. All these

data and related information can be found at the HiWATER website [44]. All the ground-based

measured data from the eighteenL↑H=iWεAbTδETR4 s+ite(s1w−erεeb)10×-mLin↓ averaged values. The longwave


radiation data were selected according to the field viewing time for the MODIS LST. The LST is

where L↑reilsatethdetoslaunrdfa-scuerfaucpe wlonegllwinavge lroadnigatwioanvaeccorraddiniagttioonth,e εSbtefiasnt-Bhoeltzsmurafnancleawbr[1o,4a5d]:band emissivity, δ







W·m+−(12 ·−K−)4 )×,

and ↓





do(5w) nwelling

longwave radiation at the surface. Therefore, the ground-measured LST can be estimated from station

longwavew-hraerdeiat↑ioisnthoebssuerrfavcaetuiopnwselbliyngthloenfgowlalovwe riandgiateiqonu,atioins :the surface broadband emissivity, is

Stefan-Boltzmann’s constant (5.67 × 10−8 W∙m−2∙K−4), and L↓ is the atmospheric downwelling

longwave radiation at the surface. Therefore, the ground-measured1 LST can be estimated from station

longwave-radiation observationsTbsy =the foLl↑lo−win(g1 e−quεabt)io×n: L↓ 4


= ↑ − (1ε−bδ ) × ↓


In Equation (6), L↑ and L↓ are obtained from the ground-based measurements. Seven narrowband

emissivities eInxisEtquinatiMonO(D6)/, M↑YaDn1d1B↓1 LarSeTo/bLtaSinEedprforodmuctthse. gεbrocuannd-bbaeseedstmimeaastuerdemfernotms. Stheveesne MODIS narrowbannadrroewmbiasnsdivemitiiesssiv[i4ti5e]s:exist in MOD/MYD11B1 LST/LSE products. can be estimated from these

MODIS narrowband emissivities [45]:

εb = 0=.2102.2212×2 ×ε29 ++00..33885599 ×× ε31++0.400.42092×9 × ε32





is the

broadband emissivity, and ε29,
is the broadband emissivity, and

εε3219,, aεn3d1,


narrow emissivity products of
the narrow emissivity products of


bands 29, 31, and 32 that are retrieved from the MODIS day/night LST algorithm (MOD/MYD11B1

MODIS bands 29, 31, and 32 that are retrieved from the MODIS day/night LST algorithm


Figure 3. Cont.

Remote Sens. 2017, 9, 1210
Remote Sens. 2017, 9, 1210

8 of 24 8 of 24

Figure 3.FiPghuroeto3.sPohfottohseof18theH1e8ihHeeiW heaWteartserhsehdedAAlllliieeddTTeelelmemetreytrEyxpEexripmeernitmaleRnetsaeal rRches(HeaiWrcAhT(EHR)iWATER) stations. Tsthateiotnws.oThpeytrwgoeopmyrgeetoemrsettehrastthwatewreeruesuesdedttoo rreeccoorrddthtehleonlognwgawvearvadeiartaiodniawtieorendwepeloreyeddeapt laonyed at an
average height of 6 m, with one facing upwards and the other facing downwards.
average height of 6 m, with one facing upwards and the other facing downwards.
3.2. Remote-Sensing Data
3.2. Remote-Sensing Data
3.2.1. MODIS Data
3.2.1. MODISADs aatacomponent of NASA’s Earth Observing System (EOS) project, two MODIS instruments
were placed onboard the Terra and Aqua satellite platforms to provide information for global
As aactommosppohenreen-,t olafnNd-ASaAnd’s EoaceratnhicO-pbrsoecersvsingstuSdyisetsem[46(]E. OMS)OpDr/oMjeYcDt,2t2w_Lo2,MMOODDIS/MinYsDt1ru1Am1e, nts were placed onMboOaDr/dMtYhDe1T1Be1rraanadnMdOADq/MuaYDsa0t7e_lLl2itearpe laaltlfdoarimlysLtSoTpprroodvuidctes ibnafsoedrmonattihoenrmfoalr-ignflroabreadl adtamtaosphere-, land- andfroomceMaOnDicI-Sp. rMoOceDs/sMYstDu1d1_ieLs2 w[4a6s]r.etrMievOedDb/yMa YgeDne2r2a_lizLe2d, sMpliOt-wDin/dMowYDalg1o1rAith1m, wMitOh D1-/kmMYD11B1 and MOD/MYD07_L2 are all daily LST products based on thermal-infrared data from MODIS. MOD/MYD11_L2 was retrieved by a generalized split-window algorithm with 1-km spatial resolution [47]. MOD/MYD11A1 is tile-based and gridded in the sinusoidal projection from MOD/MYD11_L2. MOD/MYD11B1 was obtained using the day/night LST algorithm at 5-km spatial resolution [48]. MOD/MYD07_L2 was retrieved by the atmospheric team using statistical regression methods [49]. In this study, we focus on the collection of 5 MOD/MYD11A1 products, which are more widespread. Uncertainties from the satellite measurements and improvements in the original

