frofr

finally selected for land surface temperature retrieval. The single-channel method has been applied to AHS band 75 (10.07m), which shows thehighest atmospheric transmissivity, whereas the split-window method has been applied to the combination between bands 75 and 79 (12.35m),

the collection, analysis, and modeling of environmental data,less attention has been given to the use of thermal infrared (TIR)

processes and for their characterization, analysis, and modelling(Quattrochi & Luvall, 2004). There are two fundamentalreasons why TIR data contribute to an improved understandingof land surface processes: (i) through measurement of surface

Remote Sensing of Environment 1remote sensing. With the launch of the NASA Terra suite ofwhich provides the best results according to simulated data. All the AHS thermal bands ranging from 75 to 79 have been used in the TES method.The tests conducted on the different algorithms used in this study show that single-channel and split-window methods provided similar results,with root mean square errors (RMSE) between 1.6 and 1.9K. The TES method slightly improved the results, with a RMSE of 1.4K. 2006 Elsevier Inc. All rights reserved.

Keywords: Airborne hyperspectral scanner; Visible and near infrared; Spectral band

1. Introduction

Although remote sensing is recognized as a powerful tool in

Earth remote sensing instruments in 1999, which included TIRsensors, thermal data are poised to become a major source ofquantitative and qualitative information on land surfaceGlobal Change Unit, Department of Thermodynamics, Faculty of Physics, University of Valencia, Valencia, Spainb Instituto de Agricultura Sostenible (IAS), Consejo Superior de Investigaciones Cientficas (CSIC), Crdoba, Spain

c Instituto Nacional de Tcnica Aeroespacial (INTA), Dpto. de Observacin de la Tierra, Teledeteccin y Atmsfera, Madrid, Spain

Received 24 October 2005; received in revised form 30 January 2006; accepted 1 February 2006

Abstract

The AHS (airborne hyperspectral scanner) instrument has 80 spectral bands covering the visible and near infrared (VNIR), short wave infrared(SWIR), mid-infrared (MIR), and thermal infrared (TIR) spectral range. The instrument is operated by Instituto Nacional de Tcnica Aerospacial(INTA) and it has been involved in several field campaigns since 2004. The main goal of this paper is to analyze the feasibility of retrieving landsurface temperature from the 10 AHS thermal infrared bands, from 71 to 80, located in the region between 8 and 13m. For this purpose, threedifferent methods have been considered: (i) the single-channel method, which uses only one thermal band; (ii) the split-window method, whichuses a combination of two thermal bands; and (iii) the TES (temperature and emissivity separation) method, which needs at least four thermalbands. The calibration of the AHS thermal bands and the algorithms have been tested with in situ measurements collected in the framework of theSPARC (Spectra Barrax Campaign) and EAGLE (Exploitation of AnGular effects in Land surfacE observations from satellites) field campaigns,which took place simultaneously in the agricultural area of Barrax (Albacete, Spain), and also in the framework of the AGRISPECTRA fieldcampaign, carried out over an olive orchard in Crdoba (Spain), in July 2004. AHS flights were conducted at two different altitudes, 975m and2745m above ground level. The results show that AHS bands 71 (8.18m), 72 (8.66m), and 73 (9.15m) were affected by a calibrationproblem. Taking into account that AHS bands 74 (9.60m) and 80 (12.93m) are located in an absorption region, bands from 75 to 79 have beenaJos A. Sobrino a,, Juan C. Jimnez-Muoz a, Pablo J. Zarco-Tejada b,Guadalupe Sepulcre-Cant b, Eduardo de Miguel cLand surface temperature derivedthermal in Corresponding author. Dr Moliner 50-46100 Burjassot, Valencia, Spain.Tel./fax: +34 96 354 31 15.

E-mail address: sobrino@uv.es (J.A. Sobrino).

0034-4257/$ - see front matter 2006 Elsevier Inc. All rights reserved.doi:10.1016/j.rse.2006.02.001m airborne hyperspectral scannerared data

02 (2006) 99115www.elsevier.com/locate/rsetemperatures as related to specific landscape and biophysicalcomponents and (ii) through relating surface temperatures withenergy fluxes for specific landscape phenomena or processes

g o(Quattrochi & Luvall, 1999). Different thematic areas whereTIR remote sensing data have been applied to the analysis oflandscape attributes or land surface processes could be found, asfor example: (i) landscape characterization, (ii) thermal intertiaand landscape analysis, (iii) estimation of energy fluxes, (iv)evaporation/evapotranspiration/soil moisture, (v) quantificationof energy balance or energy flux, and (vi) forest energyexchange.

In all these thematic areas and other environmental studies,land surface temperature (LST) is a key parameter which can beretrieved from TIR data. Hence, except for solar irradiancecomponents, most of the fluxes at the surface/atmosphereinterface can only be parameterized through the use of surfacetemperature. LST can play either a direct role, such as whenestimating long wave fluxes, or indirectly as when estimatinglatent and sensible heat fluxes (Kerr et al., 2004). LST can bealso used as an input data in water and energy balance studies,which is an important issue in environmental studies in order toachieve a better understanding on the exchange of heat andmoisture between the land surface and lower atmosphere, alsoleading to a better understanding on the water and carboncycles. Moreover, many other applications rely on theknowledge of LST, such as geology, hydrology, vegetationmonitoring, global circulation models, and evapotranspiration,among others.

