INVESTIGATION OF IMAGING SPECTROSCOPY FOR DISCRIMINATING URBAN LAND COVERS AND SURFACE MATERIALS Che-Ming Chen, Ph.D. student George Hepner, Professor Department of Geography University of Utah 260 S. Central Campus Drive, Room 270 Salt Lake City, UT 84112-9155 [email protected] [email protected]

ABSTRACT This study investigates the potential utility of the imaging spectroscopy approach to discriminate urban land covers and surface materials. Urban areas are highly heterogeneous so that many surface materials cover less than one pixel size of the broadband data and each material contributes its unique spectral reflectance this mixed pixel. This is socalled spectral mixing problem. Broadband scanner systems and conventional statistical analysis approaches, such as maximum likelihood classification, are incapable of discriminating many urban land covers and surface materials. In this research, the imaging spectroscopy approach was used with field spectroscopic measurement and high-altitude and low-altitude AVIRIS imagery to discriminate the land covers and surface materials in the Park City, Utah urban area. Both high-and low-altitude improved the identification of urban land covers and surface materials.

INTRODUCTION Urban growth has been recognized by many globe change agendas as not only a regional phenomenon but also a continental, global scale phenomenon. (Hepner et al., 1998). It could make profound effects on our physical and socioeconomic environments. Many urban areas face the environmental degradation such as loss of open space, natural vegetation, agricultural lands, wetlands, and natural habitat at an increasing rate. The transformation of land use and land cover is the index of urban growth. The development of the urban information systems, which contain the above baseline information and the urban growth models, is crucial for predicting regional patterns of urbanization. The implementation of the urban systems requires a variety of digital data sources and a temporal database. Urban geographers and planners recognize remote sensing data as a vital part of the total information pool (Barrett and Curtis, 1992). Although urban analysis is one of the most common applications of remote sensing, the information derived from remotely sensed data is often insufficient for operational use. One of the main problems is that the spectral and spatial resolutions of sensors are too coarse to identify the desired information for urban analysis (Hepner et al., 1998). Urban landscape is extremely heterogeneous with a variety of land cover types and surface materials mixing together within a small area. If many materials locate within one sensor pixel, each material will contribute its unique spectral characteristic to the mixed pixel and make the pixel spectrally impure. Besides, many urban ma t e r i a l ss u c ha ss oi l sa n di mpe r v i ouss u r f a c e sy i e l ds i mi l a rs pe c t r a ls i gn a t u r e s .I t ’ sdi f f i c u l tto spectrally distinguish them from each other. Therefore, broadband data and pixel-based analysis are inadequate for discriminating urban land covers and surface materials (Forster, 1985; Ridd, 1995). Both high spectral and spatial resolution are required to solve these problems. Hyperspectral data like AVIRIS which covers spectral range from 400 to 2500nm and has 224 continuous channels with 10nm bandwidth is capable of discriminating most of the terrestrial materials including urban surface materials. Generally, the FWHM of broad absorption features caused by electronic transitions is greater than 50nm. The small absorption features caused by the vibrational processes within 1000 - 2500nm region have FWHM greater than 20nm. Thus AVIRIS with 10nm sampling interval is sufficient to thoroughly describe the absorption features of urban materials (Goetz, 1992 and Clark, 1999). Imaging spectroscopy has been successfully applied to geological, aquatic, ecological and atmospheric research (Curran, 1994). Surprisingly, it has been used sparsely for the study of urban areas (Ridd et al., 1992 and Hepner et

al., 1998). The main goal of this study is to assess the hyperspectral data as a tool for discriminating land covers and surface materials in urban area.

OBJECTIVES There are two objectives for this study. The first objective is to investigate the feasibility of using the field spectra as the reference to identify the urban land covers and surface materials in the AVIRIS scene. The second objective is to evaluate the discriminability of high-altitude AVIRIS using sub-pixel analysis.

