0 evaluări0% au considerat acest document util (0 voturi)
25 vizualizări16 pagini
Synthetic data spectral cube was used to examine the accuracy of several methods to recover absolute surface reflectance data of terrestrial targets. Even under controlled and ideal conditions, the spectral recovery using HATCH provided differences of up to 40%. Applying EL on radiance data provided a severe difference of more than 200% in areas located outside the calibration target water vapor zone.
Descriere originală:
Titlu original
Quality assessment of several methods to recover surface reflectance using synthetic imaging spectroscopy data.pdf
Synthetic data spectral cube was used to examine the accuracy of several methods to recover absolute surface reflectance data of terrestrial targets. Even under controlled and ideal conditions, the spectral recovery using HATCH provided differences of up to 40%. Applying EL on radiance data provided a severe difference of more than 200% in areas located outside the calibration target water vapor zone.
Synthetic data spectral cube was used to examine the accuracy of several methods to recover absolute surface reflectance data of terrestrial targets. Even under controlled and ideal conditions, the spectral recovery using HATCH provided differences of up to 40%. Applying EL on radiance data provided a severe difference of more than 200% in areas located outside the calibration target water vapor zone.
Quality assessment of several methods to recover surface reflectance
using synthetic imaging spectroscopy data
E. Ben-Dor a, * , B. Kindel b , A.F.H Goetz b a Department of Geography and Human Environment, Tel-Aviv University, P.O Box 39040, Ramat Aviv Tel-Aviv 69978, Israel b Center for Study of the Earth from Space (CSES), Cooperative Institute for Research in Environmental Sciences (CIRES), University of Colorado at Boulder, Campus Box 216, Boulder, CO 80309-0216, USA Received 4 September 2003; received in revised form 27 January 2004; accepted 27 January 2004 Abstract A synthetic data spectral cube that represents at-sensor radiance data of AVIRIS was used to examine the accuracy of several methods to recover absolute surface reflectance data of terrestrial targets. Soil and vegetation targets, selected to represent the images of ground variation and their spectra, were retrieved using HATCH, Empirical Line (EL) and their hybrids methods. After a synthetic radiance data cube was generated, reflectance recovery was carried out and compared with the true (input) reflectance information. It was found that even under controlled and ideal conditions, the spectral recovery using HATCH code provided differences of up to 40%. The EL methods, using the two end-members that represent the scene reduced this difference to about 4%, and in some cases, even to 0.1% It was found that selecting the calibration targets over low water vapor content improved the results. Applying EL on radiance data provided a severe difference of more than 200% in areas located outside the calibration target water vapor zone. Only over similar water vapor zones were the EL methods found to reasonably recover the surface reflectance. Examining the spectral variability in the calibration targets showed that using of spectral features targets with relative spectral similarity is almost as effective as using spectrally featureless targets for the EL process. Applying EL, using external spectral information of possible known targets, revealed a relatively high difference, as compared to the true reflectance data. However, thematic analysis using a SAM classifier proved that even under non-ideal conditions, the EL correction can yield a reasonable spatial mapping capability relative to those obtained under real reflectance domains. It was concluded that EL must be run on reflectance data (generated from absolute based method) over low water vapor zones to provide the most precise reflectance information. Also, it was found that it is not mandatory to select calibration targets that are totally featureless or characterized by low or high albedo response. D 2004 Elsevier Inc. All rights reserved. Keywords: Imaging spectroscopy; Atmospheric correction; Empirical line; Synthetic data; Thematic analysis 1. Introduction The retrieval of precise and accurate spectral reflectance information from Imaging Spectroscopy (IS) data is a key factor in improving remote sensing of the Earth. The absolute reflectance properties of many solid and liquid materials in the VIS-NIR-SWIR region (0.42.5 Am) have a significant correlation with many chemical and physical parameters under laboratory and field conditions (Ben-Dor et al., 1999; Clark, 1999; Gitelson et al., 1999). This correlation is possible, mainly because sensitive spectros- copy can extract weak (but significant) spectral features that add to the few strong absorption features across the VIS- NIR-SWIR, allowing greater confidence in assessing the chemical and physical compositions of matter (Ben-Dor & Banin, 1995). In the field of remote sensing, the IS tech- nique is opening up a new frontier, by promising near- laboratory accuracy of reflectance information for every pixel in an image (Goetz, 1992). This capability enables better remote thematic mapping of the Earth from many perspectives. IS is an advantageous technique because it permits the extraction of a continuous, rather than a step spectrum for each pixel, and provides the means to distin- guish among very similar materials. Several examples support this claim, such as the ability to differentiate between dolomite and calcite (e.g., Ben-Dor et al., 1994) or between kaolin and smectite (e.g., Kruse et al., 1991) or between dry and green leaf spectral features (Gao & Goetz, 1994). Moreover, improvements in IS thematic capability 0034-4257/$ - see front matter D 2004 Elsevier Inc. All rights reserved. doi:10.1016/j.rse.2004.01.014 * Corresponding author. Tel.: +972-3-6423619; fax: +972-3-6414148. E-mail address: bendor@post.tau.ac.il (E. Ben-Dor). www.elsevier.com/locate/rse Remote Sensing of Environment 90 (2004) 389404 are emerging for similar spectral configurations, allowing more spectral information to be gained, as the signal to noise ratio (SNR) increasesinformation previously obscured by the noise. In this regard, Kruse (2002) showed that the Airborne Visible/infrared Imaging Spectrometer (AVIRIS) data of 2001 (SNR around 500) provides more landscape information than the Hyperion data of 2000 (SNR around 50) over the same area. Because IS sensors acquire the data far from the target, the retrieval of near absolute (laboratory-level) reflectance information is rather problematic and not a straightforward issue. Due to the thick atmospheric media, there is signif- icant attenuation along the sunsurfacesensor pathway. Thus the true spectral response of the surface target is heavily contaminated by atmospheric and solar signals, as well as by other factors such as sub-pixel mixture, sensor geometry and optics, adjacency and topographic effects and more. Whereas these factors are effective in all remote sensing domains, in the IS field they play a particularly important role, as they significantly cause misinterpretation of the area in question (Nadeau et al., 2002; Secker et al., 2001). Accordingly, these factors create the challenge to find ways to remove them or account for them properly, in order to recover spectral reflectance as accurately as possible. Needless to say, absolute reflectance properties are essential so that users can exploit all the laboratory-based methods to further improve the thematic classification of IS images and to better detect small features that may provide new information about the study area. Several methods and strategies to account and correct for the above-mentioned contamination factors are known in IS practice (Mustard et al., 2001). They can be divided into two categories: (1) model-based methods and (2) empirical- based methods. The first category uses methods in which the radiance at the sensor is modeled using radiative transfer models and data from detailed atmospheric and sun infor- mation archives (e.g., MODTRAN). In this procedure, field measurement is not required and basic information, such as the site height and location, flight altitude, local visibility and acquisition times, are the only required information. Several physical-based methods