Surface Emissivity Retrieval From Airborne Hyperspectral Scanner Data
180
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 9, NO. 2, MARCH 2012
Surface Emissivity Retrieval From Airborne Hyperspectral Scanner Data: Insights on Atmospheric Correction and Noise Removal Juan C. Jiménez-Muñoz, José A. Sobrino, and Alan R. Gillespie
Abstract—Airborne multispectral imagers have been used in validation campaigns in order to acquire very high spatial resolution data as a benchmark for current or future satellite data. Imagery acquired with such sensors implies specific data processing in relation to view-angle-dependent atmospheric correction and removal or minimization of stripping-based noise. It is necessary to appropriately perform this processing in order to benefit from reference imageries of surface temperature (T ) and emissivity (ε) maps retrieved from thermal infrared data. In particular, ε images generated from T /ε separation algorithms show undesirable noise that jeopardizes their photointerpretation. This letter addresses the following: 1) the removal of view-angle-dependent atmospheric effects by using ratio techniques for deriving atmospheric water vapor content in a pixel-by-pixel basis and atmospheric radiative transfer simulations to construct lookup tables (LUTs) and 2) the removal of image stripping using maximum/minimum noise fraction (MNF) transforms. For this purpose, imagery acquired with the Airborne Hyperspectral Scanner (AHS) sensor has been used. Results show that angular effects in the atmospheric correction can be addressed from AHS-derived water vapor content and LUTs, whereas due to the AHS noise specific characteristics, the MNF transform only removed part of the noise. Index Terms—Airborne Hyperspectral Scanner (AHS), emissivity, minimum noise fraction (MNF), temperature and emissivity separation (TES), thermal infrared (TIR).
I. I NTRODUCTION
L
AND surface emissivity (LSE) provides valuable information about land-cover conditions and change [1], and it is a key variable for mineral mapping [2]. Knowledge of LSE is also required to obtain accurate values of land surface temperature (LST). Currently, the unique multispectral (five-band) surface emissivity products at high spatial resolution (90 m) available for the scientific community are generated in the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Manuscript received February 28, 2011; revised June 6, 2011 and July 28, 2011; accepted July 28, 2011. Date of publication September 12, 2011; date of current version February 8, 2012. This work was supported by the European Space Agency (SEN2FLEX, 3-11291/05/I-EC), European Union (EAGLE, SST3-CT-2003-502057; CEOP-AEGIS, FP7-ENV-2007-1 No. 212921; WATCH, 036946), and Ministerio de Ciencia y Tecnología (EODIX, AYA2008-0595-C04-01). J. C. Jiménez-Muñoz and J. A. Sobrino are with the Global Change Unit, Image Processing Laboratory, University of Valencia, Valencia E-46071, Spain (e-mail: [email protected]; [email protected]). A. R. Gillespie is with the Department of Earth and Space Sciences, University of Washington, Seattle, WA 98195-1310 USA (e-mail: [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/LGRS.2011.2163699
project (standard product AST05) using the temperature and emissivity separation (TES) algorithm [3]. LSE maps at very high spatial resolution may also be obtained from airborne multispectral thermal infrared (TIR) sensors. One such sensor is the Airborne Hyperspectral Scanner (AHS), operated by the Spanish Institute of Aeronautics (INTA) and used in several field campaigns organized by the European Space Agency since 2004. Researchers involved in these field campaigns and also external users are interested in AHSderived LST and LSE products for different purposes but mainly for heat fluxes and evapotranspiration retrievals. Different validation exercises have demonstrated that TES generally performs according to specifications (around 1.5 K for LST and 0.015 for LSE) both in the ASTER case [4] and in the AHS case [5], although validation from AHS data was focused only on LST. However, undesirable noise is observed in the LSE images, which jeopardizes photointerpretation (visual inspection) of these products. It is important to highlight that noise is not introduced by the ASTER-like TES algorithm, since noise is also observed when using other TES algorithms (normalized emissivity method (NEM), alpha residuals, reference channel, etc.) and noise is not observed when using simulated data free of noise [6]. Therefore, the problem lies in the data (because of the instrumental noise) and not in the algorithms. In order to provide high-quality LSE products, it is necessary to properly address both the atmospheric correction and the instrumental noise removal. There is little published information on addressing both TIR noise reduction and atmospheric correction, particularly relative to emissivity retrieval [2], [7]– [10]. In particular, the effect of noise in AHS TIR bands in the retrieval of LST was first addressed in [11], and a preliminary analysis of in-scene atmospheric variability on AHS LST and emissivity retrievals using simulated data was presented in [6]. In this letter, we present a methodology to perform an atmospheric correction of AHS TIR bands on a pixel-by-pixel basis and accounting for the different scan view angles throughout the image in order to retrieve LSE. We also investigate the performance of noise removal techniques based on principal component analysis in order to improve the visual appearance of AHS-derived emissivity products. II. M ETHODOLOGY A. Test Imagery Two test images acquired with the AHS sensor are considered in this letter. The AHS sensor has four ports, covering the
1545-598X/$26.00 © 2011 IEEE
JIMÉNEZ-MUÑOZ et al.: SURFACE EMISSIVITY RETRIEVAL
TABLE I ATMOSPHERIC PARAMETERS O BTAINED F ROM THE L OCAL S OUNDING AND MODTRAN4 FOR THE AHS T EST I MAGES (F IG . 1). T HE VALUES OF τ AND Lu W ERE C ALCULATED F ROM THE G ROUND TO THE AHS A LTITUDE , AND THE VALUES OF Ld W ERE C ALCULATED FOR THE W HOLE ATMOSPHERE . (W V ) T OTAL ATMOSPHERIC WATER VAPOR C ONTENT
181
tracted from an atmospheric sounding balloon launched almost simultaneously with the AHS overpass and the MODTRAN v4 radiative transfer code [13] for both test cases are provided in Table I. B. TES Algorithm LST (Ts ) and LSE (ε) separation from multispectral TIR data relies on the radiative transfer equation, from which the measured radiance at the sensor level (Lsen ) for a certain band i can be written as Lsen,i = (εi Bi (Ts ) + (1 − εi )Ld,i ) τi + Lu,i
(1)
where B(Ts ) is the radiance of a blackbody at temperature Ts computed by Planck’s function, Ld is the downwelling sky radiance (downward hemispheric flux irradiance divided by π), τ is the atmospheric transmissivity, and Lu is the upwelling or path radiance. Atmospheric correction in the TIR spectral range involves the conversion of spectral radiances at sensor level to radiances at surface level (Lsur ) or land-leaving radiances, which, in turn, requires subtraction of Lu and compensation for τ : Lsur = (Lsen − Lu )/τ . Emissivity retrieval from AHS data has been performed using the TES algorithm as developed for ASTER data [3]. The following TES version was used.
