AATSR DERIVED LAND SURFACE TEMPERATURE OVER A HETEROGENEOUS REGION José A. Sobrino (1), Guillem Sòria (1), Juan C. Jiménez- Muñoz (1), Belen Franch (1), Victoria Hidalgo (1) and Elizabeth Noyes (2). (1) Global Change Unit, Departament de Física de la Terra I Termodinàmica, Facultat de Física, Universitat de Valencia, Av Dr. Moliner, 50. 46100 Burjassot, Valencia, Spain, sobrino@uv.es (2) Earth Observation Science, Department of Physics and Astronomy, University of Leicester, University Road, Leicester LE17RH, UK ABSTRACT We present a methodology to validate land surface temperature (LST) retrieved from Advanced Along- Track Scanning Radiometer (AATSR) data in heterogeneous areas. To this end, a set of algorithms based on Split-Window (SW) and Dual-Angle (DA) techniques has been proposed and evaluated with in situ measurements of LST obtained from field campaigns made in heterogeneous areas. These campaigns were carried out in Marrakech, Morocco and in Barrax, Spain, in the framework of the WATERMED, EAGLE and SPARC projects. The validation use Landsat and Chris/Proba images over the test sites. Level2 LST product has been evaluated over these heterogeneous areas. Results suggest applying LST algorithms over averaged AATSR images with a smoothing window filter of 2 x 2 pixels. Results show also that DA algorithms could not be applied over heterogeneous regions. A worse accuracy has been observed from the Level2 Product compared with other SW algorithm proposed. 1 INTRODUCTION Retrieval of Land Surface Temperature (LST) from space is of considerable importance for environmental studies, as described in [1]. In the last years, several theoretical studies have been carried out in order to develop LST algorithms through split-window (SW) [2] and [3] and dual-angle (DA) methods, as [4] and [5]. The SW method uses observations at two different spectral bands within 10-12 μm spectral region to eliminate the influence of the atmosphere. The DA method uses observation of the same channel but under two different angles to exploit the different absorption path-lengths. The benefit of the SW method is based upon the fact that the atmospheric absorption of the surface radiation varies strongly with wavelength, and so, atmospheric effects can be corrected by using data from two different spectral channels. In the case of the DA method, the atmospheric transmission and emission of the atmosphere vary as a function of the viewing angle and allows the elimination of uncertainties due to the wavelength dependency. This confirms the advantage of the DA method in comparison with the SW method if the emissivity spectral variation and the emissivity angular variation are of the same order of magnitude. The Advanced Along-Track Scanning Radiometer (AATSR) sensor, as its predecessor sensors from the ATSR serie, includes an angular viewing capability to observe each point of the Earth s surface twice, first in a forward swath with a range of variation of zenith angle between 52.4º and 55º and secondly, after 120 seconds, a nadir swath with zenith angles from 0º to 21.6º. The instantaneous fields-of-view (IFOV) at the centre of the swath are 1 km by 1 km in the nadir view and 1.5 km by 2.0 km in the forward view. The AATSR was initially design to provide sea surface temperature (SST) maps, but nowadays it is being used to obtain LST on a global scale. LST products require of additional considerations from the SST products. There is a higher heterogeneity