dedicated to retrieving reflectance information from IS data have been developed, such as ATREM (Gao et al., 1993), ATCOR (Richter, 1996; Richter & Schlapfer, 2002), ACRON (ACORN, 2001), FLAASH (Adler-Golden et al., 1998), CAM5S (ONeil et al., 1996), and lately, HATCH (Qu et al., 2000), as well as their modifications described by Sanders et al. (2001). Each of the above methods focuses on one or more issues, but all are quite similar in their absolute basis and execution, and provide similar results (e.g., Staenz et al., 2001). The empirical-based methods rely on the scene informa- tion and do not use physical information as in the model- based methods. Empirical-based methods can be divided into two categories: (1) one that relies on the raw scene data without ground reference information and (2) one that relies on the raw scene data together with ground reference information. There are two common approaches in the first category, namely, the Flat Field (FF) method (Goetz & Srivastava, 1985; Roberts et al., 1986) and the Internal Apparent Relative Reflectance (IARR) methods (Kruse et al., 1985). In both, the raw spectral data of each pixel is divided by a reference spectrum (in FF, a homogenous bright target and in IARR, an average scene spectrum). The drawback of these methods is that they are strongly artifact- and scene-dependent. An empirical method, known as the Empirical Line (EL) approach (Conel et al., 1987; Roberts et al., 1985) belongs to the second category. In this method, two (or more) targets in the scene (featureless, homogeneous, with high albedo differences, if possible; e.g., Furby & Campbell, 2001; Moran et al., 2001) are linearly correlated against corresponding field or laboratory spectra. The correlation for each band provides the relation- ship by which the entire scene is corrected, pixel by pixel, channel by channel, for all data. The advantages and disadvantages of each of the above methods are well known and have been discussed and summarized in many studies (e.g., Ben-Dor & Kruse, 1994; Ben-Dor et al., 1994; Moran et al., 2001; Mustard et al., 2001). In physical-based methods, many artifacts can be introduced (Boardman, 1998; Mustard et al., 2001), whereas in relative-based methods, the results are severely affected by artifacts, especially if ground reference infor- mation is not presented. From a survey of the literature, we concluded that more and more users are now moving towards model, rather than empirical-based methods as the former has become more user-friendly (and is installed in most of the common IS software), provide reasonable results and do not require visiting the specific site. The EL method, the most frequently used among the empirical-based methods, provides reasonable results. How- ever, it requires visits to the ground sites and care must be taken in selecting the calibration targets (sometimes more than two targets are neededpreferably spectrally feature- less and homogeneous ones that adequately represent the albedo diversity in the scene). The power of the EL method is that it can be applied to radiance data (Ben-Dor et al., 2002) or even to DN data (Moran et al., 2001) in order to overcome radiometric calibration differences, and thus can be used for orbital sensor calibration procedures (e.g., Brown et al., 2001; Teillet et al., 2001). In practice, methods that combine both model- and relative-based approaches may also be useful (e.g., Boardman, 1998; Goetz et al., 1998; Richter, 1996). With this in mind, a hybrid EL and absolute-based method was first suggested and applied by Clark et al. (1993), who achieved better results when using them in combination than when running them separately. In such a combination approach, one or two ground (or equivalent) spectra (either reflectance or irradiance) are used to boost the absolute-based atmospherically corrected data into more precise domains. The influence of the varying atmosphere conditions and/or the spectral properties E. Ben-Dor et al. / Remote Sensing of Environment 90 (2004) 389404 390 of the boost targets on the EL results still remains uncertain (Ben-Dor & Levin, 2000). Assessing a methods ability to recover reflectance prop- erties for IS data is commonly carried out on a qualitative, rather than quantitative, basis, by visually comparing field and corrected data of limited validation targets (e.g., Ben- Dor & Levin, 2000; Ben-Dor et al., 2002; Vogelmann et al., 2001). Nevertheless, some quantitative assessments are valid, such as the procedure presented by Leprieur et al. (1995), who examined the difference of retrieval reflectance information from AVIRIS data using a absolute-based approach. By changing the atmospheric conditions, they concluded that using a radiative transfer model which accounts for the multiple scattering effect, best recovered the reflectance information. Recently, Alter-Gartenberg and Nolf (2002) applied an end-to-end stochastic simulation to IS simulated data to assess the sensitivity of the model- based method ATCOR (Richter, 1996; Richter & Schlapfer, 2002) in order to retrieve reflectance for land use classifi- cation. Under various conditions (atmospheric, sensor and area), they examined the methods ability and stability to thematically map six terrestrial categories. Green (1998) used simulated AVIRIS data to study the channel response function. He concluded that 510% differences in spectral calibration caused significant, spectrally distinct differences in measured radiance that contains narrow atmosphere and solar absorptions. In another study, Green (2001) examined the IS sensitivity to account for the atmospheric water vapor content versus the accurate knowledge of surface measured objects. They found that the derived surface reflectance is strongly sensitive to the water vapor content in IS measure- ments. Based on the above, it is postulated that simulation techniques provide a perfect environment in which ques- tions regarding the quality of the IS data can be quantita- tively examined and answered under controlled conditions. One example is the MODO interface of Schlapfer (2001) which allows users to create easily with at sensor IS signals of selected scenarios. Nevertheless, to the best of our knowledge, a compre- hensive and systematic study to examine the spectral differ- ences using a full gamut of atmospheric correction methods (starting with those that are model-based and continuing with empirical-based ones and their hybrids) under a soil vegetation mixture environment has never been applied under simulation conditions. Consequently, there is still no exact information about the operational capabilities and the quality of the methods in recovering reflectance information from raw IS data under the above conditions. Therefore, the purpose of this study is to carry out a controlled study of simulated IS data, to examine the ability and accuracy of absolute- and empirical-based methods (and their hybrids), and to retrieve spectral information from IS data. For that purpose, a synthetic data set of AVIRIS at-sensor data was generated, and the spectral information was recovered from the raw AVIRIS data and quantitatively compared to true (input) reflectance information. 