Spectral Unmixing of SFSI Imagery in Nevada*

This paper was presented at the Twelfth International Conference and Workshops on Applied Geologic Remote Sensing, Denver, Colorado, 17-19 November 1997.

R.A. Neville and K. Staenz
Canada Centre for Remote Sensing Ottawa, Ontario, Canada

T. Szeredi
MacDonald Dettwiler and Associates Richmond, British Columbia, Canada

P. Hauff
Spectral International Inc. Arvada, Colorado, USA

ABSTRACT

Data acquired in Nevada in June 1995 by the SWIR Full Spectrum Imager (SFSI), an imaging spectrometer covering the short-wave infrared (SWIR) from 1220 to 2420 nm, have been analysed on the Imaging Spectrometer Data Analysis System (ISDAS). Both SFSI and ISDAS have been developed at the Canada Centre for Remote Sensing. SFSI, an airborne sensor, has been designed to acquire simultaneously a full spectrum at high spectral resolution (10.3 nm) and a full image swath (496 pixels) at high spatial resolution (1 m). ISDAS uses a look-up-table driven atmospheric correction procedure to retrieve surface reflectances from the radiance data produced by the SFSI preprocessing system. ISDAS incorporates various spatial and spectral modules to analyse the image data cubes.

An image cube of a site near Cuprite, Nevada has been processed via spectral unmixing using reflectance spectra extracted from the image. A computer assisted method was used to determine the spectral end members. These end members and the resulting mineral abundance maps are compared to ground reference information.

1.0 INTRODUCTION

The use of the Short Wave Infrared part of the spectrum for the remote sensing of minerals dates back to the launch of the Landsat Thematic Mapper and before (Goetz and Rowan, 1981). One of the first imaging spectrometers for remote sensing, the Airborne Imaging Spectrometer (AIS) (Vane and Goetz, 1988), was developed specifically for mineral exploration; it operated in the Short Wave Infrared (SWIR) with 128 contiguous bands. This development was followed by the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) (Vane et al., 1993) which collects data in 224 bands distributed across the Visible and Near Infrared (VNIR) and SWIR ranges. This latter instrument, in operation until the present, has produced numerous data sets, spurring interest in research into imagery of high spectral resolution and high spectral dimensionality.

Concurrent with these developments has been the development of a number of portable ground-based instruments for use in the field for in situ mineral identification. These ground-based instruments can be used to validate the measurements made from an airborne or space platform. The spectra from these instruments can also be used as reference spectra in spectral matching and spectral unmixing analyses of the remotely sensed imagery.

SFSI was designed and developed at the Canada Centre for Remote Sensing (Neville et al., 1995) to provide remote sensing researchers with both high spectral and high spatial resolution SWIR imagery for use in developing the methodology and promoting applications in this spectral region. In the design of the instrument, the spectral range 1220 to 2420 nm was selected to include the region 2100 nm to 2400 nm which is of specific interest for mineral identification. The SFSI design gives a nominal band width of 10.3 nm; laboratory and ground-based field data acquired at relatively high spectral resolutions do not indicate a serious need for resolutions better than 10 nm (Goetz and Rowan 1981; Curran, 1989).

Imaging spectrometers such as SFSI provide imagery in many, sometimes hundreds of, spectral bands. This gives the user a great wealth of data; the challenge is to process these data and extract the relevant information in a timely and efficient manner. The Imaging Spectrometer Data Analysis System (ISDAS) (Staenz et al., 1996a) is being developed at the Canada Centre for Remote Sensing to meet these requirements for hyperspectral data acquired with both airborne as well as future spaceborne sensors. It includes visualization tools which have been developed for displaying imaging spectrometer data together with data input/output, data processing, and information extraction modules, as well as linkages to a spectral database and a conventional image analysis system.