Remote Sens. 2017, 9, 1210

9 of 24

MODIS LSTs for cloudy days are beyond the scope of this paper. To eliminate effects from the inversion algorithm and clouds, only pixels with high-quality MODIS LSTs were selected for the evaluation based on a quality control flag value of 0. The narrow emissivity products from MOD/MYD11B1 LST/LSE were selected to estimate the land-surface broadband emissivity εb and obtain ground-based LSTs [45]. The narrow emissivities from MOD/MYD11B1 LST/LSE at 5-km resolution were resampled to 1-km spatial resolution to match the evaluated MODIS LSTs.

3.2.2. High-Spatial-Resolution Images

Land-cover images with a spatial resolution of 30 m (data doi:10.3972/hiwater.155.2014.db, downloaded from the HiWATER land cover datasets [50], which were produced by Zhong et al. [51,52], were selected to obtain the DLCT in a 1-km LST product pixel. This dataset was mainly based on charge-coupled device (CCD) data from the Huan Jing 1 (HJ-1) satellite, which was launched on 6 September 2008, by the China Center for Resources Satellite Data and Application (CRESDA). HJ-1/CCD has three visible bands and one near-infrared band [53]. This dataset provides monthly land-cover maps of the HRB from 2011 to 2015 with 30-m spatial resolution. LCTs change with the seasons, and these changes are similar across consecutive years, so the DLCT products in 2013 were collected to calculate the DLCT indicator.
Landsat 8, which is called the Landsat Data Continuity Mission, is extending the distinguished 40-year records of Landsat-series satellites and has enhanced capabilities, such as adding new spectral bands in the visible and thermal-infrared wavelengths and improving the signal-to-noise ratio and radiometric resolution of the sensor [54]. The Landsat 8 satellite includes two instruments: an Operational Land Imager (OLI) and Thermal Infrared Sensor (TIRS). High-resolution LST maps were retrieved from the Landsat 8 TIRS data and OLI data. Before retrieving the LST maps, the Landsat OLI and TIRS images were preprocessed, including radiometric calibration and atmospheric correction based on the correction model in the ENVI software. Monthly NDVI maps were obtained to assess the relationships between the monthly changes in the indicators and vegetation growth. The NDVI maps were based on the visible red band (R, band 4) and the near-infrared band (NIR, band 5) according to the following equation:
NDVI = N IR − R (8) NIR + R
In this study, the LST data were estimated from the TIRS aboard Landsat 8 based on a practical split-window (SW) algorithm developed by Du et al. [55]. The SW algorithm can be expressed as follows:

1 − ε ∆ε Ti + Tj

1 − ε ∆ε Ti − Tj


T = b0 + b1 + b2 ε + b3 ε2

2 + b4 + b5 ε + b6 ε2

2 + b7 Ti − Tj


where T is LST, Ti and Tj are the TOA brightness temperatures in the thermal-infrared channels i and j, respectively; ε is the average emissivity of the two channels (i.e., ε = 0.5(+εj)); ∆ε is the channel emissivity difference (i.e., ε = 0.5(εi − εj)); and bk(k = 0, 1, . . . 7) are the algorithm coefficients from the simulated dataset. In this algorithm, the coefficients were determined based on atmospheric water-vapor subranges, which were obtained through a modified split-window covariance-variance ratio method. The channel emissivities were acquired from newly released global land-cover products at 30 m and a fraction of the calculated vegetation cover from visible and near-infrared images that were obtained by Landsat 8.
The effect of heterogeneity changes depending on the season, so we selected Landsat 8 data from September 2013 to August 2014 with a 16-day temporal resolution to calculate the RB, and ASS. A total of 92 Landsat 8 images for all the stations in the HRB were downloaded from the following USGS website [56]. The statistical results were based on per-month averages to eliminate invalid data from cloud cover.

Remote Sens. 2017, 9, 1210

10 of 24

4. Results and Discussion
4.1. Spatial Representativeness Classification
Since the DLCT and RB can measure point-to-pixel value consistency, and ASS can assess the spatial patterns for a given station, respectively. Therefore, all three indicators were combined to describe the representativeness including point-to-pixel consistency and spatial heterogeneity. The spatial representativeness was also classified based on these three indicators. The DLCT indicator determines the representativeness of the station’s LCT in the product pixel. When the LCT in the view footprint of a station is not dominant within the product pixel, the station observations cannot be representative of the pixel, and the LST value of all other vegetation-cover types may be ignored, even if the point-to-area LST consistency is high at the station sometime. When a pixel has a large DLCT value, the RB and ASS subsequently would determine the spatial representativeness together. Presumably, the station-observed LSTs represent ideal data for LST product validation if the RB value is small and the ASS value is large. If the RB and are ASS value are both small, the station may have some spatial representation. In some extreme cases, the station observing area is an average heterogeneity sub-areas in the products pixel, which means the station observations are representative for pixels, although the surface is heterogeneous in these products pixels. Finally, if the RB is large but the ASS is small, the observations probably differ from the values of the pixel.
Reasonable thresholds for the DLCT, RB and ASS are required to determine the representativeness level of a given station’s observations. The emissivity products of the version 5 collection of MODIS LST/LSE products are retrieved based on the LCT of a pixel, and the LCT is classified as the land cover of this pixel based on the classification rule for MODIS land-cover products (MCD12Q1) if the area percent of one LCT in a pixel is higher than 60% [57]. Thus, 60% was defined as the threshold of the DLCT in this study. The RB was used to evaluate the difference between the land-based measurements within the view footprint at each station (in Table 1, the view footprints are shown in the sixth column) and the mean pixel value at the station locations. The ideal RB value is close to zero. However, the threshold of the RB is not unique and depends on the spatial resolution of the LST products, the view footprint of the station measurements, and the retrieval accuracy of the high-spatial-resolution LST maps that are used to evaluate the representativeness. Thus, a reasonable threshold for the RB was 0.5% in this study for MODIS LST products with 1-km spatial resolution, a station view footprint above 30 m and a 1-K retrieval error from the high-resolution LST map itself [55]. The ASS is calculated from variogram models based on the semi-variance in a 3-km × 3-km area that is centered at a given station and can indicate the greatest distance over which the value at a point on the surface is related to the value at another point. The ASS defines the maximum neighborhood over which control points should be selected to estimate a grid node to take advantage of the statistical correlation among observations and can describe the spatial distribution of LSTs and quantify the average LST spatial structures in the given area. In this study, the measurements from stations were used to evaluate the MODIS LSTs with a 1-km spatial resolution. Therefore, a reasonable ASS indicator should be larger than the spatial resolution of the LST products, that is, 1 km in this study.
The spatial representativeness of the station’s LST observations was classified into five different levels based on the difference-constraining degrees of the three indicators and their thresholds. The levels and their descriptions are presented in Table 2.
RepresentativenessResolutionLst ProductsStationsStation