Fig. 1. RGB-infrared composition using the AHS data acquired over two test sites. (a) Barrax agricultural area (SEN2FLEX campaign, Spain). (b) Pine forest and sand dunes in Kootwijk (EAGLE campaign, The Netherlands). The box marks the dunes.
visible and near-infrared (VNIR), short-wavelength infrared, midinfrared, and TIR spectral ranges. Port 1 includes 20 VNIR bands (0.4–1 μm, F W HM ∼ 0.03 μm), and port 4 includes 10 TIR bands (8–13 μm, F W HM ∼ 0.5 μm; see Table I). The AHS field of view (FOV) is ±45◦ , and the instantaneous FOV is 2.5 mrad. One AHS image was acquired during the SEN2FLEX field campaign [5], carried out over the agricultural area of Barrax (Albacete, Spain), and the other AHS image was acquired during the EAGLE campaign [12], carried out over a pine forest with sand dunes in Kootwijk (The Netherlands) (see Fig. 1). In terms of emissivity behavior, the Barrax area is characterized by moderate-to-low spectral contrast, whereas the Kootwijk test site includes a very high spectral contrast sample (sand dunes, high content of quartz). The AHS SEN2FLEX image was acquired on July 12, 2005, at 12:21 UTC, and the flight altitude was 1.3 km above ground level (2 km above sea level) with a resulting pixel (ground) size of 3 m at nadir. The AHS EAGLE image was acquired on June 6, 2006, at 11:34 UTC, and the flight altitude was 2.7 km above ground level (similar altitude above sea level) with a pixel (ground) size of 6 m at nadir. The values of the atmospheric parameters ex-
1) Initial guess for emissivity has been set to 0.97. 2) Iterations involved in the NEM module have been set to eight (enough to reach the convergence criterium established in the TES algorithm [3]). 3) The spectral contrast or maximum–minimum difference (MMD) threshold for low-spectral-contrast surfaces has not been applied. 4) AHS TIR bands 72, 73, and 75–79 have been used, with the relationship between εmin and MMD given by εmin = 0.999 − 0.777 × M M D0.815 [5]. Sources of error on LSE retrievals due to TES itself are discussed in [4]. C. Water Vapor Content Retrieval From AHS In order to perform an accurate atmospheric correction on a pixel-by-pixel basis, it is necessary to characterize the atmosphere at the corresponding spatial resolution. Vertical atmospheric profiles for individual pixels are not available when working at very high spatial resolution; however, atmospheric water vapor content (W V ), the main source of atmospheric absorption and emission in the TIR range, can be retrieved using a differential absorption technique provided that appropriate VNIR bands are available that bracket H2 O absorption features. This W V retrieval technique involves ratioing between the radiance at bands within the absorption feature (measurement bands) and an interpolated radiance of bands in its vicinity (reference bands). In the case of the AHS sensor, band 18 (0.948 μm) was selected as the measurement band, whereas bands 15 (0.862 μm) and 20 (1.004 μm) were selected as the reference bands. The differential absorption technique
182
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 9, NO. 2, MARCH 2012
Fig. 2. Total atmospheric water vapor content versus CIBR obtained from AHS VNIR bands for different surface reflectance spectra.
considered in this letter is the continuum interpolated band ratio (CIBR) (e.g., [14]) given by CIBR =
L18 0.394L15 + 0.606L20
is based on principal component analysis, where the signal-tonoise ratio (S/N ) is maximized, providing image components with decreasing noise fraction (or increasing S/N ratio) [15]. Noise removal techniques based on fast Fourier transform were not considered because they only provide good results over homogeneous areas (e.g., sea) and not over heterogeneous land areas. Since AHS noise showed a significant spatial correlation, both across track and along track, other common techniques based on spatial averages (e.g., low-pass filters), are not expected to provide satisfactory results. In this letter, noise removal with the following two variants of the MNF technique is considered. 1) The noisiest components were spatially filtered using a low-pass filter before transforming back to the original coordinate system. 2) The noisiest components were deleted before performing the inverse transformation to recover the original image. In both cases, noise was partially removed from the ε retrievals. Case 2) was more effective on removing part of the noise, so it was adopted in this letter.