over the land surface, besides, land surface emissivity is much variable than sea surface one. 2 ALGORITHM DEVELOPMENT The structure of the theoretical algorithms has been obtained from the radiative transference equation, at sensor considering the at-sensor radiance ( L ) for a given wavelength () as: at sensor L ε B(, T ) + (1 ε ) L atm atm = [ ] S τ + L where ε is the surface emissivity, B (, Ts) is, according to Planck s Law, the radiance emitted by a blackbody atm (BB) at temperature Ts of the surface, L = (1-τ iθ ) atm B i (T a ) is the downwelling radiance, L = (1- τ i53 )B i (T a ) is the upwelling atmospheric radiance, τ is the total transmission of the atmosphere (transmissivity), Ta is the mean temperature of the atmosphere between the surface and the highest level where the information (1) Proc. Envisat Symposium 2007, Montreux, Switzerland 23 27 April 2007 (ESA SP-636, July 2007)
comes from and τ i53 is the total atmospheric path transmittance at 53 degrees. All these magnitudes depend on the observation angle. From Eq. 1 an algorithm involving temperatures can be obtained using a first-order Taylor series expansion of the Planck s law and writing the equation for i and j (i and j being two different channels observed at the same angle, SW method, or the same channel with two different observation angles, DA method): T s = T i +A(T i -T j )-B 0 +(1-ε i )B 1 -Δε θ B 2 (2) where A and Bi are coefficients that depend on atmospheric transmittances, Ti and Tj are the radiometric temperatures for two different channels with the same view angle, SW method, or for the same channel with two different view angles, DA method, in accordance with [5]. In order to intercompare both dual-angle and splitwindow algorithms in the retrieving of surface temperature, it has been used the same mathematical structure for both methods. Reference [5] showed that the transmissivity of 11μm channel with a view angle of 53 degrees practically coincide with the transmissivity of the 12 μm channel at nadir view. So, a split-window algorithm using the 11 μm and 12 μm channels at nadir view is equivalent, in terms of atmospheric correction, to a dual-angle algorithm using the 11 μm channels at nadir and 53 degrees view angles [6]. The determination of the dual-angle and split-window coefficients has been made using MODTRAN simulations for 60 different radiosoundings extracted from the TOVS initial guess retrieval (TIGR) data base and 27 different emissivities, representative of the 90% of the Earth s landcover, obtained from [7] according to [5]. The method used to minimize the objective function is the Levenberg-Marquardt method. Error theory has been applied to all of the algorithms studied. The errors considered are: the residual atmospheric error, mod, which gives an idea of the accuracy in the ST determination; the noise error ( noise ) in the measurement process of the sensor assuming a noise temperature of 0.05 K for the AATSR channels (it should be noted that noise depends on the atmospheric water vapour content, a value of W=1 g cm -2 has been considered); the error associated with the water vapour column determination ( W ), considering a water vapour content uncertainty of 0.5 g cm -2, that error has a dependence with (T 2n -T 1n ), ε n, and Δε θ (to evaluate this error, we have taken for Ts the mean value of the database and we have chosen some representative values from the complete database for ε n and Δε θ ); the error associated with