Different approaches have been published in the last years inorder to retrieve LST from satellite data. However, most of theseapproaches have been developed for low spectral resolutionsensors, with only one or two thermal bands, as the ThematicMapper (TM) onboard the LANDSAT platform, the AdvancedVery High Resolution Radiometer (AVHRR) onboard theNOAA series, the Along Track Scanning Radiometer (ATSR)onboard ERS-1 and ERS-2 platforms, and the most recentAdvanced ATSR (AATSR) onboard the ENVISAT platform.Among these methods, we highlight the single-channel andtwo-channel or split-window algorithms, which will bedescribed below. Different spectral and spatial resolutionsensors are currently available, with several thermal bands asthe Advanced Spaceborne Thermal Emission and ReflectionRadiometer (ASTER), and the Moderate Resolution ImagingSpectrometer (MODIS) onboard Terra (or also AQUA forMODIS) satellite, or the SEVIRI (Spinning Enhanced Visibleand Infrared Imager) onboard the MSG-1 (Meteosat SecondGeneration). Airborne sensors available are the digital airborneimaging spectrometer (DAIS), the airborne hyperspectralscanner (AHS), or the airborne reflective/emissive spectrometer(ARES), among others. Despite the methods developed for lowresolution sensors that can be adapted to high resolution data,new methods have been developed in order to retrieve LSTfrom multispectral thermal data, as the TES method (Gillespieet al., 1998).

The purpose of this paper is to analyze the feasibility ofretrieving LST from AHS thermal infrared data as well as theaccuracy obtained depending on the number of thermal bands

100 J.A. Sobrino et al. / Remote Sensinused on the LST retrieval. The paper is organized as follows: inSection 2, we describe the methods used in this study for LSTretrieval; in Section 3, we present the field campaigns andground-based measurements; and in Section 4, we describe theAHS sensor. The results obtained are shown in Section 5, whichincludes the analysis of the calibration of the AHS thermalbands and the tests and validation conducted on the algorithms.Finally, we present the conclusions drawn in this study.

2. Methods for LST retrieval

Different methods or techniques have been proposed in thelast years in order to retrieved LST from thermal infrared data.A review of methods can be found in Sobrino et al. (2002), Dashet al. (2002), and Kerr et al. (2004). Basically, these methodscould be classified as (i) single-channel methods, which usedonly one thermal band; (ii) two-channel or split-windowmethods, which use a combination between two thermalbands; and (iii) two-angle methods, which use one thermalband and two view angles. There are also other methods, whichuse more than two thermal bands or based on other techniques(see for example Becker & Li, 1990a; Sun & Pinker, 2003; Wan& Li, 1997). The availability of sensors with multispectralcapabilities in the thermal infrared region has also favoured thedevelopment of methods for LST retrievals, which use severalthermal bands, as the temperature and emissivity separation(TES) method, developed by Gillespie et al. (1998), which alsoprovides surface emissivities jointly with the temperature. Inorder to retrieve LST from AHS data, we considered the single-channel, two-channel, and TES methods, which are widely usedby the scientific community and described below.

2.1. Theoretical background

Methods for LST retrieval based on the radiative transferequation, which can be written in the thermal infrared region fora certain sensor band i as:

LiTi LLLRi si Lzi 1where Li (Ti) is the radiance measured by the sensor (Ti is the at-sensor brightness temperature), i is the atmospheric transmis-sivity, and Li

is the up-welling path radiance. The term LLLR isthe land-leaving radiance (LLR) or radiance measured atground-level, which is given by:

LLLRi eiBi Ts 1 ei FAip

2

where i is the surface emissivity, Bi (Ts) is the Planck radianceat surface temperature Ts, and Fi

is the down-welling skyirradiance. In Eq. (2), the assumption of Lambertian behaviourfor the surface has been considered in order to express thereflection term as (1) 1F. The magnitudes involved inEqs. (1) and (2) are band averaged values using the spectralresponse functions.

2.2. Single-channel method

f Environment 102 (2006) 99115Single-channel methods retrieve LST only from one thermalband. Different single-channel algorithms can be found in the

(2005), providing good results (root mean square errorsthermal band for which the emissivity is maximum, in order to

g oRMSE

Fig. 1. The study area of Barrax. The image corresponds to AHS band 75(10.07m) raw data and it was acquired on 15 July 2004 at 12:43 UTC, with a

102 J.A. Sobrino et al. / Remote Sensing o25 July 2004 at 9:30 UTC, in which the olive orchard ishighlighted.

In the SPARC/EAGLE field campaign, temperaturemeasurements were made using different broadband andmultiband field thermal radiometers. Models EVEREST 3000,RAYTEK ST8, and RAYTEK MID correspond to broadbandradiometers, whereas models CIMEL CE 312-1 and CIMELCE 312-2 ASTER are multiband radiometers (they also havea broad band). In addition, two blackbodies (EVEREST 1000and GALAI 204-P) were used for calibration purposes. Table1 summarizes the main technical characteristics of the thermalradiometers. Field measurements were made over two plots ofbare soil (referred as bare soil 1 and bare soil 2) and twoplots of corn (referred as corn 1 and corn 2), one plot of

pixel size of 7m.water and one plot of grass (see Fig. 1). Two differenttechniques were considered in order to measure surface

Fig. 2. The study area of Crdoba. The image has been obtained from the at-sensor radiance in AHS band 75 (10.07m) and it was acquired on 25 July 2004at 9:30 UTC, with a pixel size of 2.5m.temperatures: (i) by means of transects, i.e., measuring whilewalking along the field, and (ii) by means of fixedmeasurements by placing the radiometers on poles with analtitude of around 2m.

A total of 10 infrared sensors Apogee model IRTS-P wereplaced on poles with an altitude of 6m over 10 olive trees in theCordoba site (AGRISPECTRA project) in order to monitorcrown temperature continuously as function of a tree waterstatus gradient obtained through drip irrigation method.Previous to the field installation, the IRT sensors were calibratedin the laboratory and under natural sun conditions to study theIRT response to the diurnal temperature variation. Temperatureover the course of the day varied between 25C and 40C,enabling a comparison between the IRT-estimated temperatureand a thermocouple type K (chromelalumel) in contact withthe water target used for calibration. The observed errors agreedwith the sensitivity of the instrument (Apogee, www.apogee-inst.com) yielding a deviation of 0.4C between the 5C and40C range. The 52 field-of-view (FOV) of the IRTS-P sensorsplaced on the top of each olive tree at 1m distance from thecrown enabled the measurement of an integrated canopytemperature for each single tree crown.