METHODS Study area The urban area of Park City, Utah is the main study area due to the representative of mixture of land covers and surface materials for western U.S. cities and other urban areas of the earth undergoing rapid urban growth (Figure 1). Although Park City area is not highly urbanized, it contains the diversity of land covers, surface materials, and vegetation associations. Several land cover types are common to nearly every urban area. These categories are light inert material, dark inert material, asphalt, water, annual weeds and natural grass areas, moist healthy vegetation, trees and shrubs, and mixed cover types (Wheeler, 1985). Most of these land covers can be found in Park City area as well as the other western U.S. cities. Park City also provides the typical example where urbanization spreads from lowlands to adjacent uplands. Therefore, the methodologies, spectral signatures and techniques investigated in this study should then be transferable to many western U.S. cities and other urban areas around the world with the similar situations.

Data acquisition In this project, the AVIRIS data is used to investigate the utility of imaging spectroscopy for discriminating the urban land covers and surface materials because of the following considerations. First, it provides the high quality hyperspectral imagery (ex. high spectral, spatial resolution and S/N ratio) and meets the requirements for this study. Second, another scientific research conducted in Park City area could provide the parameters for the surface reflectance calibration of AVIRIS data. In the Utah 1998 EPA-USGS AVIRIS Study, the USGS spectroscopy lab has developed the methodology for surface reflectance calibration (Clark et al., 1999). Besides, their project will investigate the utilities of AVIRIS data for ecosystem evaluation and environmental management, and the results of this project can be used to expand and verify the use of AVIRIS data for urban analysis. Third, the AVIRIS system ma k e spr ov i s i onf ort h es pa c e bor n ei ma g i n gs pe c t r ome t e r s .Fore x a mpl e ,NASA’ sEa r t hObs e r v i ng -1 mission has launched Hyperion hyperspectral imager on November 21st, 2000. This system has spectral coverage from 400-2500 nm with 10 nm wide contiguous bands which are identical to the AVIRIS system. Therefore, the AVIRIS data is apt to benchmark the potential of the Hyperion system for the future urban analysis. The high-altitude AVIRIS data of Park City area was obtained from the Park City flight on 5 August 1998. The data range for this imagery is from Band 1, with a bandcenter of 369.07 nm, to Band 224, with a bandcenter of 2507.50. The spatial resolution is 20m. A radiative transfer model (ATREM) provided by the University of Colorado was applied to the radiance data to remove the solar spectral response, atmospheric absorptions and atmospheric scattering. The path radiance scattering overcorrected by the ATREM was further corrected by the offset parameter provided by the USGS. Other artifacts of the spectra, which are small-scale spikes irresoluble by the ATREM, were also corrected by the multiplier parameter provided by the USGS derived from the ground calibration site (Rockwell, 2000). The low-altitude AVIRIS data with spatial resolution 2.9m was obtained on 19 October. The data range of lowaltitude data starts from Band 1, with a bandcenter of 370.43nm, to Band 224, with a bandcenter of 2510.10nm, which are slightly different from those of the high-altitude data. The only urban area common to these two datasets is located at Prospector Square area in the northeast of Park City. This area was therefore selected as the study site for the spectral analysis. Similar calibration procedure was applied to the low-altitude AVIRIS data. However, the parameters of the offset and multiplier were not available from the USGS because there was no ground calibration site located in the low-altitude AVIRIS flight line and no field spectra were obtained during the low-altitude overflight. Thus the corrected high-altitude AVIRIS data was used to simulate the ground truth for calibrating the low-a l t i t u deAVI RI Sda t a .Th es pe c t r um oft h ebr i gh tr oof t opofas u pe r ma r k e t( Da n’ sFoods )i nt h eh i gh -altitude scene was selected and then edited to remove the residual atmospheric absorptions. Relatively, the same pixels which cover approximately 6 by 6 pixels in the low-altitude scene were sampled and averaged to provide the spectrum. This low-altitude spectrum was then divided into the edited high-altitude spectrum to obtain the multiplier