the uncertainty in the value of the emissivity ( ε ) is set at 0.005. The total error has been calculated considering the different errors. Table 1 shows the algorithms that provides the lower theoretical uncertainty values. Table 1. Numerical coefficients and errors for the Split-window and Dual-angle algorithms proposed. NAME ALGORITHM SW1: quad T s = T 11n + 0.61(T 11n -T 12n ) + 0.31(T 11n -T 12n ) 2 + 1.92 1.73 0.07 1.73 SW2: quad, ε T s = T 11n + 0.76(T 11n -T 12n ) + 0.30(T 11n -T 12n ) 2 + 0.10 + 51.2(1-ε) 1.39 0.07 0.18 1.40 SW3: quad, ε, Δε SW4: (W), ε, Δε, W SW5: quad, ε, Δε, W SW6: quad(w), ε, Δε, W T s = T 11n + 1.03(T 11n -T 12n ) + 0.26(T 11n -T 12n ) 2 0.11 + 45.23(1-ε) 79.95Δε T s = T 11n + (1.01+ 0.53W)(T 11n -T 12n ) + (0.4-0.85W) + (63.4-7.01W)(1-ε) - (111-17.6W) Δε T s = T 11n + 1.35(T 11n -T 12n ) + 0.22(T 11n -T 12n ) 2 (0.82-0.15W) + (62.6-7.2W)(1-Δε) - (144-26.3W) Δε T s = T 11n + (1.97+0.2W)(T 11n -T 12n ) - (0.26-0.08W)(T 11n -T 12n ) 2 + (0.02-0.67W) + (64.5-7.35W)(1-ε) - (119-20.4W) Δε mod noise ε WV total 1.05 0.09 0.59 1.20 0.59 0.10 0.83 0.45 1.12 0.93 0.11 1.06 0.20 1.43 0.52 0.15 0.89 0.37 1.10 DA1: quad T s = T 11n + 1.36(T 11n -T 11f ) + 0.18(T 11n -T 11f ) 2 + 1.78 1.31 0.11 1.32 DA2: quad, ε T s = T 11n + 1.56(T 11n -T 11f ) + 0.15(T 11n -T 11f ) 2-0.34 + 51.9(1- ε 11n ) 0.72 0.12 0.18 0.75 DA3: quad, ε, Δε DA4: (W), ε, Δε, W DA5: quad, ε, Δε, W DA6: quad(w), ε, Δε, W T s = T 11n + 1.57(T 11n -T 11f ) + 0.15(T 11n -T 11f ) 2 0.11 + 51.7(1- ε 11n ) 25.8Δε θ 0.69 0.13 0.26 0.74 T s = T 11n + (1.62+0.3W)(T 11n -T 11f ) + (0.18-0.52W) + (70.1 7.18W)(1-ε 11n ) - (35.4-3.67W) Δε θ 0.47 0.13 0.35 0.36 0.70 T s = T 11n + 1.92(T 11n -T 11f ) + 0.12(T 11n -T 11f ) 2 (0.39+0.09W) + (71-7.55W)(1-ε 11n ) - (35.8-3.88W) Δε θ 0.57 0.15 0.36 0.17 0.71 T s = T 11n + (2.67-0.07W)(T 11n -T 11f ) - (0.29-0.09W)(T 11n -T 11f ) 2 - (0.31+0.28W) + (72.5-7.9W)(1-ε 11n ) - (35.8 4.1W) Δε θ 0.38 0.20 0.37 0.24 0.62
The algorithms are ranked according to their explicit dependence on linear difference of brightness temperature (Ti-Tj), quadratical difference of brightness temperature (quad, (Ti-Tj) 2 ), water vapour content (W), emissivity (ε) and spectral or angular emissivity difference (Δε). Table 1 shows that the error of the model is smaller when the algorithm has more degrees of freedom. Besides, dual angle algorithms give better accuracy than split window ones with the same mathematical structure. Moreover, water vapour dependent algorithms give better results than the other ones, even after including the effect of uncertainty in water vapour content error. The results are quite similar for both dual-angle and split-window models when the simplest algorithm (less input parameters) is considered, however the differences increase when increasing the input parameters (see, for instance, algorithm type 6, where SW error doubles the DA one). 3 VALIDATION 3.1 Study Area In order to validate the AATSR LST algorithms, it has been obtained in situ measurements for surfaces with some requirements of heterogeneity and low topography. So, to achieve this validation, two different field campaigns have been made; the first one was selected near Marrakech, in Morocco, in the framework of the WATERMED project. This site was divided into three different areas. A large plot of bare soil, a mixed of vegetation and bare soil and a vegetated area of barley composed of several plots with a quite good homogeneity. The second site chosen for data collection was located in Barrax in the south of Spain 20 km away from Albacete city in the framework of the EAGLE and SPARC projects. The area around Barrax has been used for agricultural research for many years and is characterised by a flat morphology and large, uniform landuse units. Differences in elevation range up to 2 m only. Also in the Barrax area, plots of 3 different samples (Bare Soil, non-green/senescent vegetation and green vegetation) were selected to carry out the validation process. 