2. Materials and methods 2.1. General To perform the quantitative analysis, we selected an AVIRIS sensor configuration, because AVIRIS represents the highest IS capability attainable from both air and space domains. AVIRIS is a Jet Propulsion Laboratory (JPL) imaging scanner sensor consisting of 224 bands across the VIS-NIR-SWIR region (10 nm band width) with a relatively very high signal-to-noise ratio (Green & Pavri, 2001). The AVIRIS is a sensor used by scientists in many fields, such as geology, agriculture, limnology, soil science, and atmo- sphere science (see AVIRIS home page http://makalu.jpl. nasa.gov/aviris.html). A synthetic AVIRIS cube was gener- ated for this study in order to have well defined conditions, and we focused on limited terrestrial and atmospheric spectral information while setting other factors (such as noise and ground effects) to zero. 2.2. Procedure for generating the synthetic AVIRIS cube The synthetic AVIRIS-like cube was generated using the atmospheric radiative transfer code MODTRAN 4.0 (Berk et al., 1999). The code was run 134 times, corresponding to 134 water vapor amounts ranging from 0.5 to 4.5 cm of a column of water vapor (providing an equal interval of 0.03 cm). A wide water vapor range was selected to illustrate extreme situations in the atmosphere and to cover all possible scenarios from arid (0.5 cm) to tropical (4.5 cm) zones. In order to generate the inputs for predicting the at-sensor radiance for each water vapor amount, the code was run three times with constant albedos of 0.0 (the path radiance term), 0.5 and 1.0. Using these results, the two-way transmittance, path radiance, and spherical albedo were generated. With the addition of the surface reflectance and exoatmospheric solar spectrum, the at-sensor radiance can be predicted (see later description). The solar geometry was chosen to replicate an average imaging spectrometer data collection, run with a solar zenith angle of about 45j. The MODTRAN model was set to the mid-latitude summer model, a constant surface elevation (0 km), a visibility of 50 km and a CO 2 mixing ratio of 360 ppm. Although AVIRIS is an airborne sensor that flies in altitudes up to 20 km, the sensor altitude was set to 100 km, in order to account for full atmosphere attenuation, as seen from orbit. The runs were performed with MODTRANs two-stream DISORT and correlated-k algorithm. The surface spectral information for the input to the synthetic data sets was obtained from two ASD-FR field spectra of soil and vegetation, acquired in Morgan County, Colorado. The vegetation and soil spectra were mixed linearly in 1% increments. The MODTRAN and Analytical Spectral Devices, Full Range (ASD-FR) spectra were con- volved to an actual AVIRIS 1999 wavelength file (band E. Ben-Dor et al. / Remote Sensing of Environment 90 (2004) 389404 391 centers and Full Width Half Maximum (FWHM)). The cube was constructed as follows: the 134 samples across the image are the 134 water vapor values; sample one contains 0.5 cm of column water vapor, sample 134 contains 4.5 cm, the top line consists of 100% vegetation, and the bottom line of 99% soil. The data were scaled by 10,000 and then converted to integer, and finally scaled by the AVIRIS scaling factors. Previous to this, all calculations were performed as floating point. The synthetic AVIRIS-like cube was then transferred to HATCH code to retrieve the re- covery surface reflectance. The HATCH code was initial- ized with the proper sensor geometry and surface elevation, using an estimate of 50 km of visibility. No information on water vapor content or spatial distribution was included the model would determine that by itself. The HATCH code was selected as it is not based directly on the MODTRAN code (which was used to create the data) to prevent output input dependency. The wavelength file given to HATCH was identical to that used to perform the spectral resampling on the MOD- TRAN and ASD-FR spectra, so that the synthetic data would have perfect spectral calibration. It should be noted that some slight radiometric differences between the MOD- TRAN model and the HATCH code appear in wavelengths shorter than 800 nm, due to different exoatmospheric solar spectra used in the two codes. The results of the above processing yielded three data sets as follows: (1) a true (input) reflectance cube data set; (2) an at-sensor radiance cube data set; and (3) a HATCH reflectance-recovery cube data set. 2.3. Data manipulation and procedures For both the at-sensor radiance cube data set and the HATCH reflectance-recovery cube data set, certain manip- ulations were applied as follows: an Empirical Line (EL) method using calibration and validation targets from data set 1 as well as another EL method using an external spectral library (more details later). On one of the manipulated data sets, a classification analysis routine, using a Spectral Angle Mapper (SAM) was employed. In the SAM classifier, the similarity of a test spectrum was examined against a reference spectrum by measuring the angular differences between the two spectra in a vector space (Kruse et al., 1993). 3. Results and discussion 3.1. General The following section will review each of the proce- dures which were performed to recover the reflectance from the synthetic radiance data. Table 1 summarizes the relevant treatments and Fig. 1 provides a schematic of the test cases used in this study. From this point on, various experiments will be discussed in the model space shown in Table 1. 3.2. HATCH correction The first step is to perform a model-based method using HATCH code. The reflectance spectra of the end-mem- bers targets, used to build the mixed surface image (soil and vegetation) are shown in Fig. 2a, while in Fig. 2b and c, the radiance spectra of the above end-members (as a product of applying the atmospheric factor to the reflec- tance data) are presented for two water vapor contents: b- low (0.5 cm) and c-high (4.5 cm). As can be seen, the synthetic radiance spectra appear to satisfactorily represent real AVIRIS data, provided from both measurements and in published literature (example: Green & Pavri, 2001). In general, the HATCH code (or other related model-based methods) is applied to radiance data in order to remove the water vapor effects, as well as other features of atmo- spheric attenuation, such as aerosols, molecular scattering, and oxygen and carbon dioxide absorptions. The aim is to extract the reflectance information of the surface for further analysis. We used the synthetic radiance data set to generate a new data cube, called the HATCH spectral recovery data, described in Table 1 as treatment a. The scaled reflectance of four targets, representing the extremes of soil and vegetation mixtures in the synthetic image (100% vegetation and 100% soil) with low and high water vapor content (0.5 and 4.5 cm water vapor, respectively), are presented along with their corresponding true (in- put) reflectance spectra in Fig. 3a (each target was generated by averaging four closed pixels in the synthetic image. Although the HATCH code was run under ideal con- ditions and data sets, without any signal limitation or other interferences (e.g., optical, BRDF, adjacency effect, precise location and measurements of the calibration sites act), it was postulated that the recovered spectra would show variations when compared to the true (input) spectra (espe- cially in the 0.91.4 Am spectral region). Moreover, for higher water vapor levels, more artifacts were found in the data. To quantitatively assess this observation, we applied a ratio manipulation technique in which the HATCH spectrum for each target was divided by its true (input-scaled) spectrum taken from the input reflectance cube. Fig. 4 shows the ratio spectra of the four selected targets presented in Fig. 3a. The noisy regions, located at the position of strong water vapor effects (around 1.4 and 1.9 and 2.5 Am), are marked with horizontal bars and were neither considered nor analyzed further. As can be seen, relatively large differ- ences were encountered outside the area of the marked barsrelative differences of up to 40% (0.1 reflectance unit) in the vegetation targets (A1 and E1;see Fig. 1 for exact location) or 25% (0.05 reflectance unit) for soil targets (A9 and E9; see Fig. 1 for exact location) in the blue-VIS region. It is important to note that the relative spectral values E. Ben-Dor et al. / Remote Sensing of Environment 90 (2004) 389404 392 cannot represent the absolute spectral variation if the albedo of the targets has not been considered. For example, comparing 20% relative difference in low albedo and high targets (e.g., VIS in vegetation or in bright soil) might be the result of small absolute differences in the low albedo target as compared to the high albedo one. Therefore albedo variation should be considered when discussing relative spectral differences. It can be concluded from the above that even under ideal conditions, when an atmospherically based model is used, the best results will not perfectly yield the reflectance information of high and low albedo targets within an average difference of about 20%. In some cases, the difference can be even higher (mostly in the VIS region). This finding is important for the analysis of the VIS region over targets with low signal output (e.g., vegetation and water). Therefore, caution must be exercised in such cases. It is assumed that, in reality, the spectral differences in retrieval reflectance information might be even higher, because other factors