used to calibrate the low-altitude data. This method had been validated by the USGS-EPA Imaging Spectroscopy Project (Rockwell et al., 1999). A spectral library containing the reference spectra and the related spectroscopic knowledge of the urban surface materials was required to perform the spectral analysis. Although the existing spectral libraries provide many urban surface materials, their diversity is not enough for this study. Besides, the ability of the reference spectra to separate plant species as well as manmade materials might vary from one region to another and might decrease as the geographic area covered increases. Thus a regional spectral library of Park City was built to better match the AVIRIS data for the spectral analysis. During the summer and autumn of 2000, 80 urban surface materials including 20 roofing materials, 12 paving materials, 23 vegetation and other 25 materials, were measured in 9 field collection sites around the study area using an ASD field spectroradiometer. The reflectance of the field spectra was rescaled to the magnitude of 0 to 1. Since the unit of wavelength, bandwidth, and bandcenter of field spectra are different from those of the AVIRIS data. The field spectra were resampled based on the wavelength and FWHM of the highaltitude and low-altitude AVIRIS data. The locations of field spectra were recorded using a GPS unit, and the digital photographs of each material were also taken using a digital camera. The field spectra were registered onto the geocoded AVIRIS images. The attributes of the field spectra were integrated into the web pages.

Spectral reflectance characteristic of urban materials The selection of the optimal spectral analysis strategy to discriminate the urban surface materials is tied to spectral reflectance characteristics of the targets. For example, if the targets have apparent absorption features, an absorption feature mapping algorithm would be appropriate (Mustard and Sunshine, 1999). The spectral signatures of the urban field spectra were therefore examined before determining the analysis strategy. Representative urban spectra in this study are shown in Figure 2 and 3. In the broadband data, water and asphalt usually confuse each other due to low reflectance. However, the field spectra show that water has apparent fluctuation in the visible region due to the underwater materials, which are water plants in this case. Asphalt also has a broad weak absorption near 2.3 µm due to the hydrocarbons. Sand, concrete, and gravel are usually indistinguishable in the broadband data due to bright reflectance. In the field spectra, they could be discriminated by the magnitude in the visible region or by the weak absorption at 2.2 µm caused by the montmorillonite, which is narrower and stronger in the sand and gravel spectra than in the concrete spectrum. Beside, the spectral curve of concrete from red to near infrared region (0.6 –1.3µm) is a concave, which is different from the convex of sand and gravel in the same spectral region. Since the high-altitude data was obtained in August and the low-altitude data was obtained in October, the plant spectra were measured both in summer and autumn to correspond with the image spectra. Figure 3 shows the leaf spectra collected in the summer. Different species could be differentiated from each other by the field spectra. The chlorophyll absorptions in the visible region, the magnitude of red edge, the broad absorptions near 1.73µm, 2.1µm, 2.3µm due to leaf biochemicals and the magnitude of the leaf water content from 2.1 to 2.3 µm apparently differentiate dry grass and sagebrush from the other plants. Sagebrush has stronger chlorophyll absorption than dry grass near 0.66 µm. Healthy grass has the highest reflectance at 0.9 and 1.1 µm. Blue spruce has the lowest reflectance from 2.1 to 2.3 µm. Besides, Figure 4 presents the leaf spectra reflectance of the same plants in the autumn. The stronger red reflectance and weaker near infrared reflectance make alfalfa and aspen could easier be discriminated in fall than in summer. The spectra of the same urban material collected in the different locations might have spectral variability. For example, paving asphalt in different ages or conditions may slightly change the spectral reflectance and thus could be detected by the field spectrometer. Figure 5 shows the paving asphalt samples measured at the different sites. Sample A is heavily wreathed. Its gravel aggregates revealed and little dirt also covered it. A weak absorption featuren e a r2. 2µm ( mon t mor i l l on i t ei ndi r torg r a v e l ) ,wh i c hdoe s n’ ta ppe a ri nSa mpl eB a n dC,wa st h e r e f or e caused by these conditions. Sample C was measured when the road was just paved. The reflectance of this brand new asphalt was 15% lower than Sample A. Besides, the spectral curve of Sample C is a concave which is different from the convex of Sample A. Sample B is medium-weathered asphalt, and the spectral curve of Sample B is flatter than Sample A and C. In contrast, the spectra of concrete illustrate more consistent reflectance than asphalt. Figure 6 depicts the spectral reflectance of three paving concrete samples collected from different sites. Their spectral curves are almost identical. Two very weak absorptions found at 0.95 and 1.15 µm on sample A might be due to the montmorillonite in the concrete or in the dust covering the sample when the measurement were taken, and they could be treated as artifacts. The two major components of concrete are the cement paste and inert materials such as sand, gravel, and crushed stone. Concrete mixtures are specified in terms of the ratios of cement, sand, and coarse aggregates used. The mixture rations can vary from 1:2:3 to 1:2:4 and 1:3:5 (Microsoft, 2001). The broad absorption of the concrete spectra near 2.2 µm could therefore be explained by the ingredients of sand and gravel. Lime is the