In the following sections we describe the sensor design, calibration and operation, and the airborne mission in Nevada. The post-flight data processing to radiance and to reflectance, including the atmospheric correction, is discussed. We describe the spectral unmixing and end member selection process, and display and discuss the results. The discussion includes a comparison of the end member spectra with reflectance spectra acquired with a ground-based instrument.

2.0 SENSOR DESCRIPTION

SFSI is a pushbroom line imager and as such uses no moving parts in the imaging process. All spectral bands for all pixels are imaged simultaneously. This is accomplished by using a two-dimensional array in which the spectra are dispersed over the 'vertical' dimension of the array and the across-track line of pixels is imaged onto the 'horizontal' dimension.

SFSI employs a two-dimensional platinum silicide Schottky barrier CCD array with 488 rows of 512 detector elements. In operation, a region of 480 lines by 496 columns is used; four adjacent lines are summed together to yield an effective array of 120 by 496 detector elements. This gives 120 spectral bands for each of 496 pixels in the across-track dimension for each integration period, which is fixed at 20 ms. The data are digitized to 13 bits of which 8 bits are selected by the operator for recording.

The optical train is f/1.8 and entirely refractive; the spectral dispersive component is a blazed transmission grating. The across-track pixel instantaneous-field-of-view (IFOV) is 0.33 mrad giving an across-track field-of-view (FOV) of 9.4°;. The along track IFOV is currently 1mrad. This is determined by the slit width which has been set so that the slit image is approximately equal in width to the effective detector size (100 µm) in the spectral dimension. This arrangement maximizes signal strength without compromising spectral resolution, which is nominally 10.3 nm at full-width-half-maximum.

The current system records data in bursts of 33.55 MB at real time rates into video RAM, to be downloaded to magneto-optical (MO) disk following acquisition. This method results in the recording of imagery segments of finite length, full swath width, and a selectable number (down to 22) of spectral bands from the full complement of 120 bands. The corresponding number of image lines contained in a cube are 3072 and 560 respectively, taking 61.4 and 11.2 seconds to acquire. All image cubes contain 496 pixels across-track. Continuing development of the recording system will give SFSI a continuous recording capability.

Combined with the target radiance which is focussed by the optics onto the appropriate detector element, is black-body radiation from the optical components and housing which are at ambient temperature. This latter radiation can become significant even at wavelengths as short as 2000 nm. To measure this radiation, as well as the signal originating in the detector itself, SFSI uses, as a 'shutter', a cooled dark target mounted externally to the spectrograph. Before commencing the acquisition of image data, a dark reference frame consisting of the dark signal (minus an electronically introduced residual offset) from each detector in the array is acquired. This dark reference frame is then subtracted from each of the image lines as they are acquired and prior to recording. In addition, a complete cube of dark reference data is periodically acquired and recorded. In the post-flight processing this dark cube is averaged and the resultant frame of offset signals is subtracted from the image data. The net sensor signals are those resulting from the radiance collected from the target.

3.0 SENSOR CALIBRATION

There are three fundamental calibrations that are required to convert sensor signal to useful data. These are spectral, radiometric, and geometric.

The spectral calibration includes the measurements of the centre wavelengths and the band widths for all the detector array elements. For SFSI this has been performed using a calibrated black body source and a series of well defined spectral filters. Ideally, each row of detector elements in the array would have the identical band centre wavelength. In reality, there is a variation in centre wavelength with across-track pixel number due to the combination of expected slit curvature (which, generally, is present in any spectrograph), small optical distortions, and any misorientation of the detector array relative to the spectrograph slit. For SFSI the deviations of the band centre wavelengths from the means for the corresponding detector rows vary, at the extremes, from -14 nm to +6 nm.

The radiometric calibration gives both the relative and absolute responsivities of SFSI's 120 by 496 detector elements. A calibrated radiance source was used to present the sensor with a uniform, accurately variable radiance target. The signal for each detector element, corrected for dark offset, divided by the known blackbody radiance at the element's centre wavelength provides the responsivity for each of the elements in the array.