not discussed here may come into play. 3.3. EL on HATCH correction As already mentioned in the Introduction, absolute reflectance is essential in order to enable users to exploit the available laboratory-based methods for further improve- ment of the thematic classification of IS images. The Table 1 A summary table for the treatments employed to the synthetic data Treatment Samples in calibration Samples in validation Symbol Description a Non C1, C3, C5, C7, C9 HATCH HATCH corrected data A5, B5, C5, D5, E5 A1, A9, E1, E9 b A1, A9 C1, C3, C5, C7, C9 EL-HATCH-A EL On HATCH corrected data, using A5, B5, C5, D5, E5 calibration targets on low WV zone A2, A8, E2, E8 (0.5 cm) located at line A in Fig. 1 c E1, E9 C1, C3, C5, C7, C9 EL-HATCH-E EL On HATCH corrected data, using A5, B5, C5, D5, E5 calibration targets on low WV zone (4.5 cm) located at line E in Fig. 1 d A1, E9 C1, C3, C5, C7, C9 EL-HATCH-AE EL On HATCH corrected data, using A5, B5, C5, D5, E5 calibration targets on low and high WV zone (0.5 and 4.5 cm) located at lines A and E in Fig. 1 e Non C1, C3, C5, C7, C9 EFFORT EFFORT on HATCH corrected data A5, B5, C5, D5, E5 A2, A8, E2, E8 f A1, A9 C1, C3, C5, C7, C9 EL-RAD-A EL On Radiance data, using calibration A5, B5, C5, D5, E5 targets on low WV zone (0.5 cm) located A2, A8, E2, E8 at line A in Fig. 1 g E1, E9 C1, C3, C5, C7, C9 EL-RAD-E EL On Radiance data, using calibration A5, B5, C5, D5, E5 targets on high WV zone (0.5 cm) A2, A8, E2, E8 located at line A in Fig. 1 h A1, E9 C1, C3, C5, C7, C9 EL-RAD-AE EL On Radiance data, using calibration A5, B5, C5, D5, E5 targets on low and high WV zones (0.5 A2, A8, E2, E8 and 4.5 cm) located at lines A and E in Fig. 1 i A2, A8 A1, E9 EL-HATCH-DIV1 EL on HATCH data using calibration targets located over low WV zone (0.5 cm) with high spectral variation j A4, A6 A1, E9 EL-HATCH-DIV2 EL on HATCH data using calibration targets located over low WV zone (0.5 cm) with low spectral variation k A2, A4 A1, E9 EL-HATCH-DIV3 EL on HATCH data using calibration targets located over low WV zone (0.5 cm) with low spectral diversity dominated by vegetation l A6, A8 A1, E9 EL-HATCH-DIV4 EL on HATCH data using calibration targets located over low WV zone (0.5 cm) with low spectral diversity dominated by soil m A1, A9 A2, A8, E2, E8 EL-HATCH-EXT EL on HATCH data using calibration targets located over low WV zone (0.5 cm) and their spectral representation using external library data base E. Ben-Dor et al. / Remote Sensing of Environment 90 (2004) 389404 393 previous HATCH result suggests that additional efforts are required in order to obtain absolute near laboratory reflec- tance information from IS data. One method suitable for such a purpose is the Empirical Line (EL) correction technique, already referred to. The EL method is basically an empirical approach in which two (or more) separate targets on the ground with well-known field (or laboratory) spectral information are selected. The reflectance and the corresponding non-corrected IS spectra are then used to construct a linear relationship between each spectral chan- nel to account for offset and gain factors that best represent the channel response. These factors are then applied to the uncorrected data on a pixel-by-pixel, channel-by-channel basis, and a corrected EL image is generated. In practice, it is recommended that this approach should be executed after selecting at least two good targets with high and low albedo characteristics with no spectral features or as few as possible. The EL method is widely used, and it frequently produ- ces good thematic results whenever radiometric or atmo- spheric information are not known or appear to be incorrect. In this regard, several studies have already shown that the EL method is able to correct IS data when other methods, including absolutely based models, have failed (e.g., Ben- Dor & Levin, 2000; Ben-Dor et al., 2002). As already pointed out, to the best of our knowledge, the EL capability, accuracy and performance have never yet been examined under controlled and systematic conditions. Therefore, we decided to examine these using the current synthetic data set. For the EL method, we selected pairs of calibration targets that represented different ground features and differ- ent atmospheric conditions. The first EL was run using two calibration targets selected to represent 100% vegetation and 100% soil, situated within the low water vapor zone (0.5 cm) (targets A1 and A9 in Fig. 1, treatment b in Table 1). The second EL was run using the previous vegetation and soil targets, but over areas in the high water vapor zone (4.5 cm) (targets E1 and E9 in Fig. 1, treatment c in Table 1). The third EL run used the previous vegetation and soil targets, but for areas in mixed low and high water vapor zones (0.5 and 4.5 cm) (targets A1 and E9 in Fig. 1, treatment d in Table 1). All of the above EL corrections were carried out on HATCH retrieved reflectance data and are summarized in Table 1. At this stage, the validation targets were selected in the low and high water vapor zones with varying amounts of vegetation and soil at each target (82.5%, 17.5%), using targets A2, E2, A8 and E8, respec- tively (see Fig. 1 and Table 1). Fig. 5a provides the spectra of the targets A2, A8, E2 and E8, after treatment b was applied, along with the corresponding true (input) spectra (2,8) for comparison. As can be seen, qualitatively, the recovery of the reflectance information after applying the above EL treatment was relatively reliable, and if a visual- Fig. 1. A schematic map presenting the position of all validation and calibration targets used in this study in the synthetic image space. Fig. 2. The original true (field) spectra of the two components (soil and vegetation) used to build the synthetic data base (a) and its at-radiance simulated response, after addition of the atmosphere attenuation under two water vapor conditions: b0.5 cm and c4.5 cm. E. Ben-Dor et al. / Remote Sensing of Environment 90 (2004) 389404 394 Fig. 4. The ratio spectra of the targets spectra (RHATCH), presented in Fig. 3, against their corresponding true (field) spectra (Rtrue). Also provided are horizontal bars in the spectral noise regions. Fig. 3. The recovered reflectance and the true (field) spectra of four targets (A1, A9, E1 and E9), representing 100% of vegetation (1) and 100% soil (9), as obtained after running the HATCH code on the synthetic radiometric data under two water vapor conditions (0.5 and 4.5 cm) (a). Also given are the spectra of the above target after running the EFFORT method (b). E. Ben-Dor et al. / Remote Sensing of Environment 90 (2004) 389404 395 ization decision was made, the correction would seem acceptable. In order to quantitatively assess the reflectance recov- ery of each method applied (at this stage, EL on HATCH data; treatments a, b, c and d in Table 1), we selected new validation targets located at the center of the syn- thetic image in both water vapor variation and soil vegetation mixture lines. These are presented in Fig. 1 as row 5 (samples A, B, C, D and E) and line C (samples 1, 3, 5, 7 and 9), respectively. Row 5 represents equal percentages of soil and vegetation (50%) with varying amounts of water vapor, and line C represents a water vapor content of 2.5 cm and varying amounts of vegeta- tion soil. For each of the validation targets, we calculated the deviation from unity under the treatment used against the true (input) ratio spectra, but omitting the noisy channels (at 1.4, 1.9 and 2.4 Am, marked as shaded areas in Fig. 4). The total number of channels for this calculation remained 163 and the Average Sum of Deviation Square (ASDS) parameter was calculated to correct for the deviation from unity according to the following equation: ASDS ARr n 1 2 =n 1 where n is the channel number and Rr n Rt n= Rtr n 2 where Rt n is the value obtained by a given treatment and Rtr n is the input corresponding reflectance value for that channel. (Plotting Rt n for all channels provides the ratio spectrum (also termed difference spectrum) shown in Fig. 4.) Plotting the ASDS value for each validation target and for every studied treatment is given in Fig. 6a. The plots were evaluated for the vegetationsoil mixture line (line C) and for the water vapor variation line (row 5). As can be seen, the ASDS values (and hence the methods difference margin) of all the EL