(2)
where L15 , L18 , and L20 are the at-sensor spectral radiances for AHS bands 15, 18, and 20, respectively. CIBR only provides ratio values, so it needs to be rescaled to give W V . For this purpose, simulated data generated from MODTRAN4 were employed. Simulations were performed using a standard midlatitude summer atmosphere with 11 different values of W V (scaling factors from 0.5 to 1.5 in steps of 0.1) and 10 different surface reflectance spectra (see the results in Fig. 2). Since the ratio technique has a low sensitivity to angular variations, a W V map dependent on the AHS view angle (θ) can be obtained using a cosine approach W V0 = W Vθ / cos θ. This is essentially scaling by the path length for different view angles. D. Atmospheric Correction Pixel by Pixel In order to perform an atmospheric correction in a pixelby-pixel basis, a lookup table (LUT) was generated including the values of the atmospheric parameters (τ , Lu , and Ld ) for the different AHS TIR bands. These values were generated from MODTRAN runs under different W V scaling factors (from 0.1 to 2.0 in steps of 0.01) applied to the MODTRAN standard atmospheres (midlatitude summer case was selected in this study). Values at each pixel are finally obtained by simple linear interpolation on the LUT. Angular effects because of the different view angles throughout the image were taken into account with the cosine approach applied to the W V retrievals, as commented in the previous section. Another valid option could be the construction of the LUT from MODTRAN runs at different view angles, then using also the AHS scan view image as input for the interpolation in the LUT. E. MNF Transform In order to analyze instrumental noise effects on ε retrievals, the minimum noise fraction (MNF) transform was applied. It
III. R ESULTS AND D ISCUSSION A. Water-Vapor-Derived Retrievals W V maps (images not shown) were obtained from the two test data sets (Fig. 1) using the CBIR method. The mean values (±1 standard deviation) of W V over the test areas were (0.87 ± 0.08) g · cm−2 for Barrax and (1.28 ± 0.14) g · cm−2 for Kootwijk, leading to W V differences (retrieved minus sounding) of 0.12 and 0.23 g · cm−2 , respectively. The atmospheric parameters (τ , Lu , and Ld ) were retrieved using the LUT and the W V image (view angle dependent) as input. The pattern of these images resembles that of the W V parameter (images not shown): τ is the highest at nadir and lower with increasing view angle, going to the sides of the image. For Lu and Ld , which covary, the pattern is inverted. B. TES Results The TES algorithm was applied to at-surface radiances after atmospheric correction made using the following: 1) only one single set of atmospheric parameters over the whole image (Table I) and 2) atmospheric parameters recovered pixel by pixel. The results are provided in Fig. 3 for AHS band 79 (12.35 μm). Fig. 3(a) and (b) clearly shows the electronic striping noise in the emissivity maps, observed also in the other AHS TIR bands, although, for band 79, the effect is more severe, since this band is located near the side of the 8–13-μm window (lower signal) and also instrumental noise is higher for this band. Fig. 3(c) and (d) shows the difference in ε for band 79 when atmospheric correction is done pixel by pixel versus once per image. The spatial distribution of the difference follows an angular pattern according to the AHS scan view angle, as expected. Differences near the borders up to 0.02 for the Barrax test site and up to 0.04 for the Kootwijk test site were found. From analyzing the AHS scan view angle, it appears that angular difference becomes significant above view
JIMÉNEZ-MUÑOZ et al.: SURFACE EMISSIVITY RETRIEVAL
183
Fig. 4. Surface emissivity (ε) retrieved with the TES algorithm for AHS band 79 (12.35 μm) after application of the MNF transform over the Kootwijk test site. Enlargements of the sand-dune area [see Fig. 1(b)] have been selected, corresponding to (a) the original ε retrieval, (b) the ε retrieved after application of MNF to at-sensor radiances, and (c) after application of MNF to the retrieved ε.
Fig. 3. LSE (ε) retrieved with the TES algorithm for AHS band 79 (12.35 μm) when the atmospheric correction is performed pixel by pixel over the (a) Barrax and (b) Kootwijk test sites. ε differences for pixel-by-pixel versus whole-scene correction for (c) Barrax and (d) Kootwijk.
angles of 40◦ in the case of Barrax and 30◦ in the case of Kootwijk. Note that the AHS image over Barrax was acquired at a lower flight altitude than the image over Kootwijk, and atmospheric and angular effects are higher at higher altitude. In terms of LST (not shown in this letter), the differences were up to 1.5 K (Barrax) and 2.0 K (Kootwijk). Note also that noise on emissivity retrievals is still present even if the atmospheric correction is performed pixel by pixel. C. Noise Removal The MNF transform was applied to at-sensor radiances measured in the ten AHS TIR bands. The first three components were selected to apply the inverse transformation and to recover a noise-filtered version of the original signal. Then, atmospheric correction was performed, and the TES algorithm was applied. An MNF transform performed directly on the emissivity retrievals was also considered, selecting in this case only the first two components because the third one showed a significant noise fraction (not observed in the radiance case). Almost negligible differences were observed when using the first three or the first two components in the emissivity case. As an example, the results are shown in Fig. 4 for the sand-dune area. Some information appears to be lost in the MNF applied directly to emissivity, so MNF applied to radiances before the TES retrieval is preferred. Since pixel values can be modified during the MNF transformation (forward and inverse), a comparison between the original and MNF-transformed data was made. The emissivity differences in band 79 (for the whole image) were 0.009 ± 0.016 for the Barrax case and 0.002 ± 0.015 for the Kootwijk case. For the rest of the bands, the differences were even lower. In terms of LST, the differences of 0.2 ± 1.3 K and 0.0 ± 1.4 K were found for the Barrax and Kootwijk test cases, respectively.
Fig. 5. Surface emissivity (ε) retrieved over the sand dunes (Kootwijk) from laboratory (FTIR spectroscopy) and in situ measurements, compared to AHSderived ε made using the TES algorithm. The atmosphere was assumed to be uniform over the scene such that it could be characterized by a single set of atmospheric parameters (AHS-single); alternatively, the atmosphere was characterized pixel by pixel from water vapor values (AHS-W V ), applying an MNF transform to at-sensor radiances (MNF-Lsen ) and applying an MNF transform directly to ε (MNF-emis).
Fig. 5 shows the emissivity spectra extracted over the sanddune site compared to laboratory measurements carried out using Fourier transform infrared (FTIR) spectroscopy at the Jet Propulsion Laboratory and to in situ measurements carried out using a CIMEL CE 312-2 radiometer [16]. It can be observed that values extracted from the AHS-derived emissivities using the TES algorithm are almost equal, independent of the atmospheric correction method (single value or pixel-by-pixel values) and also independent of the MNF transform. Since the sand-dune sample was observed by AHS at 21◦ , angular influences are not expected. Note that the main objective was to remove the noise and not to improve the values or, in other words, to remove the noise with minimal alteration of the values. IV. C ONCLUSION LSE maps at high spatial resolution provide valuable information, but they are affected by undesirable noise that degrades the visual aspect of these products. Instrumental noise in TIR bands is enhanced in the final recoveries of ε (but almost unobserved in the LST product), so efficient noise removal techniques should be applied to registered at-sensor radiances before applying TES algorithms. Atmospheric correction of TIR data is another important step in the final ε retrieval,
184
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 9, NO. 2, MARCH 2012
and it should be addressed in a pixel-by-pixel basis and also accounting for angular effects for sensors with a high FOV. The MNF transform reduced significantly the noise in our test cases, but residual striping is still present because of the correlation among AHS bands. Note that selection of a few number of components in the MNF transform (two or three in our case) could lead to a loss on spectral information, although, in the sand-dune test area used in this letter, the loss in the spectral information was not evidenced (it may be evidenced in hyperspectral systems with narrower TIR bands). This fact may jeopardize somehow the use of emissivity spectra for geological interpretation over other areas. The spectrally and spatially correlated AHS noise (different from the “white” Gaussian noise), which is not constant and should be estimated for each individual image, hinders the application of an automatic and effective noise removal tool over land surfaces. This noise is probably introduced in the signal in the electronics process, but this assumption has neither been confirmed nor modeled by the data provider (INTA). In-scene atmospheric variability had a low impact on the retrievals since test imagery was acquired over small areas (and thus, atmospheric variations were not significant). The only significant differences found were near the image borders where there were high view angles (∼ 45◦ ), which can be considered as a kind of atmospheric variability over the image. The methodology presented in this letter will allow the generation of improved surface emissivity products from AHS data by improving the visual appearance (after partial noise removal) and also improving retrievals over pixels away from the nadir view. Other commonly used noise removal techniques were not considered in this letter, as discussed in Section II-E, although it would be an interesting future analysis. ACKNOWLEDGMENT The authors would like to thank E. de Miguel and I. Carpintero [Spanish Institute of Aeronautics (INTA)] for providing useful information about Airborne Hyperspectral Scanner (AHS) instrumental noise and the European Space Agency and the rest of the INTA team for providing the AHS imagery. R EFERENCES [1] A. N. French, T. J. Schmugge, J. C. Ritchie, A. Hsu, F. Jacob, and K. Ogawa, “Detecting land cover change at the Jornada experimental range, New Mexico with ASTER emissivities,” Remote Sens. Environ., vol. 112, no. 4, pp. 1730–1748, Apr. 2008. [2] R. G. Vaughan, W. M. Calvin, and J. V. Taranik, “SEBASS hyperspectral thermal infrared data: Surface emissivity measurement and mineral mapping,” Remote Sens. Environ., vol. 85, no. 1, pp. 48–63, Apr. 2003.