major ingredient of cerement and is made by limestone and other forms of calcium carbonate (Microsoft, 2001). Since limestone consists of almost entirely of calcite, the strong absorption feature of calcite near 2.34 µm was expected to be found in the spectra of concrete too. However, this feature might be interfered by the other a g g r e g a t e di n e r tma t e r i a l sl i kes a n da n dt hu sdi dn’ te x i s ti nt h ef i e l ds pe c t r a .I ts h oul dben ot e dthat many urban ma t e r i a l sa r ema nma dea n dc ompos e dofma nyi ng r e di e n t s .Th u s ,t h e yc a nn otbec on s i de r e da s“ pu r e ”ma t e r i a l s . The strong absorption features of their individual chemical compositions might be flattened and broadened. Besides, if dirt covered the targets during the measurement, weak absorption features at 0.95, 1.15, and/or 2.2 µm might be found.

Spectral analysis strategy and procedure Lacking of consistent, strong, and well-defined absorption features, most urban surface materials are not differentiable by matching the absorption bands. The full spectral mapping methods such as Spectral Angle Mapper (Yuhas et al., 1992; Kruse et al., 1993) and Mixture Tuned Matched Filtering (Boardman, 1998), which compare the spectra using full wavelength range, might be feasible for mapping urban materials with continuum shapes and/or very broad absorptions. Imaging processing was undertaken using ENVI 3.2 software. Low-altitude AVIRIS data with fewer materials mixed in one pixel were analyzed by the pixel-based method, spectral angle mapper (SAM). Fifteen field spectra of the most common land cover types and surface materials in Park City were chosen as the reference to compare the spectral similarity with the image spectra. Pixel-based analysis is inadequate for the high-altitude data due to spectral mixing problem. Thus the mixture tuned matching filtering (MTMF), a partial-unmixing method, was adopted to analyze the high-altitude data. First, The Minimum Noise Fraction (MNF) transformation was performed to reduce the data dimensionality, and the first 40 MNF bands were selected. Then, the Pixel-Purity Index (PPI) analysis was undertaken to find the pure 803 pixels, and 11 spectral endmembers were extracted from them using the N-dimensional Visualizer. The locations of these endmembers were linked to the ground truth information to identify the materials. Finally, the abundance images of each endmember were produced by the MTMF.