treatments used were much lower than the ASDS values of the stand-alone HATCH correction. Fig. 5. The recovered reflectance spectra of four validation targets (A2, A8, E2 and E8) using the EL method on: (a) the HATCH database and (b) the radiance data base, using two calibration targets situated over a low water vapor zone (treatment b, EL-HATCH-A and treatment f, EL-RAD-A, respectively). The true (field) spectra of each sample are also provided. E. Ben-Dor et al. / Remote Sensing of Environment 90 (2004) 389404 396 Note that the best validation performance was obtained at the position where the calibration and validation targets have similar water vapor contents (EL-HATCH-A treatment b, validation target A5, and EL-HATCH-E, treatment c, validation target E5). This result suggests that the position of the calibration targets (in terms of atmospheric condition) is important for precise correction especially when no model-based method has been previously applied to the data. The more the validation targets were located in similar water vapor zones of the calibration targets, the better the results of the EL correction. The mixed calibration targets (EL-HATCH-AE, treatment d) also provided reasonable and stable performance in both the vegetationsoil mixture and water vapor line directions. In the vegetationsoil line (equal water vapor content and varying soil and vegetation percentages), it was found that the best treatment results [the best results were given] when EL was applied using targets in both low and high water vapor zones (EL-HATCH-AE, treatment d). Overall, the best spectral recovery performance occurred over the lower water vapor zone of both the calibration and validation targets. Calculating the difference in percentages for the lowest correction performance (HATCH [treatment a] in sample C1) from the ASDS data gave a value of < 10%. However, it should be pointed out that this value was an average calculation across all the spectrum channels, whereas in specific wavelengths, the relative difference reached levels of >40% (see Fig. 4 and the previous discussion). The best performance was given by the EL-HATCH-A (treatment b), using target A5 as validation, providing an average difference of <0.001% with a maximum value of 0.1%. One can conclude that any EL-HATCH treatment, i.e., application of the EL approach to HATCH corrected data, using any target in the varying water vapor space, improves the recovery of the reflectance data significantly (from >40% to >0.1% difference), and better performance was achieved in areas where the validation and calibration targets were situated on similar water vapor zones. In general, selecting the calibration targets over low water vapor zones improves the reflectance retrieval significantly. It is interesting to note that a significant increase within the ASDS values occurred in the lower water vapor interval of 1 cm (either going from 0.5 cm (row 5, line A) to 1.5 cm (row 5, line B) or from 4.5 cm (row 5, line E) to 3.5 cm (row 5 line D), see Fig. 6a). In reality, targets locating over 1 cm water vapor in the scene provide more reliable results than those locating over 4 cm. This suggests that precaution is required in using the EL correction in cases where the scenes in questions are characterized by lower water vapor variation . Based on the above, it is strongly recommended to generate a water vapor image prior to any target selection, so that the EL fine corrections can be made. 3.4. The EFFORT technique (on HATCH correction) Artifacts from atmospheric model-based correction meth- ods are well-known and discussed in the literature (see Introduction). The EFFORT technique was developed by Boardman (1998) in order to polish atmospheric model- based artifacts. The EFFORT process is similar to the EL method, in which matching reference spectra to a disturbed data set is carried out to extract slope and intercept param- Fig. 6. The Average Sum of Deviation Square (ASDS) in log space, against the position of the validation targets in the soil vegetation (line C) and the water vapor (row 5) variation lines along the synthetic image space, for treatments that used EL on HATACH data (treatments b, c and d) (a) and for treatments that used radiance data (treatments f, g and h) (b). The ASDS values of both HATCH and EFFORT on the HATCH correction (treatment a and e, respectively) are also given for comparison. E. Ben-Dor et al. / Remote Sensing of Environment 90 (2004) 389404 397 eters. In the EFFORT method, however, with its robust operational approach, no ground reference data are used, and the slope and intercept t parameters are calculated from the data itself. This is done by generating pseudo field spectra by fitting an observed spectrum with a parametric model of Legendre polynomials, optionally augmented with real spectra. A number of spectra are used, spanning the entire albedo range in order to give good leverage for the linear regression process. The data values versus modeled values are fit with a line for every band. The slope and offset of this line are used to correct the apparent reflectance data for the difference features. In the current study, we chose to examine the EFFORT method over three spectral segments free of water vapor attenuation around 1.4 and 1.9 Am, without the end of the SWIR spectral region (the regions selected were 0.41.30, 1.501.74 and 1.992.4 Am). The EFFORT treatment was run on a HATCH corrected data set (it is called treatment e in Table 1). Fig. 3b shows the result of the four selected validation targets used in Fig. 3a (A1, A9, E1 and E9 in HATCH validation) in two water vapor content zones (0.5 and 4.5 cm), along with the true (field) spectra. As can be seen, the EFFORT process yielded polished spectra, but significantly exaggerated some existing features, such as chlorophyll at 0.68 Am and liquid water at 1.32 Am. To quantitatively assess the EFFORT procedure, we applied the ASDS test for each of the calibration sites situated on the water vapor and vegetationsoil lines (line C and row 5). In Fig. 6b, we provide the ASDS values for the validation targets situated on these lines of both the EFFORT and HATCH treatments (other treatments that will be discussed later are also provided). As can be clearly seen, the EFFORT method yielded higher ASDS values (and hence differences) than the HATCH method in both the line C and row 5 directions. High water vapor and high vegetation content produced higher ASDS values in the EFFORT correction, since they consisted of strong original absorption features. Therefore without boosting spectra, the EFFORT proce- dure may be problematic, and it must be seriously consid- ered whether polishing the spectra is worthwhile. This conclusion however is exceptional only under similar con- ditions used in this study, i.e., relatively low spectral variance of soil and vegetation end-memebrs. 3.5. EL on radiance data In the practical world, EL applications are often run on radiance or DN data, rather than on reflectance (model- based corrected) data. Thus, the next step is to investigate the feasibility of applying the EL technique to the original (simulated) radiance data. We applied new treatments (f, g and h) that were similar to treatments b, c and d in all components (calibration and validation targets), but dif- fered in the basic data set used, which was the AVIRIS radiance (and not the AVIRIS HATCH corrected) data set. Fig. 5b presents the spectra of targets A2, A8, E2 and E8, after applying treatment f (see Table 1; EL-RAD-A, in which the calibration targets were selected on line A that represents low water vapor content) and using the corresponding true (input) spectra (2.8). As can be quali- tatively observed, the recovery of the reflectance informa- tion from the radiance data in this stage yielded favorable results only for targets that were located within the cali- bration targets water vapor zone (line A0.5 cm, targets A2 and A8), whereas targets located outside this zone (line E4.5 cm) were distorted by water vapor signals. How- ever, in order to assess this option quantitatively, we evaluated the ASDS test using validation samples on line C and row 5. The treatments are denoted in Table 1 as treatments f, g and h. In Fig. 6b, the ASDS results for the above examination in the validation targets were provided with the HATCH (and EFFORT) values for comparison. It