DISCRIMINATION OF VEGETATION-IMPERVIOUS SURFACE-SOIL CLASSES IN URBAN ENVIRONMENT USING HYPERSPECTRAL DATA

Transcription

1 DISCRIMINATION OF VEGETATION-IMPERVIOUS SURFACE-SOIL CLASSES IN URBAN ENVIRONMENT USING HYPERSPECTRAL DATA Shailesh Deshpande 1, Arun Inamdar 2, and Harrick Vin 3 1,3 Tata Research Development and Design Centre, 54-B Hadapsar Industrial Estate, Pune, , India ( , {shailesh.deshpande@tcs.com, harrick.vin}@tcs.com) 2 Centre of Studies in Resources Engineering, Indian Institute of Technology, Mumbai, , India ( , abi@iitb.ac.in) Correspondence: shailesh.deshpande@tcs.com ABSTRACT We present, in this paper, discrimination analysis performed for vegetation, impervious surfaces, and soil (VIS) classes in urban environment. We extracted spectral signatures of VIS classes from EO- 1 Hyperion image and used Spectral Angle Mapper (SAM) for subsequent classification. Further, we recorded field signatures of some typical urban materials such as concrete, bitumen, soils etc. Preliminary analysis indicates that spectral resolution of 10 nm is sufficient to differentiate VIS classes at different level of class granularity. Image derived signatures (and field signatures) of VIS classes are separable and show distinct spectral curves at broader level. Image derived signatures show very good classification accuracy for VIS classes (87% average overall accuracy with 98% best and 77% worst accuracies). Inter class confusion between bare soil and stone quarry, and concrete (residential) is evident from results. Keywords: Hyperspectral, EO-1, SAM, Impervious surface INTRODUCTION Impervious surface is an important artefact of changes brought by rapid urbanization. Increase in impervious surface affects water cycle significantly by reducing ground water recharge, increasing quantity of runoff, increasing non-point source pollution (Arnold and Gibbons, 1996) and so on. Urban flash floods have become a routine phenomenon now in many metros. Further, impervious surface has different heat signature as compared to the soil and/or vegetation cover it replaces. Ever increasing need to quantify fluxes of bio-physical parameters within a city area require tractable models for determining proportions of important environmental indicators such as vegetation, impervious surface, and soil (dubbed VIS). Advent of hyperspectral data provides new opportunities in this regard. Discrimination of large variety of manmade materials by comparing target spectrum with reference spectrum using hyperspectral data is possible (because of minute details available in the hyperspectral signature). Irrespective of these advantages, hyperspectral data has not been used as extensively for impervious surfaces as it is used for vegetation or water studies (Weng, 2012). Further, urban composite materials (such as concrete) change their signatures with variation in their raw materials and hence, studies using hyperspectral data require knowledgebase of local materials. Splib06a (Clark et al., 2007), ASTER (Baldridge et al., 2009) spectral libraries available in public domain cover variety of materials including minerals, vegetation, and limited numbers of urban land covers such as roofs and pavements etc. in Unites States of American (USA). Internal Average Relative Reflectance (IAR), Flat Field methods (Clark et al., 2002) are commonly used image based calibration methods that produce spectrum resembling a laboratory spectrum. Such spectra can then be used for target detection using spectral matching techniques such as Spectral Angle Mapper (SAM). IAR has been found to be suitable in arid regions (with less vegetation) for mineral mapping. Suitability of IAR for regions with heterogeneous land covers such as urban area needs further investigations. Objective of this research is to assess suitability of hyperspectral data for separating VIS classes - especially detection of impervious surfaces such as cement concrete, bitumen/asphalt concrete and so on - in Indian urban setting. Secondary goal is to begin the efforts for creating a knowledge base - comprising spectral signatures of local urban materials - in the form of a prototype spectral library.