The geometric calibration, performed on a precision indexing table, provides the view angle for all pixels and all bands of the sensor. It was found that there is a minor misregistration of the pixels from band to band caused by a small rotation of the detector array about the optic axis of the instrument.

The calibration results are applied to the sensor image data in the first stage of processing, which consists of the removal of sensor artifacts from the data and the conversion from sensor signal to radiance. Following the subtraction of the 'dark' signals, as described in Section 2, the image data are converted to radiances by dividing by the matrix of responsivities. Next, the 2-dimensional spectral-spatial frames of radiance data are remapped to give frames in which every pixel in a given row has the same wavelength. Each frame is then resampled to correct for the band misregistration measured by the geometric calibration.

4.0 IMAGERY ACQUISITION AND PROCESSING

4.1. Airborne data acquisition

The mission to Nevada in June 1995 (Hauff et al., 1996; Neville et al., 1997) provided multiple data sets over selected targets for the purpose of testing the capability of the sensor in distinguishing minerals. For this mission the sensor was mounted in an Aero Commander 500 and, for the majority of the data sets the flying altitude was 3000 m above ground level (AGL) at an aircraft speed 75 to 80 m s-1. This gave an along track sampling interval of 1.5 m and an across-track pixel size of 1.0 m.

The data used in the analysis reported in this paper were collected on 21 June, 1995 over a site located 1.7 km north of Cuprite. The coordinates are 37° 32' 38" N and 117° 11' 2" W, the time of acquisition was 23.82 h UTC, and the solar zenith angle was 52.39°. Motion effects due to aircraft attitude were minor and therefore not corrected.

4.2. Conversion from radiance to surface reflectance

In the second stage processing, the at-sensor radiances derived from the sensor signals are converted to at-surface reflectances. The surface reflectance retrieval procedure implemented in ISDAS uses a six-dimensional linear look-up table (LUT) approach with tunable breakpoints to provide additive and multiplicative coefficients for removal of scattering and absorption effects (Staenz et al., 1996b). The six dimensions are: wavelength, surface reflectance, water vapour content, aerosol optical depth, terrain elevation, and view angle. This procedure has the advantage of significantly reducing the number of radiative transfer (RT) code runs and thereby saving the time that would be required to run such a code on a pixel-by-pixel basis.

For the LUT generation for the Cuprite data, the MODTRAN3 RT code was run for two different flat reflectance spectra (ρ1=5% and ρ2=60%), an aerosol optical depth covering the prevailing atmospheric conditions, and for a fixed terrain elevation (1.554 km) averaged over the scene. A fixed value of water vapour content ( 0.6 g cm^-2 ) and a fixed CO₂ mixing ratio (200 ppm) were chosen to best compensate for their respective absorption features. The calculations made on the RT code's wavelength grid were performed at five different pixel locations across the swath to encompass the sensor geometry. The final step involved in the LUT generation is the convolution of the model output radiances with the relative spectral response profiles. SFSI's response profiles are approximated by a triangular-shaped line spread function having a full-width-half-maximum of 10.3 nm.

The third stage processing performs an empirical correction for irregularities in the reflectance data that may have originated in the sensor and escaped correction in the first stage processing from raw to radiance data processing, or that may have resulted from the approximations made in the atmospheric modelling, the selection of parameters, and the radiative transfer calculations. This involves the identification of those image spectra which are the least spectrally variable, then from the ensemble of these spectra, the derivation of a gain and offset for each spectral band. This process is based on flat spectra , although in the case used for the current image, the sought for spectra are assumed only to be slowly varying, not spectrally flat. This process is implemented on ISDAS.

4.3. Spectral unmixing

The method of unconstrained linear unmixing (Adams et al., 1986; Boardman, 1989; Boardman, 1990) expresses the spectrum of each pixel as a linear combination of N end member spectra. The process produces fractional abundance maps of the end members.