Table 1Main technical specifications for the field radiometers (FOV: field of view)

Model Bands m Range C Accuracy C FOVC

CIMEL 312-1 813 80 to 60 0.1 108.29.210.311.311.512.5

CIMEL 312-2 813 80 to 60 0.1 101111.7

10.3118.99.38.58.98.18.5

EVEREST 3000 814 40 to 100 0.5 4RAYTEK ST8 814 30 to 100 0.5 8RAYTEK MID 814 40 to 600 0.5 20

f Environment 102 (2006) 991154. The AHS sensor

4.1. Technical characteristics

The airborne hyperspectral scanner (AHS) (developed bySensyTech Inc., currently ArgonST, USA) is operated by theSpanish Institute of Aeronaoutics (INTA) and it was placedonboard the aircraft CASA 212-200 Paternina. The AHS sensoris based on the integration of many advanced technologiesdeveloped by SenSyTech under R&D contracts over the pastfew years. While the combination of these components isoffered here for the first time, each of the individual items hasbeen delivered and field-tested in operational use. The AHSincorporates advanced components to ensure high performancewhile maintaining the ruggedness to provide operationalreliability in a survey aircraft. The main AHS technicalspecifications and the arrangement of the spectral bands are80 bands in four ports (VIS, NIR, SWIR, MWIR, and LWIR);

FOV: 90 (45), IFOV: 2.5mrad, GFOV: 26m at 140ktcruise speed; scan speed: 6.25, 12.5, 18.75, 25, 31.25, 35rps; 12bits digitised; 750 samples per line; black body thermalreferences (set to 15C and 55C). The arrangement of the

4.3. AHS data processing and atmospheric correction

The processing of the AHS imagery included the raw data toradiance transformation and the atmospheric correction. AHSimages were processed to at-sensor radiance for the thermalinfrared bands (from 71 to 80) using calibration coefficientsdetermined during flight from two blackbody sources which areviewed for every mirror scan. The first (cool) blackbody hasbeen set to 15C, whereas the second (warm) blackbody hadbeen set to 55C. This procedure is performed for every thermalband on every scanline. The AHS onboard blackbodies arecopper plates covered with a black paint with emissivity greaterthan 0.99 through the 3 to 14m spectral range (note that thisspecification permits a bias

radiosoundings launched almost simultaneously with the AHSoverpass. The band values were finally obtained using the filterfunctions of the AHS thermal bands. These filter functions areshown in Fig. 3 compared to the atmospheric transmissivity forthe low flight (975m AGL), the high flight (2745m AGL), and atypical satellite altitude (700km). Bands 71 and 80, locatedaround 8 and 13m, respectively, show the highest atmosphericabsorption, whereas bands 75 to 79 are located in theatmospheric window 1012.5m, with band 75 showing thehighest atmospheric transmissivity. Band 74 is located in theregion of the ozone absorption, but in the AHS flights thisabsorption is not observed because the maximum absorptionof the ozone is usually located at atmospheric altitudes higherthan 10km.

5. Results and analysis

5.1. LST estimation from ground-based measurements

The thermal radiometers described in Section 3 and used atground-level measure the land-leaving radiance (LLLR) given by

Eq. (2), since at ground-level is assumed that 1 and L0.Surface emissivity values () can be measured in situ, forexample using the box method (Nerry et al., 1990), or extractedfrom spectral libraries, as for example the ASTER spectrallibrary (http://speclib.jpl.nasa.gov), whereas F/ can beapproximately measured by pointing with the radiometer tothe sky with a near-nadir view or, more accurately, by pointingto the sky with a view angle of 53 measured from the zenith(Kondratyev, 1969). Therefore, from Eq. (2), it is possible tofind the value of Ts from ground-based measurements and byinversion of the Planck's law. The term radiometric temper-ature (Trad) is used when Planck's law is inverted using thequantity LLLR, so the difference between the land surfacetemperature (Ts) and the radiometric temperature (Trad) is due tothe emissivity () and atmospheric effects (F), both coupled.

In order to illustrate the differences between radiometric andland surface temperatures and also the problems related with thethermal heterogeneity of the natural surfaces, Fig. 4 shows anexample of the Ts and Trad values obtained with the Raytek MIDradiometer fixed on the mast between 13:00 and 13:15 LT (localtime, GMT=LT2) the 15th of July in 2004 over bare soil (Fig.

45

46

47

48

49

50

51

52

13:

Loc

Tem

pera

ture

(C)

TradTs

(a)

13:0

oc

104 J.A. Sobrino et al. / Remote Sensing of Environment 102 (2006) 9911542

43

44

12:59 13:00 13:01 13:02 13:03 13:04 13:05 13:06

Tem

pera

ture

(C)

(b)

28.0

28.5

29.0

29.5

30.0

30.5

31.0

31.5

32.0

12:59 13:00 13:01 13:02 13:03 13:04 13:05 13:06

LFig. 4. Radiometric temperature (Trad) and surface temperature (Ts) measured with thgrass.07 13:08 13:09 13:10 13:11 13:12 13:13 13:14 13:15 13:16

al Time

7 13:08 13:09 13:10 13:11 13:12 13:13 13:14 13:15 13:16

al Time

TradTse Raytek MID radiometer located on fixed masts over (a) bare soil and (b) green