3.2 Measurements In order to obtain an average radiometric temperature of every location, a set of transects were carried out with different thermal radiometers: CIMEL CE312 radiometers, EVEREST 3000 transducers and RAYTEK MID radiometers. The radiometric temperatures of each transect within 15 minutes of the satellite overpass were examined and average and standard deviation values were calculated. The emissivity of the locations was obtained by the Emissivity Box Method. The downwelling radiance was also measured during the satellite overpass. Table 2 and 3 summarize the land surface temperatures and emissivities for each class considered in both sites, and also the water vapour content measured with a sun photometric instrumentation. Table 2. Averaged values of the surface temperature obtained from transect measurements. Campaign Date Surface temperature Overpass (UTC) Non-green/senescent Green Vegetation Bare Soil vegetation Marrakech 5 th March 2003 10:53 36.0 ± 1.5 34.1 ± 2.7 22.5 ± 1.4 Barrax 14 th July 2003 10:37 49.9 ± 1.4 45.1 ± 2.9 28.8 ± 1.2 Barrax 17 th July 2004 10:36 45.1 ± 1.5 32.9 ± 0.8 24.0 ± 0.4 Barrax 20 th July 2004 10:42 48.8 ± 2.0 39.5 ± 2.2 32.5 ± 0.6 Day Table 3. Atmospheric parameters and emissivities measured during the field campaigns. W (gcm -2 ) 05/03/2003 1.11 Emisivity Non-green / senescent Bare Soil Green Vegetation vegetation (μm) nadir forward nadir forward nadir forward 12 0.979 ± 0.006 0.956 ± 0.007 0.979 ± 0.008 0.968 ± 0.009 0.983 ± 0.007 0.970 ± 0.008 11 0.964 ± 0.007 0.950 ± 0.008 0.968 ± 0.010 0.957 ± 0.010 0.975 ± 0.007 0.962 ± 0.008 12 0.973 ± 0.006 0.959 ± 0.008 0.984 ± 0.008 0.973 ± 0.009 0.995 ± 0.007 0.982 ± 0.008 14/07/2003 2.20 11 0.970 ± 0.006 0.956 ± 0.008 0.980 ± 0.008 0.968 ± 0.009 0.990 ± 0.007 0.977 ± 0.008 17/07/2004 2.36 12 0.970 ± 0.006 0.956 ± 0.008 0.984 ± 0.008 0.973 ± 0.009 0.995 ± 0.007 0.982 ± 0.008 20/07/2004 2.30 11 0.965 ± 0.006 0.951 ± 0.008 0.980 ± 0.008 0.969 ± 0.009 0.990 ± 0.007 0.977 ± 0.008
3.3 Classification method Level1b AATSR products of the studied areas were acquired to test the LST obtained with SW and DA algorithms. The study area embraces an extent of 4 by 4 pixels in the AATSR image, in the case of Marrakech and a more extense area of 6x6 and 7x5 pixels in the case of Barrax 2003 and 2004, respectively. The pixel size in the nadir view is 1km by 1km. In order to apply the proposed algorithms, previously it is necessary to carry out a process of classification of the different sites that the AATSR pixels are made up of, as well as to achieve a statistic analysis of the proportion of every of this sites, with their particular values of temperature and emissivity [8]. With this aim in mind, images of higher spatial resolution than AATSR images have been acquired. A Landsat 5 image of 15 th march 2003 for the study area of Marrakech, with a spatial resolution of 30m in the visible bands; and two CHRIS/PROBA (Compact High Resolution Imaging