The end members were chosen from the image itself. A principal components analysis was performed on the reflectance cube and scatter plots made of pairs of principal components. The end members were chosen from those pixels occurring in the extremities of the scatter plots, often referred to as the 'purest pixels'. The results of the unmixing using the end members selected by this process are displayed in Figure 1 along with a single band image at 2275 nm.

5.0 DISCUSSION OF RESULTS

Spectra from the Portable Infrared Mineral Analyser (PIMA) were used in the analysis of the SFSI data. Seven sites within the boundaries of the imaged area had been sampled with up to six individual samples at each site. These sites were visually identified in the imagery; this was sufficient for our investigations.

The PIMA spectra were used as references against which to compare the end-member spectra selected from the image by the procedure described in Section 4.3 . Four of the six end-member spectra are displayed in Figure 2. By comparing these with the PIMA spectra, one can readily identify three of them as alunite, kaolinite, and buddingtonite. These minerals, which are indicative of hydrothermal alteration, are known to be common in the target area. The fourth end-member spectrum is very similar to the caliche and clay spectra acquired by the PIMA. A fifth is similar to calcite but is not shown here. The sixth spectrum is a 'dark shadow' spectrum. The degree to which the four end-member spectra displayed here resemble actual mineral species and other identifiable target materials is striking. Hence, the abundance images for alunite, kaolinite, and buddingtonite are expected to be similar to the end-member abundance images for the first, second and third end-members respectively. We have used the ground sample spectra as a validation of the presence of these mineral species at the imaged site and as a general check of the unmixing process. While the seven ground sample sites were insufficient to provide an adequate measure of the distribution of minerals over the whole area, the ground measurements were found generally to corroborate the spectral unmixing results. That is, the images indicated the presence of the same minerals as found in the individual ground samples for the localized regions containing the sample sites.

6.0 SUMMARY

Hyperspectral remote sensing data in the SWIR were acquired with the airborne imaging spectrometer SFSI over a site near Cuprite, Nevada in June 1995. This image cube has been corrected for sensor artifacts and has been converted, first to at-sensor radiances, and then to surface reflectances. The latter process was performed on ISDAS, a hyperspectral data analysis system which incorporates a comprehensive set of analysis tools including atmospheric correction and information extraction procedures. The SFSI image cube was analysed on this system using a principal components analysis to assist in the selection of end member spectra. These were then used in a linear spectral unmixing procedure to create abundance maps of these components. These end member spectra were matched to PIMA spectra also acquired from the image site. Three minerals, alunite, kaolinite and buddingtonite, were identified and mapped. The fact that the atmospherically corrected airborne SFSI spectra closely match the ground-based PIMA spectra supports the hypothesis that one can perform spectral unmixing of remotely acquired image data using end member mineral spectra acquired by instruments other than the airborne sensor itself, provided both data sets are processed properly to account for the instruments' spectral parameter differences and to compensate for the atmosphere.

More extensive ground measurements are required to prove the accuracy of the unmixing results obtained for the image cube processed in this exercise. Ground-based spectral reflectance measurements which more closely emulate the airborne measurements with respect to incident illumination and areal coverage would serve to strengthen the link between the airborne and ground-based measurements.

7.0 ACKNOWLEDGMENTS

The airborne SFSI data were acquired by Borstad Associates Ltd. We gratefully acknowledge the cooperation of this company in making these data available. The authors wish to thank Paul Budkewitsch for his background guidance on the geological interpretation of the site. We also thank Marie-Josée Gour of Université de Sherbrooke for the preparation of the images for presentation.

The Nevada mission was funded by the following companies: Borstad Associates Ltd., Spectral International Inc., Barrick Gold Corporation, BHP Minerals Canada Ltd., Cameco Corporation, Cominco Ltd., CRA Exploration Pty. Ltd., Homestake Mining Company, Newmont Gold Company, North Mining Inc., Placer Dome Exploration, and Western Mining Corporation. Data used in this paper have been provided courtesy of these companies.