2 We investigate here both the methods of creating spectral knowledgebase: using satellite imagery, and using field spectrometer. Further we investigate, if image based calibration method such as IAR provides accurate enough signatures for VIS classification of urban area using SAM. Remaining discussion is arranged as follows: Material and Methods section provides detailed description of field measurements, nature of samples and so on. Material and Methods sub section on computer aided image analysis provides details of calibration method for converting radiance values to reflectance values. Results and Discussion presents analysis of field measurements and provides detailed assessment of VIS classification results. MATERIAL AND METHODS Materials We used EO-1 Hyperion images for extracting spectral signatures of urban materials and further for VIS classification. Two Hyperion images on path 147 and row 47 with scene centres N, E and N, E were acquired (USGS, April 22, 2013, USGS, April 27, 2013). Coverage of ~25 km 2 is required for entire Pune city which is achieved by three Hyperion images (each 7.7 km wide). We focused on the western fringes of the city as the area has seen tremendous growth in recent past because of emerging information technology hubs. Scene centres of each strip were selected to have ~23 pixel overlap between two adjacent images. Hyperion sensor records data in 242 channels with 30 m spatial resolution and 10 nm spectral resolution over nm. Only 198 channels of 242 are calibrated (7-57 and ) (Beck, 2003). Further, we removed channels having noise due to water absorption and selected 141 of 242 channels for further processing. Overall, the range of channels removed for various reasons are: 1-8, 58-81, , , , and We used Spectra Vista Corporation s (SVC) GER 1500 field spectrometer for collecting field spectra. GER 1500 records spectral signatures in 512 channels over nm range with 3 nm spectral resolution (GER 1500 user mannual). Signatures recorded on the filed were then downloaded to personal computer (PC) using data acquisition software provided by SVC. Signature files were converted into ASCI files for further processing. Methods Field measurements: All the field measurements were taken within the window of 2 hrs. of local noon. Instrument was held ~1 m away from the body to avoid effect of scattering from clothing. Readings were taken in the direction perpendicular to azimuth (Goetz, 2012). Fresh reference (white plate) readings were taken at beginning of every new set of scans and on minor changes in atmospheric conditions (and whenever in doubt!). Samples within the image and outside the image area selected were most common surfaces occurring in urban area namely (bracket indicates codes used for these material in this paper): Taiwan lawn (tlawn), American lawn (amlawn), soil covered with dry grass (dgrass), soil covered with burnt grass (bgrass), bare soil (soil), pavement blocks (pavb), Bitumen Bound Macadam (bbm), fresh bitumen concrete road (bitroadnew), old bitumen road (bitroadold), road/parking lot concrete (concrete), plain cement concrete (pcc). Visual image interpretation: We used false colour composite (FCC) and/or side by side display of all the channels using multichannel view provided by Multispec (Biehl and Landgrebe, 2002) for visual interpretation of imagery. We used default colour assignments of Multispec for FCC that is channel 50, 27, and 17 assigned to red, blue, and green colours, respectively. Image classification (computer aided): We used IAR method to calibrate the Hyperion data and then used Spectral Angle Measurement (SAM) to classify imagery in VIS classes using image derived spectral signatures. We provide brief account, of procedure to extract spectral signatures from image, in next discussion: Selected channels within the range 1-70 were divided by 40 and were divided by 80 to convert the Digital Numbers (DNs) to radiance (W/m 2 SR µm). After converting DNs to radiance values, spectrum of each pixel within the whole image area was normalized to have a constant sum