Spectrometer / Project for On Board Autonomy) images of 14 th july 2003 and 16 th july 2004 for Barrax site, with a spatial resolution of 36 m. CHRIS operates in 63 spectral bands over the visible/near infrared band. Its images are acquired at along track angles of 55 degrees, 36 degrees and near nadir angles. Thus, for each one of the AATSR pixels studied, it can be overstruck a minimum set of 1100 Landsat pixels (a polygon of about 33 pixels by side) and a set of 784 CHRIS pixels (a polygon of about 28 pixels by side). These high resolution images have been classified through a supervised maximum likelihood classification method, taking 3 different classes (bare soil, green vegetation and non-green/senescent vegetation) as training endmembers, in order to know the proportion of each of the sites in the AATSR pixels. A statistical analysis has been carried out to obtain the proportion of the reference areas in every AATSR pixel. Figure 1 shows the classification over the Landsat 5 and CHRIS images. The in situ LST is obtained from the surface radiance by inversion of Plank s Law using the radiometric temperatures of each transect (considering from Eq. 1 τ =1 and L atm =0) and they were also corrected of the effects of the atmosphere and the emissivity by using the down-welling radiance and the values of the emissivity of every sample. These LST values have been compared (see Table 4) with the obtained from AATSR data from all the algorithms proposed in Table 1. Table 4 shows the standard deviation, mean value of the differences (bias) and the root mean square error () between LST from both SW and DA algorithms and in situ data, for both field campaigns, considering the individual pixels. This data show a better behaviour from the Marrakech campaign than from the values of the Barrax campaign. Barrax area is more heterogeneous than the Marrakech area due to the more variability of crops, and this heterogeneity has an impact in the retrieval of LST. Fig. 1. Classification images of Marrakech (a) and Barrax 2004 (b). The images have been classified taking as training endmembers: bare soil (in red), green vegetation (in green) and non-green vegetaton (in blue) Besides, Table 5 shows the results of the application of the SW and DA algorithms in all the experimental campaigns, but now, in order to minimize the Point Spread Function and the co-registration errors, a window filter of 2 x 2 pixels has been applied to the nadir image. The results show that this averaging process ameliorates the in both campaigns, especially for the Barrax campaign in 20 th july 2004.
Table 4. Validation of SW algorithms proposed considering individual pixels. Campaign Marrakech, 5 march 2003 Barrax, 14 july 2003 Barrax, 17 july 2004 Barrax, 20 july 2004 Algorithm SW 1: quad -2.2 1.2 2.5 1.8 2.8 3.3 0.4 2.7 2.7 0.1 2.7 2.7 SW 2: quad, ε -2.5 1.2 2.8 1.6 2.3 2.8 0.0 2.5 2.5-0.3 2.6 2.7 SW 3: quad, ε, Δε -1.8 1.2 2.1 2.0 2.3 3.0 0.5 2.5 2.6 0.2 2.7 2.7 SW 4: (W), ε, Δε, W -1.5 1.2 1.9 1.1 2.2 2.5 0.3 2.5 2.5 0.0 2.6 2.6 SW 5: quad, ε, Δε, W -1.3 1.2 1.8 2.3 2.3 3.2 0.7 2.5 2.6 0.4 2.7 2.7 SW 6: quad(w), ε, Δε, W -1.1 1.2 1.6 0.8 2.2 2.3 0.3 2.5 2.5 0.0 2.5 2.5 DA 1: quad -1.2 3.6 3.8 4.9 6.7 8.3 0.9 5.8 5.9-0.2 9.7 9.7 DA 2: quad, ε -1.6 3.8 4.1 4.5 5.2 6.9 0.7 6.1 6.1-1.3 10.1 10.2 DA 3: quad, ε, Δε -1.5 3.8 4.1 4.5 5.2 6.9 0.6 6.1 6.2-1.5 10.2 10.3 DA 4: (W), ε, Δε, W -1.6 3.8 4.1 3.1 4.2 5.3-0.1 5.4 5.4-4.3 11.4 12.1 DA 5: quad, ε, Δε, W -1.2 4.3 4.4 4.9 