8.0 REFERENCES

Adams, J.B., Smith, M.O., and Johnson, P.E. , "Spectral Mixture Modelling: A New Analysis of Rock and Soil Types at the Viking Lander Site," J. Geophysical Research, Vol. 91, pp. 8098-8112, 1986.

Boardman, J.W. , "Inversion of Imaging Spectrometry Data Using Singular Value Decomposition," Proceedings of the 1989 International Geoscience and Remote Sensing Symposium (IGARSS'89), and the 12th Canadian Symposium on Remote Sensing, Vancouver, British Columbia, Vol. 4, pp. 2069-2072, 1989.

Boardman, J.W. , "Inversion of High Spectral Resolution Data," Proceedings of SPIE Conference on Imaging Spectrometry of the Terrestrial Environment, Orlando, Florida, Vol. 1298, pp. 222-233, 1990.

Hauff, P., Kowalczyk, P., Ehling, M., Borstad, G., Edmundo, Kern, R., G., Neville, R., Marois, R., Perry,S., Bedell, R., Sabine, C., Crósta, A., Miura, T., Lipton, G., Sopuck, V.,Chapman, R., Tilkov, M., O'Sullivan, T., Hornibrook, M., Coulter, D., Bennett, S. , "The CCRS SWIR Full Spectrum Imager: Mission to Nevada," June 1996, Proceedings of the 11th Thematic Conference on Applied Geologic Remote Sensing, Las Vegas, Nevada, Vol. I, pp. 38-47, 1996.

Curran, P.J. , "Remote Sensing of Foliar Chemistry," Remote Sensing of Environment, Vol. 30, pp. 271-8, 1989.

Goetz, A.F.H. and Rowan, L.C. , "Geologic Remote Sensing," Science, Vol. 211, pp. 781-791, 1981.

Neville, R.A., Rowlands, N., Marois, R., and Powell, I. , "SFSI: Canada's First Airborne SWIR Imaging Spectrometer," Canadian Journal of Remote Sensing, Vol. 21, No. 3, pp. 328-336, 1995.

Neville, R.A., Staenz, K., and Szeredi, T. , "Mineral Reflectances Extracted from SFSI Imagery in Nevada," Proceedings of SPIE conference on Algorithms for Multispectral and Hyperspectral Imagery III, Orlando, Florida, Vol. 3071, (in press), 1997.

Staenz, K., Schwarz, J., and Cheriyan, J. , "Processing/Analysis Capabilites for Data Acquired with Hyperspectral Spaceborne Sensors," IAA Journal, (in press), 1996a.

Staenz, K., Williams, D.J., and Walker, B. , "Surface Reflectance Retrieval from AVIRIS Data Using a Six-Dimensional Look-Up Table," Summaries of the Sixth Annual JPL Airborne Earth Science Workshop, Vol.1, Pasadena, California, (in press), pp. 223, 1996b.

Vane, G., and Goetz, A.F.H. , "Terrestrial Imaging Spectrometry," Remote Sensing of Environment, Vol. 24, pp. 1-29, 1988.

Vane, G., Green, R.O., Chrien, T.G., Enmark, H.T., Hansen, E.G., and Porter, W.M. , "The Airborne Visible/Infrared Imaging Spectrometer (AVIRIS)," Remote Sensing of Environment, Vol. 44, pp. 127-143, 1993.

Figure 1. Single species abundance maps of the Cuprite site. Width of imaged area is 0.5 km.

Figure 2. Endmember spectra selected from the image data (in red) compared to ground reflectance spectra obtained with a PIMA spectrometer.

Click HERE to return to Papers on Geology and Mining

   
Last update 12 March 2007
The contents of this site are © copyright 1993 - 2008, Borstad Associates Ltd.