[3] A. Gillespie, S. Rokugawa, T. Matsunaga, J. S. Cothern, S. Hook, and A. B. Kahle, “A temperature and emissivity separation algorithm for Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) images,” IEEE Trans. Geosci. Remote Sens., vol. 36, no. 4, pp. 1113–1126, Jul. 1998. [4] D. E. Sabol, A. R. Gillespie, E. A. Abott, and G. Yamada, “Field validation of the ASTER temperature–emissivity separation algorithm,” Remote Sens. Environ, vol. 113, no. 11, pp. 2328–2344, Nov. 2009. [5] J. A. Sobrino, J. C. Jiménez-Muñoz, P. J. Zarco-Tejada, G. Sepulcre-Cantó, E. de Miguel, G. Sòria, M. Romaguera, Y. Julien, J. Cuenca, V. Hidalgo, B. Franch, C. Mattar, L. Morales, A. Gillespie, D. Sabol, L. Balick, Z. Su, L. Jia, A. Gieske, W. Timmermans, A. Olioso, F. Nerry, L. Guanter, J. Moreno, and Q. Shen, “Thermal remote sensing from Airborne Hyperspectral Scanner data in the framework of the SPARC and SEN2FLEX projects: An overview,” Hydrol. Earth Syst. Sci., vol. 13, pp. 2031–2037, 2009. [6] J.-C. Jiménez-Muñoz, J. A. Sobrino, and A. R. Gillespie, “Noise in emissivity images obtained from the ASTER temperature and emissivity separation (TES) algorithm: A case study of airborne imagery and implications for ASTER,” in Proc. 3rd Adv. Quantitative Remote Sens., J. A. Sobrino, Ed., 2010, pp. 770–775. [7] Z.-L. Li, F. Becker, M. P. Stoll, and Z. Wan, “Evaluation of six methods for extracting relative emissivity spectra from thermal infrared images,” Remote Sens. Environ., vol. 69, no. 3, pp. 197–214, Sep. 1999. [8] F. Jacob, X. F. Gu, J.-F. Hanocq, N. Tallet, and F. Baret, “Atmospheric corrections of single broadband channel and multidirectional airborne thermal infrared data: Application to the ReSeDA experiment,” Int. J. Remote Sens., vol. 24, no. 16, pp. 3269–3290, 2003. [9] F. Jacob, F. Petitcolin, T. Schmugge, E. Vermote, A. French, and K. Ogawa, “Comparison of land surface emissivity and radiometric temperature derived from MODIS and ASTER sensors,” Remote Sens. Environ., vol. 90, no. 2, pp. 137–152, Mar. 2004. [10] A. R. Gillespie, E. A. Abbott, L. Gilson, G. Hulley, J.-C. Jiménez-Muñoz, and J. A. Sobrino, “Residual errors in ASTER temperature and emissivity standard products AST08 and AST05,” Remote Sens. Environ., to be published. [11] E. de Miguel, R. García, and A. Fernández-Renau, “The effect of noise in AHS thermal bands in the retrieval of pixel temperature,” in Proc. 10th Int. Symp. Phys. Meas. Spectral Signatures Remote Sens.—ISPRS, Davos, Switzerland, 2010, vol. XXXVI, p. 6, Part 7/C50. [12] Z. Su, W. J. Timmermans, C. van der Tol, R. Dost, R. Bianchi, J. A. Gómez, A. House, I. Hajnsek, M. Menenti, V. Magliulo, M. Esposito, R. Haarbrink, F. Bosveld, R. Rothe, H. K. Baltink, Z. Vekerdy, J. A. Sobrino, J. Timmermans, P. van Laake, S. Salama, H. van der Kwast, E. Claassen, A. Stolk, L. Jia, E. Moors, O. Hartogensis, and A. Gillespie, “EAGLE 2006—Multi-purpose, multi-angle and multi-sensor in-situ and airborne campaigns over grassland and forest,” Hydrol. Earth Syst. Sci., vol. 13, pp. 833–845, 2009. [13] A. Beck, G. P. Anderson, P. K. Acharya, J. H. Chetwynd, L. S. Bernstein, E. P. Shettle, M. W. Matthew, and S. M. Adler-Golden, MODTRAN4 User’s Manual. Hanscom AFB, MA: Air Force Res. Lab., 1999. [14] D. Schläpfer, C. C. Borel, J. Keller, and K. I. Itten, “Atmospheric precorrected differential absorption technique to retrieve columnar water vapour,” Remote Sens. Environ., vol. 65, no. 3, pp. 353–366, Sep. 1998. [15] A. A. Green, M. Berman, P. Switzer, and M. D. Craig, “A transformation for ordering multispectral data in terms of image quality with implications for noise removal,” IEEE Trans. Geosci. Remote Sens., vol. 26, no. 1, pp. 65–74, Jan. 1988. [16] J. A. Sobrino, C. Mattar, P. Pardo, J. C. Jiménez-Muñoz, S. J. Hook, A. Baldridge, and R. Ibáñez, “Soil emissivity and reflectance spectra measurements,” Appl. Opt., vol. 48, no. 19, pp. 3664–3670, Jul. 2009.