3 (sum over all the channels of a given pixel). Normalization removes illumination differences because of topographical (or similar other) factors within the image. Net effect of normalization is; shift in a given spectrum (upwards or downwards) so that all spectra within the scene represent same relative brightness (Kruse, 1988). Further, we remove additive component of atmospheric effects using novel extension of improved dark object technique. Improved dark object technique (Chavez, 1988) assumes that relation between path radiance and wavelength is not the same for all atmospheric conditions. It assumes different atmospheric models (equation determining relation between path radiance and wavelength) for different atmospheric conditions. To begin, user selects the model that best represents the atmospheric condition at the time of data collection and then selects a base path radiance value from a selected channel using histogram (or dark object). In the next step, path radiance for other channels is calculated by scaling the selected path radiance. The scaling factor is determined by the wavelength of the new channel and selected atmospheric model. For example, for simple hypothetical model where path radiance is inversely proportional to wavelength, path radiance of channel µm is 2.67 ({1/.300}/{1/.800}) times path radiance of µm channel. Improved dark object technique has been successfully used for Landsat data (Chavez, 1996). We extend the approach to hyperspectral data; scaling factors for each atmospheric model for all Hyperion channels are calculated using band centres. Other option could have been, using Gaussian model of spectral response and integrating line results for each central wavelength of a given channel within 10 nm width (Full Width Half Maximum). We resorted to former approach for its simplicity and computational efficiency. In the present implementation, we selected very clear atmospheric model (i.e. path radiance is inversely proportional to fourth power of the wavelength) and chose channel 52 ( nm) for selecting initial haze value. We selected 0.5 (W/m 2 SR µm) as a starting haze value to avoid overcorrection of the channels beyond visible range. Correction factors for all wavelengths were calculated for very clear atmospheric condition. Initial haze value is then multiplied by correction factor for each channel to arrive at radiance value that needs to be subtracted from each channel, respectively. We then used average spectrum over the whole scene as a reference and divided each pixel within the image by the reference to calculate relative (to average) reflectance (IAR). Average spectrum, in ideal situations where there is no vegetation or any other dominant material in the scene, indicates absorptions due to atmosphere and can be used as a reference spectrum to calculate the reflectance values. Flat Field (Roberts and Yamaguchi, 1986) and Log Residuals (Green and Craig, 1985) are other image based methods for calculating relative reflectance values from radiance values. IAR produces spectra resembling laboratory spectra and hence can be used to detect the material using spectral matching techniques such as SAM. We then used SAM to detect VIS classes in the acquired images. We provide preliminaries of SAM in following sections. SAM is one of the primary techniques that uses spectral library to classify the pixels in unsupervised manner. The goal is to identify (classify) a pixel by comparing the spectrum of that pixel with the spectrum of known material. If the pixel spectrum matches with a particular reference spectrum of the known material in the library, we can assign that pixel to that particular material class. The matching or closeness of the two spectra is decided using angle between the two spectral vectors with m dimensions where m is number of bands in a spectrum and hence the name. More formally expressed as: Let L = {l i=1, l 2, l 3 l n, } set of library members where each member l 1, l 2,... l n are the vectors indicating a member in the spectral library - say for example, each member represent signature for asphalt, concrete, gravel surface and so on. Each member l 1, l 2,... l n is represented as a vector of reflectance values in each band as l i = {r k=1, r 2, r 3 r m } where m equals number of bands. Similarly, each pixel in the image can be represented as m dimensional vector as p j = {r k=1, r 2, r 3 r m }. Then the cosine of angle between pixel vector and library member is given by: l i. p j l i p j (1)

4 Experimental set up for classification experiments We performed various experiments to study accuracy of the VIS classification, and intra class - inter class confusion if any using SAM. All reference signatures were taken from pure pixels (verified by field trip and/or from high resolution imagery) but few intentional mixture signatures for thematic classes. As a strategy, reference signatures were extracted from small regions and tested on large area. We selected test region (pixels) that are away from reference pixels by large distance margin. All the experiments were performed on the April 22 image. Detail experimental set up is as explained below: Experiment 1-2 (Baseline): We extracted reference signatures from most representative materials of VIS classes. For example vegetation signature is taken from trees, soil signatures from bare soil area with little or low grass cover, and impervious surface signatures are taken from heavy residential area