5.4 7.3 0.7 6.4 6.4-2.0 11.2 11.4 DA 6: quad(w), ε, Δε, W -1.8 4.1 4.4 2.0 3.3 3.8-0.4 4.7 4.7-5.4 11.3 12.5 Table 5. Validation of SW algorithms proposed considering a window filter of 2 x 2 pixels. Campaign Marrakech, 5 march 2003 Barrax, 14 july 2003 Barrax, 17 july 2004 Barrax, 20 july 2004 Algorithm SW 1: quad -1.2 1.0 1.5 1.6 1.2 2.0 1.1 2.2 2.4 0.0 1.1 1.1 SW 2: quad, ε -1.5 0.9 1.8 1.4 1.2 1.9 0.7 2.3 2.4-0.5 1.3 1.4 SW 3: quad, ε, Δε -0.7 0.9 1.1 1.8 1.2 2.2 1.2 2.3 2.6 0.0 1.3 1.3 SW 4: (W), ε, Δε, W -0.4 0.9 1.0 1.0 1.1 1.5 1.0 2.1 2.3-0.1 1.2 1.2 SW 5: quad, ε, Δε, W -0.3 0.9 0.9 2.1 1.2 2.4 1.5 2.3 2.7 0.3 1.3 1.3 SW 6: quad(w), ε, Δε, W -0.1 0.9 0.9 0.6 1.0 1.2 0.9 2.0 2.2-0.2 1.1 1.1 DA 1: quad 1.7 3.1 3.6 5.3 3.7 6.5-0.5 4.9 5.0-3.5 4.5 5.7 DA 2: quad, ε 1.5 3.3 3.6 4.7 3.7 5.9-1.3 5.1 5.3-4.6 5.0 6.8 DA 3: quad, ε, Δε 1.6 3.3 3.7 4.6 3.7 5.9-1.4 5.1 5.3-4.7 5.1 6.9 DA 4: (W), ε, Δε, W 1.5 2.9 3.2 3.6 2.9 4.6-1.8 5.1 5.4-6.4 5.9 8.7 DA 5: quad, ε, Δε, W 2.1 3.6 4.2 5.1 3.7 6.4-1.4 5.4 5.6-5.3 5.8 7.8 DA 6: quad(w), ε, Δε, W 1.5 0.8 1.7 2.5 2.2 3.4-1.8 4.6 5.0-6.6 5.8 8.8 The algorithms obtained by DA method incorporate the angular effects due to variation of the emissivity, water vapour content and radiometric temperature with the view angle. But this method supposed that the surface observed at different angles is the same, an assumption that is not always reliable. In the AATSR images, nadir pixels have a nominal size of 1.0 km x 1.0 km; instead of this, forward pixels have a size of 1.5 km x 2.0 km (according to the sensor specifications). In this way, when AATSR data is processed from Level 0 product to Level 1b product, pixels are re-gridded onto a regular grid of a nominal pixel size of 1 x 1 km 2, so, values of some pixels are allocated according to values of neighbour pixels. This effect would be negligible when evaluating DA algorithms in homogeneous surfaces. However, in heterogeneous surfaces, it is of principal importance to know the area of each pixel. Additionally, temporal variation of LST from the forward acquisition of the AATSR data to nadir one could be another source of uncertainty in the LST retrieval using DA algorithms. 4 COMPARISON LEVEL 2 AND ALGORITHMS LST values have been compared from the SW algorithms and the provided by the Level 2 product from ESA. LST-Level 2 product is retrieved using a SW algorithm that combines the 11 μm and 12 μm spectral bands. The emissivity is obtained through a global classification according a set of biomes. This SW algorithms is dependent with seasonal values of the fractional vegetation cover and the precipitable water with a spatial resolution of 1ºx1º and 2.5ºx2.5º, respectively. According with these parameters, the SW 6 algorithm in nadir view has been selected to compare the LST retrieved with the LST provided by the Level 2 product. Some problems were found in the Product image: not all AATSR pixels of the area were processed
to Level 2 product, only few clusters of 3x3 pixels in every image. The rest of the pixels have been clasified as cloudy. Measurements of the downwelling sky radiances showed a Sky Temperature of less than -49ºC, this demonstrate that the area was not cloudy. Table 6 shows the effective mean values of the 3x3 clusters measured in situ that has been used to validate the LST from satellite. The Level 2 product LST values are shown with also the LST obtained from the SW 6 algorithm proposed using emissivity and water vapour from satellite data and also from field measurements. The comparison between the in situ data and the satellite data shows that the SW algorithm shows a bias of -2.78 and -1.59 in the first case and 0.56 and 1.11 in the second case. In the case of the Level 2 LST values, the differences are greater than using the SW 6 algorithm. Table 6.