4a) and grass (Fig. 4b) in the Barrax site. The differencebetween Ts and Trad is higher for bare soil (around 2K) than forgreen grass (around 1K), due to the higher emissivity for greengrass. Surfaces with high surface emissivity minimize the effectof the emissivity and atmospheric correction. The graphsincluded in Fig. 4 also show a higher heterogeneity for baresoil than for green grass. Hence, the difference between themaximum and minimum value for bare soil is 5K, whereas forgreen grass is 1.2K. Moreover, differences between two conse-cutive measurements (in steps of 1min) are higher than 1.5K forbare soil but lower than 0.5K for green grass. The heterogeneityof the surfaces can be also shown by calculating the mean valueand the standard deviation of the measurements. Values of Ts=(321.01.5) K for bare soil and Ts= (304.10.4) K for greengrass have been obtained, showing a higher standard deviationfor bare soil. Another example is presented in Fig. 5, whichshows a comparison between the values measured with theRaytek MID radiometer located on fixed masts and the valuesmeasured with the Raytek ST8 radiometer by making transectsaround the same sample for bare soil and green grass. This

figure shows a significant variability in the measurementscarried out by transects, as has been stated before. Thedifferences between the fixed measurements and the transectmeasurements are also lower for green grass, which is a morehomogeneous surface. The variability shown in the radiometrictemperatures can be also due to turbulence induced fluctuationsof wind speed at the surface (Balick et al., 2003). Thesefluctuations can be larger than 1K, so accuracies worse than 1Kare always expected in the validation procedures. Theseexamples illustrate the difficulties involved in the fieldmeasurements and the validation of satellite products, mainlyover heterogeneous surfaces.

5.2. Analysis of the AHS thermal data: calibration and optimalbands

Eqs. (1) and (2) can be used in order to reproduce theradiance measured by the sensor (Li) or the at-sensor brightnesstemperature (Ti) (by inversion of the Planck's law) if theatmospheric parameters (i, Fi

, and Li), the emissivity spectrum

40

42

44

46

48

50

0

oca

Rad

iom

etric

Tem

pera

ture

(C)

RAYTEK MID (Fixed)RAYTEK ST8 (Transect)

(a)

13oca

105J.A. Sobrino et al. / Remote Sensing of Environment 102 (2006) 991153812:51 12:54 12:57 13:0

L

Rad

iom

etric

Tem

pera

ture

(C)

(b)

25

26

27

28

29

30

31

32

33

34

35

13:24 13:26 13:27 13:29 13:30LFig. 5. Comparison between the radiometric temperature measured with the Raytek Mwith the Raytek ST8 radiometer by making transects over (a) bare soil and (b) gree13:03 13:06 13:09 13:12

l Time

:32 13:33 13:35 13:36 13:37 13:39l Time

RAYTEK MID (Fixed)RAYTEK ST8 (Transect)ID radiometer located on fixed masts and the radiometric temperature measuredn grass.

(i), and the LST (Ts) are known. Values of at-sensor brightnesstemperatures have been compared with the extracted from theAHS images (by selecting boxes of mn pixels located aroundthe measurement sites) in order to check the calibration of theAHS sensor. The atmospheric parameters have been extractedfrom the radiosoundings launched in situ and the MODTRAN 4code. The emissivity spectra for water and grass have beenobtained from the ASTER spectral library, whereas for the cornplot a constant value of 0.99 have been assumed due to its highcoverage. One sample of soil was collected at the field and sentto the Jet Propulsion Laboratory (JPL) in order to measure itsemissivity spectrum. Fig. 6 shows all the emissivity spectra forthe AHS thermal bands considered in this study. The analysis isonly shown for the SPARC/EAGLE campaign, since in theAGRISPECTRA campaign the atmosphere was not character-ized. From all the AHS images and ground-based measurementsavailable, only the plots measured simultaneously with thesensor overpass have been selected for checking the calibration

emissivity and reflection terms (1)1F involved in Eq. (2)have a minimal contribution. The bare soil plot provides a goodshape for the spectrum; however, high differences (>4K) havebeen found, except for the bare soil observed in the AHS imageacquired on 18 July at 11:03 UTC (low flight, Fig. 8). Anexplanation of the different behaviour of this bare soil and theothers could be found in the different radiometers used in themeasurements. This last bare soil plot was measured with theCIMEL radiometer, which is the most accurate thermalradiometer. Problems on the ground-based measurements havebeen also found in the corn plots for the low flight carried out on18 July at 11:03 UTC (see Fig. 8). Differences when comparingwith ground-based measurements higher than 4K have beenobtained. The corn plot for the AHS image acquired at 10:50UTC (also included in Fig. 8) shows ground-based brightnesstemperatures substantially different to those measured at 11:03UTC. Taking into account that the corn plot was in a fullycovered stage and well irrigated, temperatures between the two

vele

106 J.A. Sobrino et al. / Remote Sensing of Environment 102 (2006) 99115of the AHS thermal bands, leading to a total amount of 19 plotsacquired over different surfaces (bare soil, grass, corn, andwater). The results obtained are shown in Fig. 7 for the highflight (975m AGL) and in Fig. 8 for the low flight. All thegraphs shown in these figures have been scaled in steps of 1K(Y-axis, horizontal lines) for a better comparison of thedifferences obtained over the different plots and flights.According to the results obtained, AHS band 73 (9.15m)seems to be affected by a calibration problem, because thevalues obtained with this band do not follow the expectedtendency in the shape of the brightness temperatures versus thewavelength. AHS band 71 (8.183m) also shows somecalibration problems in the low flights (see Fig. 8), whereasband 72 (8.659m) seems to show a random behaviour.Looking to the different plots considered in this analysis, theresults obtained over water provide, in general, the betteraccordance with the ground-based measurements. It should benoted that the water behaviour is similar to a blackbody, so the

0.94

0.95

0.96

0.97

0.98

0.99

1.00

8.0 8.5 9.0 9.5 10.0Wa

Emis

sivi

tyFig. 6. Emissivity spectra for different plots. Emissivities for water and grass have becollected at the field and measured in the Jet Propulsion Laboratory (JPL). A constanthigh cover.consecutive flights should be similar. Significant differenceshave been found also in some green grass plots. Despite theseplots have been labelled as green, a visual inspection at thefield site shows a mixed contribution composed by green andalso senescent grass, leading to a less homogeneous plot.Despite these problems, the AHS thermal bands provides goodresults in comparison with ground-based measurements for thebands located in the split-window region from 10 to 13m(AHS bands from 75 to 79).