with concrete roof tops. In experiment 2, we change the reference signature of impervious surface to concrete pavement and see if all impervious surface targets are detected. Experiment 3-5: In experiment 3, we extracted reference signatures of the subclasses of impervious surfaces concrete and industrial roof tops and then further investigated (in experiment 4 and 5) if different income zones can be detected spectrally by extracting reference mixture signatures from respective zones. Experiment 6-7: Similar to experiment for impervious surfaces, we successively extracted reference signatures of various subclasses in soil and performed VIS classification. Expr8: We divided vegetation reference signatures into two groups and performed VIS classification. We repeated experiment 1 and 2 with different plain test regions (Experiment 1+, 2+). We provide below description of regions (and their codes) used to extract the reference signatures and to calculate accuracy of classification. All the locations are in and around of Pune city, India: Trees - Urban tree cover (TRLP, TRUN); Farms, green grass etc. - Farms, green grass and other miscellaneous low laying vegetation (GGRF, GGHB, GGHI); Plains - Common open areas with brown soil, typically with varying degree of dry grass cover (PLMU, PLCT, PLCI, PLTA), Open ground - An open ground with compact bare soil without grass cover, like play grounds (OGFC, OGPA), Quarry - A large stone quarry area mainly with exposed basaltic rock surface (QAMO), Dry grass - Open areas covered with dry grass shoots or grass on slopes cut to the ground level (DGKO, DGKA), Forest fire marks - Burned grass pastures and/or meadows at different locations (BGKH, BGCH, BGKO) Concrete - Bright road grade cement concrete (CONP, COCP); Industrial roof - Industrial grade tin roof tops (IRCU, IRTA, IRKI); Residential - A mixture signature of area covered predominantly with concrete slab roofs without little vegetation (REPG, REKR); Residential 2 - A mixture signature of upmarket housing with good tree cover (RUNA, RUMO) We measured accuracy of results for all the experiments using standard confusion matrix and further calculated Producer s and User s accuracy. Producer s accuracy indicate fraction of reference pixels that are correctly identified by classifier (omission errors), and User s accuracy indicate fraction of pixels assigned to a particular class (by classifier) that are true class pixels (commission errors) (Congalton, 1991). For example, Experiment 1 reference signatures for VIS classes are extracted from, GGPA, REPG, and PLMU (Table 1), and accuracy is calculated using TRUN, REKR, and PLCI regions (Table 2), respectively. Overall accuracy of classification for experiment 1 is: 95% with Producer s and User s accuracy of 100% for Vegetation class, User s accuracy of 99% and Producer s accuracy of 93% for Impervious Surface class and so on (Table 3). RESULTS AND DISCUSSION Field signatures analysis Most of the impervious surfaces do not show distinct diagnostic features in the wavelength range of nm and show very similar shape. All the road surfaces like bitumen and concrete differ largely in amplitude (reflectance) values. Only exception to this general observation is pavement blocks (pavb). Pavement blocks show sharp increase in reflectance in the rage nm and show highest reflectance of all the surfaces in nm range. Though there are little differences between road cement concrete and bitumen concrete signatures, concavity of concrete (and pcc) is noticeable. Concrete and plain cement concrete (pcc) show much higher overall reflectance than bitumen surfaces. Bitumen Bound Macadam (bbm) and old bitumen road (bitroadold) show almost identical spectral signature. The difference in shape that is convexity of fresh asphalt road and concavity of old asphalt roads found by Herold et al. (2004) is very subtle and unnoticeable in our observations. It could be

5 because of small range of spectral coverage in our study and/or differences in raw materials used in respective composite materials (Figure 1). Soil signatures change significantly with percentage of green or dry vegetation. Many of the open areas and hill slopes on the fringes of the city are covered with dry grass shoots or grass cut to ground level (at the time of image recording). The dry grass (dgrass) sample signature represents such bright spots. Spectral signature of soil dominated with dry grass more than 90% has distinct shape with continuously increasing reflectance from visible to near infrared range. Soil covered with burnt marks - made by fire used for remineralisation of pastures and meadows shows lowest overall reflectance of all and is very similar to fresh bitumen road surface. Burnt marks (bgrass) show very flat spectral