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 9, NO. 2, MARCH 2012
Surface Emissivity Retrieval From Airborne Hyperspectral Scanner Data: Insights on Atmospheric Correction and Noise Removal Juan C. Jiménez-Muñoz, José A. Sobrino, and Alan R. Gillespie
Abstract—Airborne multispectral imagers have been used in validation campaigns in order to acquire very high spatial resolution data as a benchmark for current or future satellite data. Imagery acquired with such sensors implies specific data processing in relation to view-angle-dependent atmospheric correction and removal or minimization of stripping-based noise. It is necessary to appropriately perform this processing in order to benefit from reference imageries of surface temperature (T ) and emissivity (ε) maps retrieved from thermal infrared data. In particular, ε images generated from T /ε separation algorithms show undesirable noise that jeopardizes their photointerpretation. This letter addresses the following: 1) the removal of view-angle-dependent atmospheric effects by using ratio techniques for deriving atmospheric water vapor content in a pixel-by-pixel basis and atmospheric radiative transfer simulations to construct lookup tables (LUTs) and 2) the removal of image stripping using maximum/minimum noise fraction (MNF) transforms. For this purpose, imagery acquired with the Airborne Hyperspectral Scanner (AHS) sensor has been used. Results show that angular effects in the atmospheric correction can be addressed from AHS-derived water vapor content and LUTs, whereas due to the AHS noise specific characteristics, the MNF transform only removed part of the noise. Index Terms—Airborne Hyperspectral Scanner (AHS), emissivity, minimum noise fraction (MNF), temperature and emissivity separation (TES), thermal infrared (TIR).
I. I NTRODUCTION
L
AND surface emissivity (LSE) provides valuable information about land-cover conditions and change [1], and it is a key variable for mineral mapping [2]. Knowledge of LSE is also required to obtain accurate values of land surface temperature (LST). Currently, the unique multispectral (five-band) surface emissivity products at high spatial resolution (90 m) available for the scientific community are generated in the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Manuscript received February 28, 2011; revised June 6, 2011 and July 28, 2011; accepted July 28, 2011. Date of publication September 12, 2011; date of current version February 8, 2012. This work was supported by the European Space Agency (SEN2FLEX, 3-11291/05/I-EC), European Union (EAGLE, SST3-CT-2003-502057; CEOP-AEGIS, FP7-ENV-2007-1 No. 212921; WATCH, 036946), and Ministerio de Ciencia y Tecnología (EODIX, AYA2008-0595-C04-01). J. C. Jiménez-Muñoz and J. A. Sobrino are with the Global Change Unit, Image Processing Laboratory, University of Valencia, Valencia E-46071, Spain (e-mail: [email protected]; [email protected]). A. R. Gillespie is with the Department of Earth and Space Sciences, University of Washington, Seattle, WA 98195-1310 USA (e-mail: [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/LGRS.2011.2163699
project (standard product AST05) using the temperature and emissivity separation (TES) algorithm [3]. LSE maps at very high spatial resolution may also be obtained from airborne multispectral thermal infrared (TIR) sensors. One such sensor is the Airborne Hyperspectral Scanner (AHS), operated by the Spanish Institute of Aeronautics (INTA) and used in several field campaigns organized by the European Space Agency since 2004. Researchers involved in these field campaigns and also external users are interested in AHSderived LST and LSE products for different purposes but mainly for heat fluxes and evapotranspiration retrievals. Different validation exercises have demonstrated that TES generally performs according to specifications (around 1.5 K for LST and 0.015 for LSE) both in the ASTER case [4] and in the AHS case [5], although validation from AHS data was focused only on LST. However, undesirable noise is observed in the LSE images, which jeopardizes photointerpretation (visual inspection) of these products. It is important to highlight that noise is not introduced by the ASTER-like TES algorithm, since noise is also observed when using other TES algorithms (normalized emissivity method (NEM), alpha residuals, reference channel, etc.) and noise is not observed when using simulated data free of noise [6]. Therefore, the problem lies in the data (because of the instrumental noise) and not in the algorithms. In order to provide high-quality LSE products, it is necessary to properly address both the atmospheric correction and the instrumental noise removal. There is little published information on addressing both TIR noise reduction and atmospheric correction, particularly relative to emissivity retrieval [2], [7]– [10]. In particular, the effect of noise in AHS TIR bands in the retrieval of LST was first addressed in [11], and a preliminary analysis of in-scene atmospheric variability on AHS LST and emissivity retrievals using simulated data was presented in [6]. In this letter, we present a methodology to perform an atmospheric correction of AHS TIR bands on a pixel-by-pixel basis and accounting for the different scan view angles throughout the image in order to retrieve LSE. We also investigate the performance of noise removal techniques based on principal component analysis in order to improve the visual appearance of AHS-derived emissivity products. II. M ETHODOLOGY A. Test Imagery Two test images acquired with the AHS sensor are considered in this letter. The AHS sensor has four ports, covering the
1545-598X/$26.00 © 2011 IEEE
JIMÉNEZ-MUÑOZ et al.: SURFACE EMISSIVITY RETRIEVAL
TABLE I ATMOSPHERIC PARAMETERS O BTAINED F ROM THE L OCAL S OUNDING AND MODTRAN4 FOR THE AHS T EST I MAGES (F IG . 1). T HE VALUES OF τ AND Lu W ERE C ALCULATED F ROM THE G ROUND TO THE AHS A LTITUDE , AND THE VALUES OF Ld W ERE C ALCULATED FOR THE W HOLE ATMOSPHERE . (W V ) T OTAL ATMOSPHERIC WATER VAPOR C ONTENT
181
tracted from an atmospheric sounding balloon launched almost simultaneously with the AHS overpass and the MODTRAN v4 radiative transfer code [13] for both test cases are provided in Table I. B. TES Algorithm LST (Ts ) and LSE (ε) separation from multispectral TIR data relies on the radiative transfer equation, from which the measured radiance at the sensor level (Lsen ) for a certain band i can be written as Lsen,i = (εi Bi (Ts ) + (1 − εi )Ld,i ) τi + Lu,i
(1)
where B(Ts ) is the radiance of a blackbody at temperature Ts computed by Planck’s function, Ld is the downwelling sky radiance (downward hemispheric flux irradiance divided by π), τ is the atmospheric transmissivity, and Lu is the upwelling or path radiance. Atmospheric correction in the TIR spectral range involves the conversion of spectral radiances at sensor level to radiances at surface level (Lsur ) or land-leaving radiances, which, in turn, requires subtraction of Lu and compensation for τ : Lsur = (Lsen − Lu )/τ . Emissivity retrieval from AHS data has been performed using the TES algorithm as developed for ASTER data [3]. The following TES version was used.