can be clearly seen from this figure that, in contrast to treatments b, c and d (EL on HATCH, Fig. 6a), the majority of the validation points in each treatment fell to highest values than the HATCH result itself. The highest ASDS value in this stage was >100% and at a specific point, even reached a value of 600% (in EL-RAD-AE treatment h). Two exceptions are noticeable, located on the vegetationsoil mixture line at rows A and E. To further examine this result with more targets domains, we enlarged the validation targets enve- lope by selecting two more samples for each of the water vapor zones: A2 and A8 for 0.5 cm and E2 and E8 for 4.5 cm (see Fig. 1 for exact location). Table 2 summarizes the ASDS values of HATCH and the current EL-Radiance treatments, calculated for the above validation targets. (The calibration targets were selected in A (0.5 cm) position (A1, A9 in calibration), E (4.5 cm) position (E1, E9 in calibration) and in mixed positions (AE, 0.5 and 4.5 cm) (A1, E8 in calibration).) From Table 2 it appears that the EL-Radiance treatments provided lowest ASDS values as compared to the corresponding HATCH values at any stage. As in the EL-HATCH corrections the best perfor- mance was obtained when the calibration and validation targets were selected to represent zones have similar/the same water vapor. Therefore, it can be concluded at this stage that targets located along similar water vapor zones of the calibration Table 2 The MS values of several treatment used to recover the reflectance in four selected targets Treatment Sample in validation A2 A8 E2 E8 EL (f/g) a 1.92 10 5 (A1, A9) b 1.34 10 5 (A1, A9) b 2.09 10 5 (E1, E9) b 6.90 10 5 (E1, E9) a EL (h) a 0.0049 (A1, E9) b 0.0965 (A1, E9) b 0.21 (A1, E9) b 0.069 (A1, E9) b HATCH (a) a 0.0014 0.0042 0.0084 0.0041 a Treatment symbol according to Table 1. b Targets in the EL Calibration. E. Ben-Dor et al. / Remote Sensing of Environment 90 (2004) 389404 398 targets were recovered effectively, even when radiance data were used. Outside this zone, significant distortion by water vapor signals occurred. If the area is characterized by a non- homogenous water vapor spatial distribution (e.g,. varying terrain area), then caution must be applied when using the EL method on radiance (or DN) data. It should be pointed out that the best ASDS values in this stage (EL-RAD-A, treatment f, targets on A position and EL-RAD-E treatment g, targets on E position) were still lower than the best corresponding EL-HATCH treatment (EL-HATCH treat- ment a) values. It is interesting to note that there were similar trends in both databases (EL-HATCH and EL-RAD) with regard to the variation of ASDS as the water vapor level changed, i.e., locating the validation and calibration samples under similar water vapor zones increased the retrieval accuracy. In this regard, HATCH better recovered the reflectance than the EL-RAD treatment, outside the water vapor zone of the calibration EL targets. The conclusions to be drawn from this section are: application of the EL approach to radiance data (or DN data as well) can introduce significant differences based on water vapor variation in the scene. To improve the calibra- tion significantly and provide data sets that present almost absolute reflectance values, it is mandatory to use HATCH or any other model-based method. 3.6. EL using spectral variability targets The next stage in the EL examination is to investigate the quality of the chosen calibration targets, in terms of their spectral variability. Based on our previous results, we selected a single (low) water vapor zone area to examine. The area selected was line A (0.5 cm of water vapor), and new calibration targets were set (A2, A4, A6 and A8, see Fig. 1 for exact locations). In this stage, A1 and A9 were used as validation targets for all treatments, whereas other targets were manipulated in pairs as calibration targets (see the following discussion). Fig. 7 presents the true (input) spectra of the above selected calibration targets, showing significant spectral variation, ranging from high (A2) to low (A8) vegetation content. The treatments used in this stage are summarized in Table 1 as: i-EL on the HATCH data set with calibration targets presenting high spectral diversity (A2 and A8 as calibration targets; EL-HATCH- DIV1), j-EL on the HATCH data set with calibration targets presenting low spectral diversity (A4 and A6 as calibration targets; EL-HATCH-DIV2), k-EL on the HATCH data set with calibration targets presenting low spectral diversity with vegetation domination (A2 and A4 as calibration targets; EL-HATCH-DIV3), and l-EL on the HATCH data set with calibration targets presenting low spectral diversity with soil domination (A6 and A8 as calibration targets; EL-HATCH-DIV4). Fig. 8a and b shows the ratio spectra obtained for the two validation sites (A1 and A9, representing 100% vegetation and 100% soil, respectively), at each of the EL calibration targets (treatments) used, along with their calculated ASDS values. In all of the examined treatments, it appeared that the recovery spectrum was better in the soil-dominant rather than in the vegetation-dominant targets. The best result was obtained in treatment i (large spectral differences in the calibration [A2, A8 in calibration]) for recovery of the soil spectrum. Nevertheless, for all of the treatments, the difference was significantly lower than those obtained by the HATCH code alone. Therefore, better EL results can be obtained when the calibration targets maintain high spectral diversity. Howev- er, lower spectral diversity does not affect the results significantly, as occurs when using the HATCH code alone. Thus, if absolute reflectance recovery must be attained, it is recommended to use good targets along the scene to boost the model-based results. Another conclusion is that the selection of spectrally featureless targets for the EL correction is not a mandatory step. 3.7. EL with no field data One of the basic goals of using model-based methods for reflectance recovering is often related to the fact that no ground visit to the area in question is required. Therefore, the classical EL method has a significant limitation, since a visit to the area is required. However, as in the EFFORT technique, users can often select boost spectra obtained from previous data and from existing spectra library to correct the data. To check the feasibility of such a correc- tion, we used the HATCH corrected data, together with the external vegetation and soil target spectra, selected to represent, as much as possible, the end-members in the synthetic image, taken from the ENVI spectral library (jhu_lib; Haplustalf 87P3468). Fig. 9 presents the spectra of the selected library (a) versus the true (input) spectra (b) of the calibration targets (A1 and A9 in Fig. 1, and treatment m in Table 1). It can be seen that the selection of the targets from the spectral library was not precise, but more or less Fig. 7. The true (field) spectra of the four targets, selected for examining EL feasibility under diverse spectral representation. E. Ben-Dor et al. / Remote Sensing of Environment 90 (2004) 389404 399 represented the spectral variations of the two selected end- members. Applying the EL algorithm to the corrected HATCH data, using A1 and A9 as calibration targets, provided a corrected EL reflectance cube termed EL- HATCH-EXT, where validation targets were selected as A2, A4, A6 and A8. Fig. 10a provides the reflectance Fig. 8. The ratio spectra of the validation targets spectra (A1, A9) against the corresponding true (field) spectra in the EL runs on the HATCH database over the low water content zone allows examination of the diverse spectral representation of the calibration targets: (a) A6, A8 represents low spectral diversity with soil dominant (in calibration); (b) A2, A4 represents low spectral diversity with vegetation dominant (in calibration), (c) A4, A6 represents low spectral diversity (in calibration) and (d) A2, A8 represents high spectral diversity (in calibration). The ASDS values for each ratio are also provided. Fig. 9. The spectra end-members of the two main components in the synthetic image, as extracted from an external library (a) and the corresponding true (field) spectra of these targets (b). E. Ben-Dor et al. / Remote Sensing of Environment 90 (2004) 389404 400 spectra of the