RESULTS Figure 7 shows the results of the SAM analysis with the low-altitude AVIRIS data. The SAM analysis successfully discriminate the surface materials using the field spectra. The general categories of vegetation such as healthy grass, dry grass, conifer, and deciduous were classified using the representative field spectra of lawn grass, dry grass, blue spruce, and maple (Figure 7b, 7c, 7d, 7e). Turbid and clear water bodies were differentiated using the field spectra collected from a pond and a swimming pool (Figure 7f, 7g). Soil of the dirt road and red soil of the softball field were separated (Figure 7j, 7k). Three tennis courses were found using the paint spectrum (Figure 7l). Four types of roofing materials, asphalt-based roof, metal roof, wood shingle, and membrane, were discriminated (Figure 7m, 7n, 7o, 7p). Since roofing asphalt and paving asphalt are basically the same material, paving asphalt was not subtracted from the asphalt roof image when the field spectrum of paving asphalt was applied (Figure 7m). The similar situation was found in the image of wood shingle, which contains some pixels of dry brush located on the f oot h i l l .SAM di dn’ tdi s c r i mi n a t et h emi n ors pe c t r a ldi f f e r e n c ebe t we e nt h e m( Fi gu r e7o) . Su r pr i s i ng l y ,pa v i ng concrete has very bright reflectance, but it was confused with asphalt in Figure 7i. This confusion might be due to the SAM algorithm which measures the spectral vector angle but ignores the vector lengths. The SAM algorithm may fail to discriminate the spectra with very similar spectral shape; even they have high contrast of brightness. The abundance images derived from the high-altitude AVIRIS data using MTMF are shown in Figure 8. MTMF analysis demonstrates that the target can be detected against a variety of background materials. Three general categories of vegetation, healthy grass, trees, and mountain brush were identified (Figure 8b, 8c, 8d). The highest concentration of water corresponds to the pond in the Park City Golf Course (Figure 8f). Soil concentrations are located at ski trails and a construction site (Figure 8e). The metal roof of the Park City Public Works was detected in Figure 7h. Four different types of roofing membrane were differentiated (Figure 8i, 8j, 8k, 8l). However, not all of the common urban materials were extracted by means of the MNF, PPI, and N-dimensional Visualizer. For example, asphalt was not qualified as an endmember during the process.

CONCLUSIONS AND DISCUSSION The imaging spectroscopy approach with both high- and low-altitude AVIRIS data makes it possible to discriminate the urban land covers and surface materials, which usually spectrally confuse each other in the broadband data. The low-altitude AVIRIS data with 2.9m spatial resolution offers a significant improvement of the identification of certain classes of materials in the highly heterogeneous urban scene. The high-altitude AVIRIS data, even with a coarse 20m spatial resolution, can detect many urban land covers and surface materials with minor spectral difference by taking advantage of the high spectral resolution along with the sub-pixel analysis. The uncertainties of the discrimination resulted from the spectral variability of the field spectra and the spectral analysis algorithm applied. The spectral variability of the similar materials increases the uncertainty of spectral analysis, but it also provides the potential for more urban applications, which can take advantage of the spectral variability to discriminate more detailed materials. Hybrid spectral analyses are required to resolve the spectral a mbi gu i t ywhi c hc ou l dn’ tber e s ol v e d by the single algorithm.