signatures without any diagnostic absorption in the recorded range. Gravel and soil signatures show typical signatures of the classes and are separable using shape information. Two common species of lawns (Local trade names Taiwan Lawn and American Lawn) used for private and public gardening show very similar spectral shape but are separable because of different green, red, and near infrared reflectance peaks (Figure 1). Image classification using SAM General comments The signatures of VIS classes - with 10 nm spectral resolution are sufficient to differentiate and detect them. Preliminary results using image derived signatures show very good classification accuracy for VIS classes (87% overall accuracy). While broad level VIS classes are separable spectrally, some of the subclasses show similar signatures. Road cement concrete and bitumen concrete show very similar signatures with flat reflectance values without any diagnostic absorption. Concrete show higher reflectance values than that from the bitumen road. Though the soil covered bright spots (e.g. play grounds) in the image have different spectral shape, they are at times confused with built-up classes. Spectral matching technique such as NS 3 (Nidamanuri and Zbell, 2011) that take into account reflectance information as well would be helpful in such a scenario. Accuracy of classification is largely determined by the test region than reference region: as an average of all the pixels are taken as a reference while each individual pixel similarity is calculated in classification process (Table 1, 2 and 3). Effect of mixture signature as a reference Reference signatures from pure pixel provide more accurate results (Baseline experiments 1 and 2). There is jump of 3% in accuracy when concrete signature is used to identify residential area (predominantly concrete roof tops). The slight improvement in accuracy is consistent for all further experiments wherein we divide impervious surface and soil surface (Experiment 5, 6, 7, and 8). Further, if reference signature is taken from large area representing mixture signature (of pure and/or mixed pixels), accuracy is not affected significantly (baseline experiments 1 and 2 for both the test regions). Producer s accuracy is also very high for impervious surfaces (above 90%) except cases where two bright signatures namely stone quarry and concrete are considered as references at the same time. Irrespective of high user s accuracy for stone quarry area (100), user s accuracy of 58% is achieved respectively (Experiment 6). Similar confusion occurs when large test region for soil contains some bright spots in it they are confused with bright signatures of concrete (Experiment 2+). Effect of breaking impervious surfaces, soil and vegetation Further considering subclasses of vegetation, impervious surfaces, and soil did not affect accuracy by large margin. Results showed very high overall accuracy with high user s and producer s accuracy with occasional dips for vegetation and soil classes. Vegetation class confusion arises when thematic class (rather than material class) like upmarket residential zone is considered which is mixture of concrete and tree cover pixels. Confusion between soil and impervious surface arises when bright pixels of stone quarry are added in reference signatures. Some of the bright pixels in the residential area comprised of concrete roof tops are misclassified as stone quarry area. Intra class confusion is maximum for trees and resulted in poor producer s and user s accuracy for the class. Water hyacinth is identified accurately with good user s and producer s accuracy of 0.74 and 0.81 respectively. Only two sub classes are considered at present: Trees and Water hyacinth

6 (Experiment 8). Further investigation is required for other vegetation classes such as green grass subtypes. Vegetation class is generally not confused with soil or impervious surface class. In case of large test region (for upmarket residential class) which is composed of concrete roof tops and tree covers, some of the true (possibly) tree pixels are labeled correctly affecting user s accuracy for tree class (Experiment 5). Though producer s accuracy is lowered for thematic class (upmarket residential), at material level these results are still true as test area indeed has lots of tree cover. Among the two impervious material classes considered, industrial roofs and residential roofs (concrete), there is no intra class confusion. Industrial roof tops are identified with 100% user s and producer s accuracy all the time (Experiment 3, 4, 5, 7, and 8). Adding reference spectra of subclasses of vegetation and soil does not affect accuracy of industrial roof top detection (Experiment 5, 7, and 8) the way soil classes affect residential area