5.3. Simulated data for single-channel and split-windowalgorithms

In order to obtain operative algorithms for retrieving LSTwith single-channel or split-window methods, different atmo-spheric conditions have been simulated. For this purpose theMODTRAN 4 radiative transfer code (executed in the thermalradiance mode) and a set of 54 radiosoundings extracted from

10.5 11.0 11.5 12.0 12.5 13.0ngth (m)

WATERBARE SOILGRASSCORNen extracted from the ASTER spectral library, whereas the bare soil sample wasvalue equal to 0.99 has been considered for the corn plot, due to its greenness and

00

1elen

g o15 July 2

290

295

300

305

310

315

320

8.0 8.5 9.0 9.5 10.0Wav

At-s

enso

r Brig

htne

ss T

empe

ratu

re (K

)

BARE SOIL 1

GRASS

WATER

J.A. Sobrino et al. / Remote Sensinthe TIGR (TOVS Initial Guess Retrieval) database (Scott &Chedin, 1981) and described in Sobrino et al. (1993b) havebeen used. In this way, spectral values of the atmosphericparameters , F, and L have been obtained. These values havebeen averaged according to the AHS filter functions in order toobtain the effective values for the thermal bands from 75 to 79(see Eq. (3)). Fig. 3 showed that the highest is obtained forAHS band 75. For this reason, band 75 (10.07m) has beenchosen as the optimal band for applying the single-channelalgorithm given by Eq. (3). The dependence of this algorithmwith the atmospheric parameters have been avoided by findingrelations with w. Taking into account the set of the 54radiosoundings mentioned before, the following results havebeen obtained for the low flight (975m AGL):

s75 0:0505w2 0:0691w 1:0010R2 0:998; r 0:003 8Lz75 0:5638w2 0:4498w 0:0164R2 0:998; r 0:03 Wd m2d Am1d sr1 9

18 July 200

290

295

300

305

310

315

8.0 8.5 9.0 9.5 10.0 1Wavele

At-s

enso

r Brig

htne

ss T

empe

ratu

re (K

)

WATER

GRASS

Fig. 7. Comparison between the ground-based and the AHS at-sensor brightness tem0.5 11.0 11.5 12.0 12.5 13.0gth (mm)4 - 12:43 UTC

ground-basedAHS

107f Environment 102 (2006) 99115Lz75 0:2746w2 3:5473w 0:2231R2 0:981; r 0:3 Wd m2d Am1d sr1 10

and for the high flight (2745m AGL):

s75 0:0143w2 0:0678w 1:0006R2 0:997; r 0:007 11Lz75 0:1531w2 0:4900w 0:0513R2 0:995; r 0:07 Wd m2d Am1d sr1 12LA75 0:0431w2 1:8217w 0:1503R2 0:976; r 0:3 Wd m2d Am1d sr1 13where LF/, R2 is the correlation coefficient squared, and the standard error of estimation.

In order to select the best combination between the bands iand j involved in the split-window algorithm (see Eq. (16)),the numerical coefficients ak have been obtained for all thecombinations between AHS bands from 75 to 79, and then a

4 - 12:17 UTC

0.5 11.0 11.5 12.0 12.5 13.0ngth (m)

ground-basedAHS

perature for the high flights (2745m above ground level) over different surfaces.

15 July 2004 - 11:57 UTC

2908.0 8.5 9.0 9.5 10.0 10.5 11.0 11.5 12.0 12.5 13.0

295

300

305

310

315

320

325

Wavelength (m)

8.0 8.5 9.0 9.5 10.0 10.5 11.0 11.5 12.0 12.5 13.0Wavelength (m)

8.0 8.5 9.0 9.5 10.0 10.5 11.0 11.5 12.0 12.5 13.0Wavelength (m)

At-s

enso

r Brig

htne

ss T

empe

ratu

re (K

) ground-basedAHS

WATER

GRASS

BARE SOIL 1

18 July 2004 - 10:30 UTC

292

293

294

295

296

297

At-s

enso

r Brig

htne

ss T

empe

ratu

re (K

) ground-basedAHS

WATER

18 July 2004 - 10:50 UTC

292

294

296

298

300

302

304

306

308

At-s

enso

r Brig

htne

ss T

empe

ratu

re (K

) WATER: ground-basedWATER: AHSCORN 1: ground-basedCORN 1: AHSCORN 2: ground-basedCORN 2: AHS

Fig. 8. Comparison between the ground-based and the AHS at-sensor brightness temperature for the low flights (975m above ground level) over different surfaces.

108 J.A. Sobrino et al. / Remote Sensing of Environment 102 (2006) 99115

g oJ.A. Sobrino et al. / Remote Sensinsensitivity analysis according to Sobrino et al. (2004) (see Eqs.(10) to (14) in the cited paper) has been carried out in order toestimate the error on the LST retrieved with these algorithms.

18 July 2004

290

295

300

305

310

315

320

325

330

At-s

enso

r Brig

htne

ss T

empe

ratu

re (K

)

WATER

GRASS

BARE SOIL 2

18 July 2004

298

300

302

304

306

At-s

enso

r Brig

htne

ss T

empe

ratu

re (K

)

18 July 200

290

295

300

305

310

315

At-s

enso

r Brig

htne

ss T

empe

ratu

re (K

)

WATER

GRASS

8.0 8.5 9.0 9.5 10.0 1Wavele

8.0 8.5 9.0 9.5 10.0 1Wavele

8.0 8.5 9.0 9.5 10.0 1Wavele

Fig. 8 (con109f Environment 102 (2006) 99115At-sensor brightness temperatures for the bands i and jinvolved in Eq. (5) are simulated using the radiative transferequation (see Eqs. (1) and (2)), in which atmospheric

- 11:03 UTC ground-basedAHS

- 11:03 UTC

CORN 1: ground-basedCORN 1: AHSCORN 2: ground-basedCORN 2: AHS

4 - 11:28 UTCground-basedAHS

0.5 11.0 11.5 12.0 12.5 13.0ngth (m)

0.5 11.0 11.5 12.0 12.5 13.0ngth (m)

0.5 11.0 11.5 12.0 12.5 13.0ngth (m)tinued).