Fig. 1. RGB-infrared composition using the AHS data acquired over two test sites. (a) Barrax agricultural area (SEN2FLEX campaign, Spain). (b) Pine forest and sand dunes in Kootwijk (EAGLE campaign, The Netherlands). The box marks the dunes.
visible and near-infrared (VNIR), short-wavelength infrared, midinfrared, and TIR spectral ranges. Port 1 includes 20 VNIR bands (0.4–1 μm, F W HM ∼ 0.03 μm), and port 4 includes 10 TIR bands (8–13 μm, F W HM ∼ 0.5 μm; see Table I). The AHS field of view (FOV) is ±45◦ , and the instantaneous FOV is 2.5 mrad. One AHS image was acquired during the SEN2FLEX field campaign [5], carried out over the agricultural area of Barrax (Albacete, Spain), and the other AHS image was acquired during the EAGLE campaign [12], carried out over a pine forest with sand dunes in Kootwijk (The Netherlands) (see Fig. 1). In terms of emissivity behavior, the Barrax area is characterized by moderate-to-low spectral contrast, whereas the Kootwijk test site includes a very high spectral contrast sample (sand dunes, high content of quartz). The AHS SEN2FLEX image was acquired on July 12, 2005, at 12:21 UTC, and the flight altitude was 1.3 km above ground level (2 km above sea level) with a resulting pixel (ground) size of 3 m at nadir. The AHS EAGLE image was acquired on June 6, 2006, at 11:34 UTC, and the flight altitude was 2.7 km above ground level (similar altitude above sea level) with a pixel (ground) size of 6 m at nadir. The values of the atmospheric parameters ex-
1) Initial guess for emissivity has been set to 0.97. 2) Iterations involved in the NEM module have been set to eight (enough to reach the convergence criterium established in the TES algorithm [3]). 3) The spectral contrast or maximum–minimum difference (MMD) threshold for low-spectral-contrast surfaces has not been applied. 4) AHS TIR bands 72, 73, and 75–79 have been used, with the relationship between εmin and MMD given by εmin = 0.999 − 0.777 × M M D0.815 [5]. Sources of error on LSE retrievals due to TES itself are discussed in [4]. C. Water Vapor Content Retrieval From AHS In order to perform an accurate atmospheric correction on a pixel-by-pixel basis, it is necessary to characterize the atmosphere at the corresponding spatial resolution. Vertical atmospheric profiles for individual pixels are not available when working at very high spatial resolution; however, atmospheric water vapor content (W V ), the main source of atmospheric absorption and emission in the TIR range, can be retrieved using a differential absorption technique provided that appropriate VNIR bands are available that bracket H2 O absorption features. This W V retrieval technique involves ratioing between the radiance at bands within the absorption feature (measurement bands) and an interpolated radiance of bands in its vicinity (reference bands). In the case of the AHS sensor, band 18 (0.948 μm) was selected as the measurement band, whereas bands 15 (0.862 μm) and 20 (1.004 μm) were selected as the reference bands. The differential absorption technique
182
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 9, NO. 2, MARCH 2012
Fig. 2. Total atmospheric water vapor content versus CIBR obtained from AHS VNIR bands for different surface reflectance spectra.
considered in this letter is the continuum interpolated band ratio (CIBR) (e.g., [14]) given by CIBR =
L18 0.394L15 + 0.606L20
is based on principal component analysis, where the signal-tonoise ratio (S/N ) is maximized, providing image components with decreasing noise fraction (or increasing S/N ratio) [15]. Noise removal techniques based on fast Fourier transform were not considered because they only provide good results over homogeneous areas (e.g., sea) and not over heterogeneous land areas. Since AHS noise showed a significant spatial correlation, both across track and along track, other common techniques based on spatial averages (e.g., low-pass filters), are not expected to provide satisfactory results. In this letter, noise removal with the following two variants of the MNF technique is considered. 1) The noisiest components were spatially filtered using a low-pass filter before transforming back to the original coordinate system. 2) The noisiest components were deleted before performing the inverse transformation to recover the original image. In both cases, noise was partially removed from the ε retrievals. Case 2) was more effective on removing part of the noise, so it was adopted in this letter.
(2)
where L15 , L18 , and L20 are the at-sensor spectral radiances for AHS bands 15, 18, and 20, respectively. CIBR only provides ratio values, so it needs to be rescaled to give W V . For this purpose, simulated data generated from MODTRAN4 were employed. Simulations were performed using a standard midlatitude summer atmosphere with 11 different values of W V (scaling factors from 0.5 to 1.5 in steps of 0.1) and 10 different surface reflectance spectra (see the results in Fig. 2). Since the ratio technique has a low sensitivity to angular variations, a W V map dependent on the AHS view angle (θ) can be obtained using a cosine approach W V0 = W Vθ / cos θ. This is essentially scaling by the path length for different view angles. D. Atmospheric Correction Pixel by Pixel In order to perform an atmospheric correction in a pixelby-pixel basis, a lookup table (LUT) was generated including the values of the atmospheric parameters (τ , Lu , and Ld ) for the different AHS TIR bands. These values were generated from MODTRAN runs under different W V scaling factors (from 0.1 to 2.0 in steps of 0.01) applied to the MODTRAN standard atmospheres (midlatitude summer case was selected in this study). Values at each pixel are finally obtained by simple linear interpolation on the LUT. Angular effects because of the different view angles throughout the image were taken into account with the cosine approach applied to the W V retrievals, as commented in the previous section. Another valid option could be the construction of the LUT from MODTRAN runs at different view angles, then using also the AHS scan view image as input for the interpolation in the LUT. E. MNF Transform In order to analyze instrumental noise effects on ε retrievals, the minimum noise fraction (MNF) transform was applied. It
III. R ESULTS AND D ISCUSSION A. Water-Vapor-Derived Retrievals W V maps (images not shown) were obtained from the two test data sets (Fig. 1) using the CBIR method. The mean values (±1 standard deviation) of W V over the test areas were (0.87 ± 0.08) g · cm−2 for Barrax and (1.28 ± 0.14) g · cm−2 for Kootwijk, leading to W V differences (retrieved minus sounding) of 0.12 and 0.23 g · cm−2 , respectively. The atmospheric parameters (τ , Lu , and Ld ) were retrieved using the LUT and the W V image (view angle dependent) as input. The pattern of these images resembles that of the W V parameter (images not shown): τ is the highest at nadir and lower with increasing view angle, going to the sides of the image. For Lu and Ld , which covary, the pattern is inverted. B. TES Results The TES algorithm was applied to at-surface radiances after atmospheric correction made using the following: 1) only one single set of atmospheric parameters over the whole image (Table I) and 2) atmospheric parameters recovered pixel by pixel. The results are provided in Fig. 3 for AHS band 79 (12.35 μm). Fig. 3(a) and (b) clearly shows the electronic striping noise in the emissivity maps, observed also in the other AHS TIR bands, although, for band 79, the effect is more severe, since this band is located near the side of the 8–13-μm window (lower signal) and also instrumental noise is higher for this band. Fig. 3(c) and (d) shows the difference in ε for band 79 when atmospheric correction is done pixel by pixel versus once per image. The spatial distribution of the difference follows an angular pattern according to the AHS scan view angle, as expected. Differences near the borders up to 0.02 for the Barrax test site and up to 0.04 for the Kootwijk test site were found. From analyzing the AHS scan view angle, it appears that angular difference becomes significant above view
JIMÉNEZ-MUÑOZ et al.: SURFACE EMISSIVITY RETRIEVAL
183
Fig. 4. Surface emissivity (ε) retrieved with the TES algorithm for AHS band 79 (12.35 μm) after application of the MNF transform over the Kootwijk test site. Enlargements of the sand-dune area [see Fig. 1(b)] have been selected, corresponding to (a) the original ε retrieval, (b) the ε retrieved after application of MNF to at-sensor radiances, and (c) after application of MNF to the retrieved ε.