some time (Experiment 5 and 6). Overall soil classes show moderate intra class and low interclass confusion. Especially, bare soil (like playground) creates intra class confusion and interclass confusion with impervious class like concrete. Intra class confusion lowered the user s accuracy for soil class to 0.78 and producer s accuracy to 0.56 (Experiment 7). Intra class confusion might be because of heterogeneous cover in plain test region which might have some sub areas with less vegetation cover resembling bare soil. This confusion is less harmless and can be avoided by choosing test region with more uniform cover appropriately. While residential class accuracy was unaffected in presence of bare soil reference signatures, stone quarry signatures reduced accuracy of residential classes by large margins (user s accuracy 0.82, and producer s accuracy 0.28). Bright regions in the residential zones are mistaken to be stone quarry. Gross inaccuracy at the thematic level in concrete roofs might be because gravel possibly came from the same stone quarry. More detail study is required to see relation between raw material fraction and spectral signatures especially to asses if gravel is the only dominant factor controlling its signature or not. Irrespective of high producer s accuracy for stone quarry area user s accuracy of 58% is achieved (Experiment 6). Figure 1. Field signatures of impervious surfaces (left subplot), vegetation, and soil covers (right subplot) in urban area, Pune, India Table 1. Region codes of reference signatures for VIS classification using SAM Description Expr. Vegetation Impervious Soil No. Baseline1 1 GGPA REPG PLMU Baseline2 2 GGPA CONP PLMU +Industrial 3 TRUN CONP IRKI PLMU +Residential 4 TRLP REPG RUNA IRTA PLMU Soil test train 5 TRLP REPG RUNA IRTA PLCI +Quarry, Dry grass 6 TRLP CONP PLMU QOMO DGKO +Open grounds, 7 TRLP REPG IRTA PLMU OGFC DGKO BGKO Burnt grass 3+Hyacinth 8 GGPA GGHB CONP IRTA PLMU DGKO

7 Table 2. Area codes used for testing and overall accuracy (OA) Expr. Vegetation Impervious Soil OA No. 1 TRUN REKR PLCI TRUN REPG PLCI GGPA REPG IRTA PLCI TRUN REKR RUMO IRKI PLCI TRUN REKR RUMO IRKI PLMU TRUN REPG PLCI QOMO DGKA TRUN REKR IRKI PLCI OGPA DGKA BGKH TRUN GGHI REPG IRKI PLCI DGKA TRUN REKR PLCT TRUN REPG PLCT 0.77 Table 3. User s accuracy and / Producer s for various experiments with overall accuracy Expr. Vegetation Impervious Soil OA No / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / CONCLUSION Test area should be covered with a particular target material more uniformly especially for soil surfaces accuracy is affected as some of the bright spots are confused with residential areas or vice versa. As reference signature is average of all the pixels in a given area, slight variations in pixels (materials) for reference area do not affect the accuracy of the results. Different types of vegetation covers can be differentiated (for example experiment 8, hyacinth and tree), but further experiments with more vegetation types in urban area are required to understand the possibilities. In presence of bare soil reference spectra, interclass confusion (with concrete or residential) is reduced as bright spots in plain region are mistaken to be open grounds (instead of concrete roofs). Similarly, some of the bright spots in the residential area (concrete roof tops) are mistaken to be surfaces similar to stone quarry area. Mixed signature (of pixels) with known dominant materials (that is confirmed with due supervision) can be effectively used for impervious surfaces in absence of pure material signature. In fact, it reduces confusion resulted, because of bright pure signatures at times. Further, mixed signatures can represent thematic classes, for example, upmarket residential and residential. Techniques that take into account amplitude information like NS 3 would be more useful as spectral shapes of some of the impervious surfaces are similar. In addition, robust spectral features are required to identify thematic classes such as economic zones within the city. Instead of taking reference signature using multiple pixels (pure or mixed), fractions of tree, open area, road network, and roof types within a window of say 10X10 pixels would be more fruitful attributes. This method would be very suitable for unmixing techniques as fractions are sufficient within a given area without attributing a material to a particular pixel. ACKNOLEDGEMENT Shailesh Deshpande would like to thank Sachin Gupte and Priya Deshpande for their timely assistance during the trips conducted for field measurements and the survey of land use land covers for this study. He would like to also thank Samee Azmi at CSRE for providing useful tips for handling field spectrometer and its data acquisition software.