5.4. Simulated data for the TES method: relation minMMDand testing

The MMD module included in the TES method uses anempirical relationship between the minimum emissivity (min)and the spectral contrast (MMD) in order to recover thesurface emissivities (Gillespie et al., 1998). This relationship,initially obtained for the ASTER sensor, needs to berecalculated when other sensors are used. The emissivityspectra (299 samples) described in the previous section havebeen employed to find the relation between min and MMDwhen AHS bands 75 to 79 are considered. Fig. 9 shows theplot of min versus MMD for the 299 emissivity spectra ofnatural surfaces. The following relation has been finallyobtained for AHS data:

g of Environment 102 (2006) 99115parameters are obtained again with the MODTRAN 4 codeand the set of 54 radiosoundings. In this simulation, surfaceemissivity values are also needed. A total amount of 299emissivity spectra extracted from the ASTER spectral library(http://speclib.jpl.nasa.gov) and averaged according to theAHS filter functions have been used in this simulation. All theemissivity spectra correspond to natural surfaces, as is shownin Table 3. Using 54 radiosoundings and 299 emissivityspectra, we obtain 16146 data simulating different atmosphericconditions over different natural surfaces. With this amount ofsimulated data, statistical fits are applied in order to recoverthe numerical coefficients involved in the split-windowalgorithm. The simulated data is also used in the sensitivityanalysis (Eqs. (10) to (14) in Sobrino et al., 2004), in order toobtain the theoretical error of the algorithm over differentconditions. The mean value of the error and the standarddeviation are used in order to compute the root mean squareerror (RMSE), assumed to be the final theoretical error of the

Table 3Emissivity spectra extracted from the ASTER spectral library and used in thesimulation procedure

Class Subclass Samples

Rocks Igneous 100Metamorphics 76Sedimentary 68

Soils Alfisol 9Aridisol 14Entisol 10Inceptisol 7Mollisol 9

Vegetation Green grass 1Dry grass 1Conifers 1Decideous 1

Watersnowice Water 1Ice 1

Total: 299

110 J.A. Sobrino et al. / Remote Sensinalgorithm. The results obtained are shown in Table 4 for thedifferent flight altitudes and also for a typical satellite altitude.Regarding to the low and high flights, the best combination isobtained with AHS bands 75 (10.07m) and 79 (12.35m).When a satellite altitude is considered, the best combination isobtained with AHS bands 76 (10.59m) and 79 (12.35m),which agrees with typical split-window combinations used inother thermal sensors as NOAA-AVHRR or ENVISAT-AATSR. In all the cases (low, high, and satellite altitude),the error is around 1K. The split-window algorithms with thenumerical coefficients for the low and high flight are given inEqs. (14) and (15), respectively:

Ts T75 0:485T75 T79 0:0068T75 T792 0:0798 47:15 10:80W 1 e 49:05 21:53wDe 14

Ts T75 0:734T75 T79 0:0096T75 T792 0:1198 47:46 5:20W 1 e 61:82 14:97wDe 15emin 0:986 1:350MMD1:019 16

with a correlation coefficient squared of R2 =0.93 and astandard error of estimation of =0.019. It should be notedthat the MMD values involved in Eq. (16) have beenobtained from AHS bands 75 to 79, located in the 1012m region. In theory, Eq. (16) works better when bandslocated in the 810m region are also included. AHSbands located in this region (71 to 74) have not beenincluded due to the calibration problems found (Section5.2). Despite some problems could arise when using onlyAHS bands 75 to 79 over high MMD surfaces (rocks),accurate results are expected over low MMD surfaces asagricultural areas. AHS images with no calibration problemswill require a new computation for the relationship betweenmin and MMD.

The TES method has been tested using the simulated datadescribed in the previous section for different flight altitudes:700km, 975m, and 2745m. The LST retrieved with the methodhas been compared with the LST included in the simulated dataand extracted from the radiosoundings. In all the cases, the TESmethod provided a RMSE

5 0.20 0.25 0.30 0.35M

299 emissivity spectraR2 = 0.93

= 0.019

ontr

111g of Environment 102 (2006) 991155.5. Algorithms validation with ground truth data

The algorithms described in the previous sections have beenvalidated using the field measurements carried out in theframework of the different field campaigns. Single-channel andtwo-channel methods have been tested in the framework of theSPARC/EAGLE and AGRISPECTRA field campaigns. In theSPARC/EAGLE campaign, a total amount of 17 plotscomposed by water, grass, bare soil, and corn have been used,whereas in the AGRISPECTRA campaign a total amount of 30plots have been used (10 trees per flight). In order to apply theTES method, an accurate estimation of the atmosphericparameters is needed, since it uses atmospherically correcteddata. In the SPARC/EAGLE campaign, different radiosound-ings were launched almost simultaneously with the sensoroverpass. However, in the AGRISPECTRA campaign, radio-

0.5

0.6

0.7

0.8

0.9

1.0

0.00 0.05 0.10 0.1

min

Fig. 9. Empirical relationship between minimum emissivity (min) and spectral ccorrelation coefficient squared and is the standard error of estimation).

J.A. Sobrino et al. / Remote Sensinsounding data or atmospheric characterizations were notavailable, so TES method has been only tested in the frameworkof the SPARC/EAGLE campaign. The values of atmosphericwater vapour content needed in order to apply the single-channel and split-window algorithms are given in Table 5.These values have been obtained using the radiosoundingslaunched in situ in the framework of the SPARC/EAGLEcampaigns and the MODTRAN 4 radiative transfer code. In theAGRISPECTRA campaign, no radisoundings were launched,as has been commented before, so in this case we only have thevalue measured with a sunphotometer in the El Arenosillosite, located around 250km far from the field and which isincluded in the AERONET (Aerosol Robotic Network). Detailscan be found at http://aeronet.gsfc.nasa.gov. The total watervapour content extracted from the AERONET data has beenscaled to the altitude of the flight (980m AGL) using theMODTRAN 4 code and assuming the atmospheric profileincluded in the midlatitude summer atmosphere.