- Effective LST measured in situ and a comparison with the LST obtained from Level 2 product and SW 6 algorithm using emissivity and water vapour from satellite data and also from field measurements. Data from the field campaigns of Barrax 14th july 2003 and 17th, 20th july 2004. Day / cluster 14a 14b 17a 17b 17c 20 in situ 317.40 317.64 307.30 312.60 313.54 319.74 Level2 312.35 311.68 308.94 313.23 316.15 323.54 In situ - Level2 5.05 5.97-1.64-0.63-2.61-3.80 0.39 4.11 4.13 SW 6 320.17 319.24 306.73 311.16 313.36 318.64 In situ - SW6-2.78-1.60 0.56 1.44 0.18 1.10-0.18 1.65 1.66 SW 6 + ε, W 316.83 316.52 307.29 310.76 315.27 317.84 In situ - SW6 0.57 1.12 0.01 1.84-1.73 1.9 0.62 1.36 1.50 5 CONCLUSIONS A set of dual-angle and split-window algorithms to estimate land surface temperature from AATSR data has been proposed. A supervised maximum likelihood classification method has been used to validate AATSR LST in heterogeneous sites. A thorough comparison using ground truth data shows a better than 1.7 K for the split-window algorithms proposed. When an averaging process is carried out over the AATSR images using a window filter of 2 x 2 pixels, the results are improved from 3.6 to 1.1 K. LST obtained from DA algorithms present a worse accuracy for heterogeneous surfaces due to the process of re-gridded of the forward AATSR data. Thus, further investigation is required to solve this problem in the overlap of AATSR nadir and forward pixels over heterogeneous areas. In general, LST retrieved from Level 2 product shows a worse accuracy over heterogeneous surfaces than other SW algorithms like the proponed. 6 REFERENCES 1. Schmugge, T., French, A., Ritchie, J. C., Rango, A. and Pelgrum, H. (2002). Temperature and emissivity separation from multispectral thermal infrared observations, Remote Sens. Environ., 79, 189-198 2. Prata, A. J. (1993). Land surface temperatures derived from the AVHRR and ATSR, 1 Theory, J. Geophys. Res., 89D9, 16689-16702 3. Sobrino, J.A., Li, Z.-L., Stoll, M.P., and Becker, F. (1994). Improvements in the split-window technique for the land surface temperature determination. IEEE Trans. Geosci. Remote Sens., Vol.32, No.2, 243-253 4. Sobrino, J.A., Li, Z.-L., Stoll, M.P., and Becker, F. (1996). Multi-channel and multi-angle algorithms for estimating sea and land surface temperature with ATSR data. Int. J. Remote Sensing, 17, 2089-2114. 5. Sobrino, J. A., Sòria, G. and Prata, A. J. (2004). Surface temperature retrieval from Along Track Scanning Radiometer 2 data: Algorithms and validation, J. Geophys. Res., 109, D11101, doi:10.1029/2003jd004212. 6. Prata, A.J. (1994). Land surface temperatures derived from the advanced very high resolution radiometer and the along-track scanning radiometer 2. Experimental results and validation of AVHRR algorithms. J. Geophys. Res., vol. 99, no. D6, 13025-13058 7 Salisbury J. W. and D Aria, D. M., (1992). Emissivity of terrestrial materials in the 8-14 mm atmospheric window, Remote Sens. Environ., 42, pp. 83-106. 8. Soria G. and Sobrino, J.A., (2007) ENVISAT/AATSR derived Land Surface Temperature over a heterogeneous region. Remote Sens. Environ. In press.