8 REFERENCES [1] Arnold C. L. Jr. and Gibbons C. J., 1996, Impervious Surface Coverage: The Emergence of a Key Environmental Indicator, Journal of the American Planning Association, (62) [2] Baldridge A. M., Hook S. J., Grove C. I., and Rivera G., 2009, The ASTER Spectral Library Version 2.0, Remote Sensing of Environment, (113) [3] Beck R. A., 2003, EO-1 User Guide v. 2.3, Satellite Systems Branch USGS Earth Resources Observation Systems Data Center (EDC), nd Street, Sioux Falls, S.D., Available online: (accessed on 19 th August 2013). [4] Biehl L. and Landgrebe D., 2002, MultiSpec-a Tool for Multispectral-Hyperspectral Image Data Analysis, Computers and Geosciences, (28) [5] Clark R. N., Swayze G. A., Livo K. E., Kokaly R. F., King T. V. V., Dalton J. B., Vance J. S., Rockwell B. W., Hoefen T., and McDougal R. R., 2002, Surface Reflectance Calibration of Terrestrial Imaging Spectroscopy Data: a Tutorial Using AVIRIS, Proceedings of the 10th Airborne Earth Science Workshop, JPL Publication [6] Clark R. N., Swayze G. A., Livo K. E., Hoefen T., Kolkaly R., Sutley S. J., 2007, USGS Digital Spectral Library splib06a: U.S. Geological Survey, Digital Data Series 231, Available online: (accessed on19 th August 2013). [7] Congalton R. G., 1991, A Review of Assessing the Accuracy of Classifications of Remotely Sensed Data, Remote Sensing of Environment, (37) [8] GER 1500 User Manual 3.7 ed., Spectra Vista Corporation, 29 Firemen s Way, Poughkeepsie, New York [9] Goetz A. F. H., 2012, Making Accurate Field Spectral Reflectance Measurements, ASD Incorporated, Boulder, Colorado, 80301, USA, Available online: (accessed on 19 th August 2013). [10] Green A. A. and Craig M. D., 1985, Analysis of Aircraft Spectrometer Data with Logarithmic Residuals, Proceedings of the Airborne Imaging Spectrometer Data Analysis Workshop, JPL Publications [11] Herold M., Roberts D. A., Gardner M. E., and Dennison P. E., 2004, Spectrometry for Urban Area Remote Sensing - Development and Analysis of a Spectral Library from 350 to 2400 nm., Remote Sensing of Environment, (91) [12] Kruse F. A., 1988, Use of Airborne Imaging Spectrometer Data to Map Minerals Associated with Hydrothermally Altered Rocks in the Northern Grapevine Mountains, Nevada, and California, Remote Sensing of Environment, (24) [13] Nidamanuri R. R. and Zbell B., 2011, Normalized Spectral Similarity Score (NS 3 ) as an Efficient Spectral Library Searching Method for Hyperspectral Image Classification, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, (4) [14] Chavez P. S. Jr., 1988, An Improved Dark-Object Subtraction Technique for Atmospheric Scattering Correction of Multispectral Data, Remote Sensing of Environment, (24) [15] Chavez P. S. Jr., 1996, Image-Based Atmospheric Corrections - Revisited and Improved, Photogrammetric Engineering and Remote Sensing, (62) [16] Roberts D. A. and Yamaguchi Y., 1986, Comparison of Various Techniques for Calibration of AIS Data, Second Airborne Imaging Spectrometer Data Analysis Workshop, JPL Publications 86-35, [17] U.S. Geological Survey, 2013, April 22, EO1 Hyperion Scene: EO1H KZ_PF1_01, for Target Path 147, and Target Row 47, U.S. Geological Survey, Sioux Falls, SD. [18] U.S. Geological Survey, 2013, April 27, EO1 Hyperion Scene: EO1H PZ_PF2_01, for Target Path 147, and Target Row 47, U.S. Geological Survey, Sioux Falls, SD. [19] Weng, Q., 2012, Remote Sensing of Impervious Surfaces in the Urban Areas: Requirements, Methods, and Trends, Remote Sensing of Environment, (117)