The test of the single-channel method is shown in Fig. 10afor the SPARC/EAGLE and in Fig. 10b for AGRISPECTRA,with RMSE values of 1.6K and 1.9K, respectively. The Biashas been calculated as the mean value for the differencebetween the LST retrieved with the algorithm and the measuredin situ. In both cases, the Bias is positive, which indicates thatthe algorithm overestimates the LST. This overestimation ismore clearly shown in the results achieved in the AGRISPEC-TRA campaign, with a Bias of 1.7K. We attribute this to thepoor characterization of the atmosphere during this experiment,in which a water vapour value measured 250km far from thefield site has been chosen. The results obtained with the single-channel algorithm proposed in this paper agree with thoseobtained by Jimnez-Muoz and Sobrino (2005) also from highresolution data over the Barrax test site, with RMSE

T in

g o290

295

300

305

310

315

320

325

330

290 295 300 305LS

LST

sing

le-c

hann

el (K

)

BIAS = 1.1 KSt. Dev. = 1.1 KRMSE = 1.6 K

(a)

(b)

112 J.A. Sobrino et al. / Remote Sensinresults have been obtained with split-window and single-channel algorithms, the split-window technique provides betterresults over a world-wide scale, overall in wet atmospheres(w>3g/cm2), whereas single-channel methods only providegood results for low atmospheric absorption (w

290

295

300

305

310

315

320

325

330

290 295 300 305 310 315 320 325 330LST in situ (K)

LST in situ (K)

LST

split

-win

dow

(K)

Low flight (975m AGL)High Flight (2745m AGL)

BIAS = 1.2 KSt. Dev. = 1.5 KRMSE = 1.9 K

(a)

300

305

310

315

320

300 305 310 315 320

LST

split

-win

dow

(K)

Flight 1 (7:30 GMT)Flight 2 (9:30 GMT)Flight 3 (12:30 GMT)

BIAS = 1.3 KSt. Dev. = 0.9 KRMSE = 1.6 K

(b)

Fig. 11. Comparison between the land surface temperature retrieved with the split-window method and the one measured in situ in the framework of the (a) SPARC/EAGLE and (b) AGRISPECTRA field campaigns.

290

295

300

305

310

315

320

325

330

290 295 300 305 310 315 320 325 330LST in situ (K)

LST

TES

(K)

Low Flight (975m AGL)High Flight (2745m AGL)

BIAS = 1.0 KSt. Dev. = 1.0 KRMSE = 1.4 K

Fig. 12. Comparison between the land surface temperature retrieved with the TES method and the one measured in situ in the framework of the SPARC/EAGLE fieldcampaign.

113J.A. Sobrino et al. / Remote Sensing of Environment 102 (2006) 99115

g ocontext of precision agriculture and vegetation stress monitor-ing. The airborne hyperspectral scanner (AHS) sensor, with 10

298 318 300 324 2 5

(a) (b) (c) Fig. 13. (a) At-sensor brightness temperature obtained from the AHS band 75(T75), (b) land surface temperature (LST) obtained with the TES method and (c)difference image (T=LSTT75). The AHS image was acquired on 15 July2005 over the Barrax test site (see Fig. 1) at 12:43 UTC and the spatial resolutionis 7m. Temperature values are given in K.

114 J.A. Sobrino et al. / Remote Sensinthermal bands (from 71 to 80) covering the range between 8 and13m, is an example of the new generation of airborne sensorswith multispectral thermal infrared capabilities.

The feasibility of retrieving LST from the 10 AHS thermalbands has been analyzed in the framework of the SPARC andEAGLE field campaigns, which took place simultaneously inthe Barrax (Albacete, Spain) test site, and in the framework ofthe AGRISPECTRA project, which took place in an orchardcrop field in Crdoba (Spain), all of them in July 2004. For thispurpose, the single-channel, split-window, and TES (tempera-ture and emissivity separation) methods have been adapted tothe AHS characteristics. Previously, the at-sensor brightnesstemperatures extracted from the AHS thermal bands werecompared with in situ measurements carried out over differentplots (bare soil, water, grass, and corn) in order to assess theoptimal selection of bands for LST retrieval. This comparisonhas shown that the AHS sensor had a calibration problem withbands 71, 72, and 73. Bands from 75 to 79, located in theatmospheric window between 10 and 12m, seem to agreebetter with in situ data, so these bands were finally selected. Thesingle-channel method, which only uses one thermal band, wasapplied to AHS band 75 because this band showed the highestatmospheric transmissivity. The split-window algorithm, whichused a combination between two thermal bands, was applied toAHS bands 75 and 79, because this combination showed thelowest error on the LST. Both algorithms were validated usingground truth data, showing similar results, with root meansquare errors (RMSE) lower than 2K. The good results obtainedwith the single-channel method are attributed to the lowatmospheric absorption during the flights, with total atmospher-ic water vapour contents lower than 1.8g/cm2 for the highflights (2745m above ground level) and lower than 1g/cm2 forthe low flights (945m above ground level). With higher watervapour contents and, therefore, with higher atmosphericabsorption, the split-window algorithms are expected to providebetter results, whereas single-channel algorithms becomealmost unusable (Sobrino & Jimnez-Muoz, 2005). The TESmethod, developed originally for the ASTER sensor byGillespie et al. (1998), was applied to AHS bands from 75 to79, and it was also tested with simulated data, with aRMSE

Jimnez-Muoz, J. C., & Sobrino, J. A. (2003). A generalized single-channelmethod for retrieving land surface temperature from remote sensing data.Journal of Geophysical Research, 108(D22). doi:10.1029/2003JD003480.

Jimnez-Muoz, J. C., & Sobrino, J. A. (2005). Atmospheric water vapourcontent retrieval from visible and thermal data in the framework of theDAISEX campaigns. International Journal of Remote Sensing, 26(15),31633180.