REFERENCES Barrett, E. C. and Curtis, L. F. (1992). Demography and Social Change, Introduction to Environmental Remote Sensing, London, Chapman & Hall, 363. Boardman, J. W. (1998). Leveraging the high dimensionality of AVIRIS data for improved sub-pixel target unmixing and rejection of false positives: mixture tuned matched filtering, Summaries of the Seventh JPL Airborne Earth Science Workshop, Jet Propulsion Laboratory, Pasadena, CA., Vol. 1, p. 55. Clark, R. N. (1999). Chapter 1: Spectroscopy of rocks and minerals, and principles of spectroscopy, Manual of Remote Sensing, 3rd ed., John Wiley and Sons, Inc., New York. Clark, R. N., et al. (1999). Surface Reflectance Calibration of Terrestrial Imaging Spectroscopy Data: a Tutorial Using AVIRIS, USGS Spectroscopy Lab, http://speclab.cr.usgs.gov/PAPERS.calibration.tutorial/calibntA.html Forster, B. C. (1985). An examination of some problems and solutions in monitoring urban areas from satellite platforms, International Journal of Remote Sensing, 6(1): 139-151. Goetz, A. F. H. (1992). Imaging spectrometry for earth remote sensing, Imaging Spectroscopy: Fundamentals and Prospective Applications, ECSC, EEC, EAEC, Brussels and Luxembourg, Netherlands, pp. 1-19. Hepner, G. F., Houshmand, B., Kulikov, I., and Bryant, N. (1998). Investigation of the integration of AVIRIS and IFSAR for urban analysis, Photogrammetric Engineering and Remote Sensing, 64, pp. 813-820. Kruse, F. A., Lefkoff, A. B., Boardman, J. W., Heidebrecht, K. B., Shapiro, A. T., Barloon, P. J., and Goetz, A. F. H. (1993). The spectral imaging processing system (SIPS) –Interactive Visualization and Analysis of Imaging Spectrometer Data, Remote Sensing of Environment, 44:145-163. Microsoft. (2001). Microsoft Encarta Encyclopedia CD-ROM. Mustard, J. F., and Sunshine, J. M. (1999). Chapter 5: Spectral analysis for Earth science: investigations using remote sensing data, Manual of Remote Sensing, 3rd ed., John Wiley and Sons, Inc., New York, pp. 251-306. Ridd, M. K., Ritter, N. D., Bryant, N. A. and Green, R. O. (1992). AVIRIS data and neural networks applied to an urban ecosystem, Third Airborne Geoscience Workshop, National Aeronautics and Space Administration, Jet Propulsion Laboratory, Pasadena, Ca., 1, pp. 129-131. Ridd, M.K. (1995). Exploring the VIS Model for Urban Ecosystems Analysis Through Remote Sensing, Int. J. of Remote Sensing, 16(12):2165-2186. Rockwell, B. W., Clark, R. N., Livo, K. E., McDougal, R. R., Kokaly, R., and Vance, J. S. (1999). Preliminary materials mapping in the Park City region for the Utah USGS-EPA Imaging Spectroscopy Project using both high and low altitude AVIRIS data, Summaries of the Eighth JPL Airborne Earth Science Workshop, Jet Propulsion Laboratory, Pasadena, CA., pp. 365-376. Rockwell, B. W. (2000). AVIRIS Data Calibration Information in Park City Region, USGS Spectroscopy Lab, http://speclab.cr.usgs.gov/earth.studies/Utah-1/park_city_calibration.html Wheeler, D. J. (1985). Evaluation of thematic mapper data for determining urban land cover, Ph.D. dissertation, University of Utah, Department of Geography, Salt Lake City, Utah. Xiao, Q., Ustin, S. L., McPherson, E. G., and Peper, P. J. (1999). Characterization of the structure and species composition of urban trees using high resolution AVIRIS data, Summaries of the Eighth JPL Airborne Earth Science Workshop, Jet Propulsion Laboratory, Pasadena, CA., pp. 451-460.

Yuhas, R. H., Goetz, A. F. H., Boardman, J. W. (1992). Discrimination among semi-arid landscape endmembers using the spectral angle mapper (SAM) algorithm, Third Annual JPL Airborne Geoscience Workshop, Jet Propulsion Laboratory, Pasadena, pp.147-149.

a.

b.

c.

Figure 1. AVIRIS Scenes of the Park City, Utah a) high-altitude AVIRIS Scene showing white box that outlines the study area where both covered by the high- and low-altitude AVIRIS overflights b) high-altitude AVIRIS scene resampled to match the low-altitude AVIRIS coverage c) low-altitude AVIRIS scene

Figure. 2 Reflectance Spectra of the Common Urban Surface Materials

Figure 3. Reflectance Spectra of the Vegetation in Summer

Figure 4. Reflectance Spectra of the Vegetation in Autumn

Figure. 5 Reflectance Spectra of the Asphalt in Different Ages

Figure 6. Reflectance Spectra of the Concrete Measured at Different Sites

Figure 7. SAM Similarity Images for the Urban Land Covers and Surface Materials Derived from Low-altitude AVIRIS Data (plotted at a smaller scale)

Figure 8. MTMF Abundance Images for the Urban Land Covers and Surface Materials Derived from High-altitude AVIRIS Data (plotted at a larger scale)