validation targets, along with the true (field) spectra, while Fig. 10b provides the difference (ratio) spectra of the validation targets. As can be seen, the difference in all cases was significantly higher than any treatment used so far, where a constant shift from unity was encountered (based on different albedo characteristics of the data sets). In real life, users may not know quantitatively which is a good target, and thus have to select calibration targets that are as precise as possible, based on their best knowl- edge. As seen in Fig. 10, the abundance trends in the vegetation soil mixture were reasonably maintained, and thus, we could examine the thematic results of data obtained from such a correction. We, therefore, ran a Fig. 10. The recovered spectra of targets A2, A4, A6 and A8 using EL on the HATCH data (a) and its ratio spectra (b) at the lower water vapor position, using the end-members in Fig. 8a with the true (field) spectra of each target (treatment n). Fig. 11. The abundance images using the SAM classifier and end-members presented in Fig. 9, for the best spectral recovery correction found in this study (EL- HATCH-A, treatment b) and for the worst recovery correction EL-HATCH-EXT treatment m). The true (synthetic) image of the two selected end-members is also provided for comparison (denoted as a, b and c, respectively, on the images). E. Ben-Dor et al. / Remote Sensing of Environment 90 (2004) 389404 401 SAM classifier on both b (the best correction treatment found in this study) and m (the current treatment) data sets (HATCH-EL-A and HATCH-EL-Ex, respectively) using the 100% soil and vegetation targets as end-members (targets A1, A9) for the SAM method (see the Materials and method). Fig. 11ac provides the image abundance of the vegetation and the soil end-members in each of the treatments used (a-EL-HATCH-A, treatment b; b-EL-Ext- A, treatment m), along with the original reflectance image (c-True). It is apparent that a very good agreement exists between the two data sets at most water vapor contents, but in both, some problems arose at relatively high water vapor levels. This observation suggests that even if the wrong targets for calibration are selected, the results may maintain a reasonable representation of the real spectra, and the thematic result might not be harmed by the EL correcting method. Although it is not recommended to do so, these results may explain why, in many cases, EL correction methods do provide reasonable thematic results, although the absolute difference in such a process may be signifi- cantly higher than a reasonable value. We can conclude that absolute reflectance retrieval using external boost targets, may be problematic. However, precise knowledge of the area in question, along with a good spectral library, may reduce difference If appropriate cali- bration targets are selected to represent the spectral variation in the image, then the thematic mapping will be reasonably maintained. 4. Summary and conclusions Even under optimal conditions, the model-based meth- od used to recover spectral reflectance properties can introduce relative differences of more than 40% in several spectral regions. In reality, where other factors are in- volved, this difference may be even higher. The external effects that may increase the difference in recovering reflectance information from radiance data are: electronic noise, ground topography, variation in the sun exoatmo- sphere solar spectra, differences in spectral and radiomet- ric calibration, BDRF and adjacency effects. In addition, problems in the precise location of the ground and image calibration sites may introduce differences in the final results. The EL method was found to be a reliable technique, in which the original model-based difference could be reduced to values lower than 4% especially over areas with narrow range of water vapor content . It was found that application of the EL to model-based correc- tion data was much more effective than applying it to other data, such as the original radiance. This result may support the utilization of the inflight calibration meth- odology and can be helpful in processes where vicarious calibration in operational environment is taking place (e.g., Secker et al., 2001). Use of the EFFORT method to polish atmospherically based artifacts led to an increase in overall spectral differences. Although applying the EL to radiance data can be problematicif the water vapor in the scene is homogenous, then EL correction may yield reasonable results. It was shown that high spectral vari- ation was not significant, targets with less spectral vari- ation were reasonably recovered to their reflectance values. Also, from the results it seems that the selection of featureless targets as calibration sites is not mandatory. We assume that precise representation of the ground targets on the image (both spatially and spectrally) is much more important than selection of featureless targets. In cases where ground reference sampling is not possible, knowledge of the area is a key factor in application of the EL correction with external boost spectra. In this manner, selection of targets that best represent the area is highly recommended. If caution is used, the thematic mapping results seem to be reasonable in all treatment cases. In summary, applying the EL approach to atmo- spherically corrected data using (any) targets located in areas within similar water vapor content can significantly recover the absolute reflectance in IS data. It should be remembered, however, that under real IS environment, as well as other factors which in this study were set to zero do exist. Nevertheless, the current results provide a strong indication of how and where EL methods should be applied and what is the recovery limit using atmospher- ically based models and commercially available polishing techniques. Acknowledgements This work was conducted at CSES/CIRES University of Colorado under a CIRES fellowship visiting program. The authors are also grateful to the anonymous reviewers who helped to sharpen the issue presented and to present the paper better. References ACORNk (2001). Atmospheric correction now. Analytical imaging and geophysics LLC version 3.12. Boulder, CO. Adler-Golden, S., Berk, A., Bernstein, L. S., & Richtsmeier, S. (1998). FlAASH, A MODTRAN4 atmospheric correction package for hyper- spectral data retrievals and simulations. Summaries of the seventh JPL airborne earth science workshop. JPL-Pub., vol. 97-21 (pp. 19). Pasadena, CA: JPL Pub. Alter-Gartenberg, R., & Nolf, S. R. (2002). A sensitivity assessment of terrestrial identification in remote sensing. International Journal of Re- mote Sensing, 23, 825849. Ben-Dor, E., & Banin, A. (1995). Near infrared analysis (NIRA) as a simultaneously method to evaluate spectral featureless constituents in soils. Soil Science, 159, 259269. Ben-Dor, E., Irons, J. A., & Epema, A. (1999). Soil spectroscopy. In A. Rencz (Ed.), Manual of remote sensing (third edition) (pp. 111189). New York: Wiley. Ben-Dor, E., & Kruse, F. A. (1994). The relationship between the sub E. Ben-Dor et al. / Remote Sensing of Environment 90 (2004) 389404 402 spatial subset of GER 63 channel scanner and the quality of the Internal Average Relative Reflectance (IARR) correction technique. Internation- al Journal of Remote Sensing, 3, 683690. Ben-Dor, E., Kruse, F. A., Lefkoff, A. B., & Banin, A. (1994). Comparison of three calibration techniques for the utilization of GER 63 channel scanner data of Makhtesh Ramon, Negev, Israel. Photogrammetric En- gineering and Remote Sensing, 60, 13391354. Ben-Dor, E., & Levin, N. (2000). Determination of surface reflectance from raw hyperspectral data without simultaneous ground truth measurements. A case study of the GER 63-channel sensor data acquired over Naan, Israel. International Journal of Remote Sensing, 21, 20532074. Ben-Dor, E., Patkin, K., Banin, A., & Karnieli, A. (2002). Mapping of several soil properties using DAIS-7915 hyperspectral scanner data A case study over clayey soils in Israel. International Journal of Remote Sensing, 23, 10431062. Berk, A., Anderson, G. P., Bernstein, L. S., Acharya, P. K., Dothe, H., Matthew, M. W., Adler-Golden, S. M., Chetwynd Jr., J. H., Richtsmeier, S. C., Pukall, B., Allred, C. L., Jeong, L. S., & Hoke, M. L. (1999). MODTRAN 4 radiative transfer modeling for atmospheric