Kerr, Y. H., Lagouarde, J. P., Nerry, F., & Ottl, C. (2004). Land surfacetemperature retrieval techniques and applications: Case of AVHRR. In D. A.Quattrochi, & J. C. Luvall (Eds.), Thermal remote sensing in land surfaceprocesses (pp. 33109). Florida, USA: CRC Press.

Kondratyev, K. Y. (1969). Radiation in the atmosphere. New York: AcademicPress.

McClain, E. P., Pichel, W. G., & Walton, C. C. (1985). Comparativeperformance of AVHRR-based multichannel sea surface temperatures.Journal of Geophysical Research, 90(11), 587601.

McMillin, L. M. (1975). Estimation of sea surface temperature from two infraredwindow measurements with different absorption. Journal of GeophysicalResearch, 80, 51135117.

Moreno, J., Calera, A., Caselles, V., Cisneros, J. M., Martnez-Lozano, J. A.,

Sepulcre-Cant, G., Zarco-Tejada, P.J., Jimnez-Muoz, J.C., Sobrino, J.A., deMiguel, E., Villalobos, F.J., in press. Within-field thermal variabilitydetection as function of water stress inOlea europaea L. orchards with high-spatial remote sensing imagery. Agricultural and Forest Meteorology(available online).

Sobrino, J. A., Caselles, V., & Coll, C. (1993). Theoretical split windowalgorithms for determining the actual surface temperature. Il NuovoCimento, 16, 219236.

Sobrino, J. A., Coll, C., & Caselles, V. (1991). Atmospheric correction for landsurface temperature using NOAA-11 AVHRR Channels 4 and 5. RemoteSensing of Environment, 38, 1934.

Sobrino, J. A., & Jimnez-Muoz, J. C. (2005). Land surface temperatureretrieval from thermal infrared data: An assessment in the context of theSurface Processes and Ecosystem Changes Through Response Analysis(SPECTRA) mission. Journal of Geophysical Research, 110(D16103).doi:10.1029/2004JD005588.

Sobrino, J. A., Jimnez-Muoz, J. C., El-Kharraz, J., Gmez, M., Romaguera,M., & Sria, G. (2004). Single-channel and two-channel methods for landsurface temperature retrieval from DAIS data and its application to theBarrax test site. International Journal of Remote Sensing, 25(1), 215230.

Sobrino, J. A., Li, Z. -L., Soria, G., & Jimnez, J. C. (2002). Land surfacetemperature and emissivity retrieval from remote sensing data. RecentResearch Developments on Geophysics, 4, 2144.

Sobrino, J. A., Li, Z. -L., & Stoll, M. P. (1993). Impact of the atmospherictransmittance and total water vapor content in the algorithms for estimating

115J.A. Sobrino et al. / Remote Sensing of Environment 102 (2006) 99115final results workshop, SP-499 (pp. 4351). Noordwijk (Holland): ESAPublications Division.

Nerry, F., Labed, J., & Stoll, M. P. (1990). Spectral properties of lands surfacesin the thermal infrared band: Part II. Field method for spectrally averagedemissivity measurements. Journal of Geophysical Research, 95,70277044.

Prata, A. J. (1993). Land surface temperatures derived from the AVHRR andATSR: 1. Theory. Journal of Geophysical Research, 89(16), 689702.

Price, J. C. (1984). Land surface temperature measurements from the splitwindow channels of the NOAA 7 AVHRR. Journal of GeophysicalResearch, 89, 72317237.

Qin, Z., Karnieli, A., & Berliner, P. (2001). A mono-window algorithm forretrieving land surface temperature from Landsat TM data and its applicationto the IsraelEgypt border region. International Journal of Remote Sensing,22(18), 37193746.

Quattrochi, D. A., & Luvall, J. C. (1999). Thermal infrared remote sensing datafor analysis of landscape ecological processes: Methods and applications.Landscape Ecology, 14(6), 577598.

Quattrochi, D. A., & Luvall, J. C. (2004). Thermal remote sensing in landsurface processes (pp. 440). Boca Raton, FL, USA: CRC Press.

Scott, N. A., & Chedin, A. (1981). A fast line by line method for atmosphericabsorption computations: The authomatized atmospheric absorption atlas.Journal of Meteorology, 20, 802812.satellite sea surface temperatures. IEEE Transactions on Geoscience andRemote Sensing, 31(5), 946952.

Sobrino, J. A., Li, Z. -L., Stoll, M. P., & Becker, F. (1994). Improvements in thesplit-window technique for land surface temperature determination. IEEETransactions on Geoscience and Remote Sensing, 32, 243253.

Sobrino, J. A., & Raissouni, N. (2000). Toward remote sensing methods for landcover dynamic monitoring. Application to Morocco. International Journalof Remote Sensing, 21, 353366.

Sun, D., & Pinker, R. T. (2003). Estimation of land surface temperature from aGeostationary Operational Environmental Satellite (GOES-8). Journal ofGeophysical Research, 108(D11). doi:10.1029/2002JD002422.

Wan, Z., & Li, Z. -L. (1997). A physics-based algorithm for retrieving landsurface emissivity and temperature from EOS/MODIS data. IEEE Transac-tions on Geoscience and Remote Sensing, 35(4), 980996.Meli, J., et al. (2001). The measurement programme at Barrax. DAISEX

Land surface temperature derived from airborne hyperspectral scanner thermal infrared dataIntroductionMethods for LST retrievalTheoretical backgroundSingle-channel methodTwo-channel algorithmsTES method

Field dataThe AHS sensorTechnical characteristicsImagery acquiredAHS data processing and atmospheric correction

Results and analysisLST estimation from ground-based measurementsAnalysis of the AHS thermal data: calibration and optimal bandsSimulated data for single-channel and split-window algorithmsSimulated data for the TES method: relation minMMD and testingAlgorithms validation with ground truth data

Summary and conclusionsAcknowledgementsReferences