Fig. 3. LSE (ε) retrieved with the TES algorithm for AHS band 79 (12.35 μm) when the atmospheric correction is performed pixel by pixel over the (a) Barrax and (b) Kootwijk test sites. ε differences for pixel-by-pixel versus whole-scene correction for (c) Barrax and (d) Kootwijk.
angles of 40◦ in the case of Barrax and 30◦ in the case of Kootwijk. Note that the AHS image over Barrax was acquired at a lower flight altitude than the image over Kootwijk, and atmospheric and angular effects are higher at higher altitude. In terms of LST (not shown in this letter), the differences were up to 1.5 K (Barrax) and 2.0 K (Kootwijk). Note also that noise on emissivity retrievals is still present even if the atmospheric correction is performed pixel by pixel. C. Noise Removal The MNF transform was applied to at-sensor radiances measured in the ten AHS TIR bands. The first three components were selected to apply the inverse transformation and to recover a noise-filtered version of the original signal. Then, atmospheric correction was performed, and the TES algorithm was applied. An MNF transform performed directly on the emissivity retrievals was also considered, selecting in this case only the first two components because the third one showed a significant noise fraction (not observed in the radiance case). Almost negligible differences were observed when using the first three or the first two components in the emissivity case. As an example, the results are shown in Fig. 4 for the sand-dune area. Some information appears to be lost in the MNF applied directly to emissivity, so MNF applied to radiances before the TES retrieval is preferred. Since pixel values can be modified during the MNF transformation (forward and inverse), a comparison between the original and MNF-transformed data was made. The emissivity differences in band 79 (for the whole image) were 0.009 ± 0.016 for the Barrax case and 0.002 ± 0.015 for the Kootwijk case. For the rest of the bands, the differences were even lower. In terms of LST, the differences of 0.2 ± 1.3 K and 0.0 ± 1.4 K were found for the Barrax and Kootwijk test cases, respectively.
Fig. 5. Surface emissivity (ε) retrieved over the sand dunes (Kootwijk) from laboratory (FTIR spectroscopy) and in situ measurements, compared to AHSderived ε made using the TES algorithm. The atmosphere was assumed to be uniform over the scene such that it could be characterized by a single set of atmospheric parameters (AHS-single); alternatively, the atmosphere was characterized pixel by pixel from water vapor values (AHS-W V ), applying an MNF transform to at-sensor radiances (MNF-Lsen ) and applying an MNF transform directly to ε (MNF-emis).
Fig. 5 shows the emissivity spectra extracted over the sanddune site compared to laboratory measurements carried out using Fourier transform infrared (FTIR) spectroscopy at the Jet Propulsion Laboratory and to in situ measurements carried out using a CIMEL CE 312-2 radiometer [16]. It can be observed that values extracted from the AHS-derived emissivities using the TES algorithm are almost equal, independent of the atmospheric correction method (single value or pixel-by-pixel values) and also independent of the MNF transform. Since the sand-dune sample was observed by AHS at 21◦ , angular influences are not expected. Note that the main objective was to remove the noise and not to improve the values or, in other words, to remove the noise with minimal alteration of the values. IV. C ONCLUSION LSE maps at high spatial resolution provide valuable information, but they are affected by undesirable noise that degrades the visual aspect of these products. Instrumental noise in TIR bands is enhanced in the final recoveries of ε (but almost unobserved in the LST product), so efficient noise removal techniques should be applied to registered at-sensor radiances before applying TES algorithms. Atmospheric correction of TIR data is another important step in the final ε retrieval,
184
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 9, NO. 2, MARCH 2012
and it should be addressed in a pixel-by-pixel basis and also accounting for angular effects for sensors with a high FOV. The MNF transform reduced significantly the noise in our test cases, but residual striping is still present because of the correlation among AHS bands. Note that selection of a few number of components in the MNF transform (two or three in our case) could lead to a loss on spectral information, although, in the sand-dune test area used in this letter, the loss in the spectral information was not evidenced (it may be evidenced in hyperspectral systems with narrower TIR bands). This fact may jeopardize somehow the use of emissivity spectra for geological interpretation over other areas. The spectrally and spatially correlated AHS noise (different from the “white” Gaussian noise), which is not constant and should be estimated for each individual image, hinders the application of an automatic and effective noise removal tool over land surfaces. This noise is probably introduced in the signal in the electronics process, but this assumption has neither been confirmed nor modeled by the data provider (INTA). In-scene atmospheric variability had a low impact on the retrievals since test imagery was acquired over small areas (and thus, atmospheric variations were not significant). The only significant differences found were near the image borders where there were high view angles (∼ 45◦ ), which can be considered as a kind of atmospheric variability over the image. The methodology presented in this letter will allow the generation of improved surface emissivity products from AHS data by improving the visual appearance (after partial noise removal) and also improving retrievals over pixels away from the nadir view. Other commonly used noise removal techniques were not considered in this letter, as discussed in Section II-E, although it would be an interesting future analysis. ACKNOWLEDGMENT The authors would like to thank E. de Miguel and I. Carpintero [Spanish Institute of Aeronautics (INTA)] for providing useful information about Airborne Hyperspectral Scanner (AHS) instrumental noise and the European Space Agency and the rest of the INTA team for providing the AHS imagery. R EFERENCES [1] A. N. French, T. J. Schmugge, J. C. Ritchie, A. Hsu, F. Jacob, and K. Ogawa, “Detecting land cover change at the Jornada experimental range, New Mexico with ASTER emissivities,” Remote Sens. Environ., vol. 112, no. 4, pp. 1730–1748, Apr. 2008. [2] R. G. Vaughan, W. M. Calvin, and J. V. Taranik, “SEBASS hyperspectral thermal infrared data: Surface emissivity measurement and mineral mapping,” Remote Sens. Environ., vol. 85, no. 1, pp. 48–63, Apr. 2003.