correction. Summaries of the eighth JPL airborne earth science workshop. JPL- Pub., vol. 99-17 (pp. 2331). Pasadena, CA: JPL Pub. Boardman, J. W. (1998). Post-ATREM polishing of AVIRIS apparent reflec- tance data using EFFORT: A lesson in accuracy versus precision. Sum- maries of the seventh JPL airborne earth science workshop, vol. 1 (p. 53). Pasadena, CA: JPL Pub. Brown, S. W., Johnson, B. C., Johnson, H. W., Yoon, J. J., Butler, R. A., Barnes, S. B., Spyak, P., Thome, K., Zalewski, E., Helmlinger, M., Bruegge, C., Schiller, S., Fedosejevs, G., Gauthier, R., Tsuchida, S., & Machida, S. (2001). Support of the 1997 Lunar Lake, Nevada exper- iment to determine surface reflectance and top-of-atmosphere radiance. Remote Sensing of Environment, 77, 367376. Clark, R. N. (1999). Spectroscopy of rocks and minerals and principle of spectroscopy. In A. Rencz (Ed.), Manual of remote sensing (third edi- tion) (pp. 359). New York: Wiley. Clark, R. N., Swayze, G., Heidebrecht, K. B., Goetz, A. F. H., & Green, R. O. (1993). Comparison of methods for calibrating AVIRIS data to ground reflectance. Proceedings of the fourth annual JPL airborne geoscience workshop, October 2529, 1993, vol. I (pp. 3536). Pasa- dena, CA: JPL Pub. Conel, J. E., Green, R. O., Vane, G., Bruegge, C. J., Alley, R. E., & Curtiss, B. J. (1987). Airborne imaging spectrometer-2: Radiometric spectral characteristics and comparison of ways to compensate for the atmo- sphere. Proceedings of SPIE, 834, 140157. Furby, S. L., & Campbell, N. A. (2001). Calibration images from different dates to like volume digital counts. Remote Sensing of Environment, 77, 186196. Gao, B. -C., & Goetz, A. F. H. (1994). Extraction of dry leaf spectra features from reflectance spectra of green vegetation. Remote Sensing of Environment, 47, 369374. Gao, B. -C., Heidebrecht, K. B., & Goetz, A. F. H. (1993). Derivation of scaled surface reflectance from AVIRIS data. Remote Sensing of Envi- ronment, 44, 145163. Gitelson, A. A., Schalles, J. F., Rundquist, D. C., Schiebe, F. R., & Yacobi, Y. Z. (1999). Comparative reflectance properties of algal cultures with manipulated densities. Journal of Applied Phycology, 11(4), 345354. Goetz, A. F. H. (1992). Imaging spectrometry for Earth observation. Epi- sodes, 15, 114. Goetz, A. F. H., & Srivastava, V. (1985). Mineralogical mapping in the Cuprite mining district, Nevada. Proceedings of the airborne imaging spectrometer data. JPL Publication, vol. 85-41 (pp. 2229). Goetz, A. F. H., Heidebrecht, K., & Kindel, B. (1998). Using ground spectral irradiance for model correction of AVIRIS data. Summaries of the seventh JPL airborne earth science workshop. JPL-Pub., vol. 97-21 (pp. 159168). Pasadena, CA: JPL Pub (97-21) Green, R., & Pavri, B. (2001). AVIRIS inflight calibration experiment measurements, analysis and results in 2000. Proceedings of the tenth JPL airborne earth science workshop. JPL-Pub., vol. 01-21 (pp. 205219). Pasadena, CA: JPL Pub. Green, R. O. (1998). Spectral calibration requirement doe Earth-looking imaging spectrometers in the solar-reflected spectrum. Applied Optics, 37, 683692. Green, R. O. (2001). Atmospheric water vapor sensitivity and compen- sation requirement for Earth-looking imaging spectrometers in the solar-reflected spectrum. Journal of Geophysical Research, 106, 1744317452. Kruse, F. A. (2002). Comparison between AVIRIS and Hyperion hyper- spectral mineral mapping. Proceedings of the 11th JPL airborne earth science workshop, JPL Publication 034 December 2002, Pasadena, CA ( pp. 171180). Kruse, F. A., Lefkoff, A. B., Boardman, J. W., Heidebrecht, K. B., Shapiro, A. T., Barloon, P. J., & Goetz, F. H. A. (1993). The spectral image processing system (SIPS)interactive visualization and analysis of imaging spectrometer data. Remote Sensing of Environment, 44, 145163. Kruse, F. A., Raines, G. I., & Watson, K. (1985). Analytical techniques for extracting geologic information from multichannel airborne spectrora- diometer and airborne imaging spectrometer data. Proceedings of the 4th thematic conference on remote sensing for exploration geology. ERIM ( pp. 309324). Ann Arbor, MI: ERIM. Kruse, F. A., Thiry, M., & Hauff, P. L. (1991). Spectral identification (1.2 2.5 Am) and characterization of Paris basin kaolinite/smectite clays using a field spectrometer. Proceedings of the 5th International Collo- quium, Physical Measurements and Signatures in Remote Sensing, Courchevel, France, vol. I (pp. 181184). Noordwijk, The Nether- lands: ESA Publication ESTEC. Leprieur, C., Carrere, V., & Gu, X. F. (1995). Atmospheric correction and ground reflectance recovery for Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) data: MAC Europe 91. Photogrammetric Engi- neering and Remote Sensing, 61, 12331238. Moran, M. S., Bryant, R., Thome, K., Ni, W., Nouvellon, Y., Gonzalez- Dugo, M. P., Qi, J., & Clarke, T. R. (2001). A refined empirical line approach for reflectance factor retrieval from Landsat-5 TM and Land- sat-7 ETM+ . Remote Sensing of Environment, 78, 7182. Mustard, J. F., Staid, M. I., & Fripp, W. J. (2001). A semi analytical approach to calibration of AVIRIS data to reflectance over water application in a temperate estuary. Remote Sensing of Environment, 74, 335349. Nadeau, C., Neville, R. A., Staenz, K., Owneill, N. T., & Royer, A. (2002). Atmospheric effects on the classification of surface minerals in an arid region using Short-Wave Infrared (SWIR) hyperspectral imagery and spectral unmixing technique. Canadian Journal of Remote Sensing, 28, 738749. ONeil, N.T., Royer, A., & Nguyyen, M.N. (1996). Canadian Advanced Modified 5S (CAM5S). Internal Report, CARTEL-1996-0202 Centre dapplications et de recherches en teledetection (CARTELl), Universitie de Sherbrooke, Sherbrooke, Quebec, Canada. 46 pp. Qu, Z., Goetz, A. F. H., & Heidbrecht, K. B. (2000). High accuracy atmo- sphere correction for hyperspectral data (HATCH). Proceedings of the ninth JPL airborne earth science workshop. JPL-Pub., vol. 00-18 (pp. 373381). Pasadena, CA: JPL Pub. Richter, R. (1996). A spatially adaptive fast atmosphere correction algo- rithm. International Journal of Remote Sensing, 11, 159166. Richter, R., & Schlapfer, D. (2002). Geo-atmospheric processing of airborne imaging spectrometry data: Part 2. Atmospheric/topo- graphic correction. International Journal of Remote Sensing, 23, 26312649. Roberts, D. A., Yamaguchi, Y., & Lyon, R. J. P. (1985). Calibration of airborne imaging spectrometer data to percent reflectance using field spectral measurements. Proceedings of the 19th international sympo- sium on remote sensing of environment ( pp. 295298). Ann Arbor, MI: ERIM. Roberts, D. A., Yamaguchi, Y., & Lyon, R. J. P. (1986). Comparison of E. Ben-Dor et al. / Remote Sensing of Environment 90 (2004) 389404 403 various techniques for calibration of AIS data. Proceedings of the 2nd AIS workshop. JPL Publication, vol. 86-35 (pp. 2130). Pasadena, CA: JPL Pub. Sanders, L. C., Schott, J. R., & Raqueno, R. (2001). AVNIR/SWIR atmo- spheric correction algorithm for hyperspectral imagery with adjacency effects. Remote Sensing of Environment, 78, 252263. Secker, J., Stanez, K., Gouthier, R. P., & Budkewitsch, P. (2001). Vicarious calibrations of airborne hyperspectral sensors in operational environ- ments. Remote Sensing of Environment, 76, 8192. Schlapfer, D. (2001). MODO: An Interface to MODTRAN for the simu- lation of imaging spectrometry at-sensor signals. Proceedings of the 10th annual. JPL airborne earth science workshop ( pp. 343350). Pasadena, USA: JPL. Staenz, K., Nadeau, C., Neville, R. A., & Secker, J. (2001). Effect of radiative transfer codes for mineral mapping using hyperspectral data. Proceedings of the 8th international symposium on physical measure- ments and signatures in remote sensing, France January 812. Teillet, P. M., Barker, J. L., Markham, B. L., Irish, R. R., Fedosejevs, G., & Storey, J. C. (2001). Radiometric cross-calibration of Landsat-7 ETM+ and Landsat-5 T sensors based on tandem data sets. Remote Sensing of Environment, 78, 3954. Vogelmann, J. E., Helder, D., Morfitt, R., Cjoate, M., Merchant, J. W., & Bulley, H. (2001). Effect of Landsat 5 Thematic Mapper and Landsat 7 Enhanced Thematic Mapper Plus radiometric and geometric calibra- tions and corrections on landscape characterization. Remote Sensing of Environment, 78, 5570. E. Ben-Dor et al. / Remote Sensing of Environment 90 (2004) 389404 404