[3] A. Gillespie, S. Rokugawa, T. Matsunaga, J. S. Cothern, S. Hook, and A. B. Kahle, “A temperature and emissivity separation algorithm for Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) images,” IEEE Trans. Geosci. Remote Sens., vol. 36, no. 4, pp. 1113–1126, Jul. 1998. [4] D. E. Sabol, A. R. Gillespie, E. A. Abott, and G. Yamada, “Field validation of the ASTER temperature–emissivity separation algorithm,” Remote Sens. Environ, vol. 113, no. 11, pp. 2328–2344, Nov. 2009. [5] J. A. Sobrino, J. C. Jiménez-Muñoz, P. J. Zarco-Tejada, G. Sepulcre-Cantó, E. de Miguel, G. Sòria, M. Romaguera, Y. Julien, J. Cuenca, V. Hidalgo, B. Franch, C. Mattar, L. Morales, A. Gillespie, D. Sabol, L. Balick, Z. Su, L. Jia, A. Gieske, W. Timmermans, A. Olioso, F. Nerry, L. Guanter, J. Moreno, and Q. Shen, “Thermal remote sensing from Airborne Hyperspectral Scanner data in the framework of the SPARC and SEN2FLEX projects: An overview,” Hydrol. Earth Syst. Sci., vol. 13, pp. 2031–2037, 2009. [6] J.-C. Jiménez-Muñoz, J. A. Sobrino, and A. R. Gillespie, “Noise in emissivity images obtained from the ASTER temperature and emissivity separation (TES) algorithm: A case study of airborne imagery and implications for ASTER,” in Proc. 3rd Adv. Quantitative Remote Sens., J. A. Sobrino, Ed., 2010, pp. 770–775. [7] Z.-L. Li, F. Becker, M. P. Stoll, and Z. Wan, “Evaluation of six methods for extracting relative emissivity spectra from thermal infrared images,” Remote Sens. Environ., vol. 69, no. 3, pp. 197–214, Sep. 1999. [8] F. Jacob, X. F. Gu, J.-F. Hanocq, N. Tallet, and F. Baret, “Atmospheric corrections of single broadband channel and multidirectional airborne thermal infrared data: Application to the ReSeDA experiment,” Int. J. Remote Sens., vol. 24, no. 16, pp. 3269–3290, 2003. [9] F. Jacob, F. Petitcolin, T. Schmugge, E. Vermote, A. French, and K. Ogawa, “Comparison of land surface emissivity and radiometric temperature derived from MODIS and ASTER sensors,” Remote Sens. Environ., vol. 90, no. 2, pp. 137–152, Mar. 2004. [10] A. R. Gillespie, E. A. Abbott, L. Gilson, G. Hulley, J.-C. Jiménez-Muñoz, and J. A. Sobrino, “Residual errors in ASTER temperature and emissivity standard products AST08 and AST05,” Remote Sens. Environ., to be published. [11] E. de Miguel, R. García, and A. Fernández-Renau, “The effect of noise in AHS thermal bands in the retrieval of pixel temperature,” in Proc. 10th Int. Symp. Phys. Meas. Spectral Signatures Remote Sens.—ISPRS, Davos, Switzerland, 2010, vol. XXXVI, p. 6, Part 7/C50. [12] Z. Su, W. J. Timmermans, C. van der Tol, R. Dost, R. Bianchi, J. A. Gómez, A. House, I. Hajnsek, M. Menenti, V. Magliulo, M. Esposito, R. Haarbrink, F. Bosveld, R. Rothe, H. K. Baltink, Z. Vekerdy, J. A. Sobrino, J. Timmermans, P. van Laake, S. Salama, H. van der Kwast, E. Claassen, A. Stolk, L. Jia, E. Moors, O. Hartogensis, and A. Gillespie, “EAGLE 2006—Multi-purpose, multi-angle and multi-sensor in-situ and airborne campaigns over grassland and forest,” Hydrol. Earth Syst. Sci., vol. 13, pp. 833–845, 2009. [13] A. Beck, G. P. Anderson, P. K. Acharya, J. H. Chetwynd, L. S. Bernstein, E. P. Shettle, M. W. Matthew, and S. M. Adler-Golden, MODTRAN4 User’s Manual. Hanscom AFB, MA: Air Force Res. Lab., 1999. [14] D. Schläpfer, C. C. Borel, J. Keller, and K. I. Itten, “Atmospheric precorrected differential absorption technique to retrieve columnar water vapour,” Remote Sens. Environ., vol. 65, no. 3, pp. 353–366, Sep. 1998. [15] A. A. Green, M. Berman, P. Switzer, and M. D. Craig, “A transformation for ordering multispectral data in terms of image quality with implications for noise removal,” IEEE Trans. Geosci. Remote Sens., vol. 26, no. 1, pp. 65–74, Jan. 1988. [16] J. A. Sobrino, C. Mattar, P. Pardo, J. C. Jiménez-Muñoz, S. J. Hook, A. Baldridge, and R. Ibáñez, “Soil emissivity and reflectance spectra measurements,” Appl. Opt., vol. 48, no. 19, pp. 3664–3670, Jul. 2009.
