This article was downloaded by: [Colorado College]On: 29 October 2014, At: 18:17Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,37-41 Mortimer Street, London W1T 3JH, UK

Canadian Journal of Remote Sensing: Journal canadiende télédétectionPublication details, including instructions for authors and subscription information:http://www.tandfonline.com/loi/ujrs20

The Impact of Spectral Band Characteristics onUnmixing of Hyperspectral Data for Monitoring MineTailings Site RehabilitationJ. Lévesque, K. Staenz & T. SzerediPublished online: 31 Jul 2014.

To cite this article: J. Lévesque, K. Staenz & T. Szeredi (2000) The Impact of Spectral Band Characteristics on Unmixing ofHyperspectral Data for Monitoring Mine Tailings Site Rehabilitation, Canadian Journal of Remote Sensing: Journal canadien detélédétection, 26:3, 231-240, DOI: 10.1080/07038992.2000.10874772

To link to this article: http://dx.doi.org/10.1080/07038992.2000.10874772

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) containedin the publications on our platform. However, Taylor & Francis, our agents, and our licensors make norepresentations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of theContent. Any opinions and views expressed in this publication are the opinions and views of the authors, andare not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon andshould be independently verified with primary sources of information. Taylor and Francis shall not be liable forany losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoeveror howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use ofthe Content.

This article may be used for research, teaching, and private study purposes. Any substantial or systematicreproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in anyform to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://www.tandfonline.com/page/terms-and-conditions

( anudian .Iournal ot Remote Scnsiug/.JournaJ canadien dc tClellCtcction

The Impact of Spectral Band Characteristicson Unmixing of Hyperspectral Data for

Monitoring Mine Tailings Site Rehabilitation

by J. Levesque • K. Staenz • T. Szeredi

RESUMECette etude demontre l'importance de la position, dunombre, et de la largeur des bandes spectrales sur lesresultats de deconvolution spectrale des donnees dereflectance du Compact Airborne Spectrographic Imager(casi). Les donnees ont ete acquises dans le visible et leproche infrarouge pour le site de rejets miniers de la mineCopper Cliff pres de Sudbury (Ontario, Canada). Afind 'etudier I'effet du nombre de bandes spectrales, lesdonnees casi ont ete reduites de facon systematique de 65it 33, 17, 8 et 4 bandes en maintenant la largeur de bandeinitiale de 8,7 nm it mi-hauteur. Pour etudier l'effet de lalargeur des bandes, une serie de donnees interpolees parune fonction cubique a ete creee it partir des spectres desdonnees originales casi de 65 bandes. Une image de sixbandes spectrales pre-selectionnees a ete produite pourchaque largeur de bande it mi-hauteur variant entre 8,5nm et 76,5 nm avec un intervalle de 8,5 nm. Les donneesmultispectrales it haute resolution spatiale Ikonos 2(comportant des fenetres spectrales plus larges et moinsnombreuses que casi) ont ete simulees afin d'etudierl'effet combine du nombre et de la largeur de bande. Unedeconvolution spectrale lineaire a ete appliquee sur lesdonnees it partir de la signature spectrale de 5composantes elementaires : vegetation, chaux, rejetsminiers oxydes. et deux types de surfaces d'eau. Lesresultats montrent une difference absolue moyenne allantjusqu'a 0,02 (ecart-type: ±O,05) entre les fractions descomposantes elementaires des donnees simulees et cellesprovenant de la serie de donnees originales casi de 65bandes lorsque que les bandes spectrales sontselectionnees selon des proprietes physiques etspectrales. La difference absolue moyenne est beaucoupplus elevee lorsque les bandes sont positionnees de faconaleatoire. En general, la deconvolution spectrale desdonnees simulees produit des resultats semblables it ceuxobtenus avec la serie de donnees originale casi de 65bandes dans la mesure oil la position des bandesspectrales est determinee d'apres des proprietesphysiques et spectrales. De tels resultats, entre autre ceuxde la simulation de bandes larges, sont realisables engrande partie parce que les composantes elementaires ontdes signatures spectrales tres distinctes.

SUMMARYThis paper demonstrates the impact of bandcharacteristics on spectral unmixing of CompactAirborne Spectrographic Imager (casi) reflectance dataacquired in the visible near-infrared spectral range overthe Copper Cliffmine tailings site near Sudbury (Ontario,Canada). Spectral unmixing is used to monitor therehabilitation status. For this purpose, the bands werereduced systematically from 65 to 33, 17, 8 and 4 bandsusing the original casi bandwidth of8. 7 nm full width athalfmaximum (FWHM). An interpolated data set using acubic spline was derivedfrom the 65-band casi spectra tostudy the effect of the bandwidth. The bandwidth atFWHM was varied between 8.5 nm and 76.5 nm inincrements of 8.5 nm for six selected band positionsacross the casi wavelength range. High spatial resolutionIkonos 2 multispectral sensor image data with less bandsand larger bandwidths than casi were simulated in orderto investigate the combined effect ofnumber ofbands andbandwidth. These data cubes were unmixed with aconstrained linear approach involving the endmembersgreen vegetation, lime, oxidized tailings, and two differentwater types. A comparison of the unmixing resultsindicates a mean absolute difference of up to 0.02(standard deviation: ±0.05) between the endmemberfractions of the simulated data versus the 65-bandbenchmark data if the bands are selected according tophysical spectral properties. The errors are much largerif the bands are not positioned properly. In general, theunmixing ofthe simulated data, including thefour broad­band Ikonos simulation case, produced similar results aswith the full 65-band casi as long as the bands arepositioned with respect to physical spectral properties.Such a result, especially for the broad-band simulations,was achieved mainly because the endmembers arespectrally very distinct and do not show subtledifferences.

• J. Levesque is with MIR Teledetection Inc., 2182 de laProvince, Longueuil, QC, J4G 1R7, Canada.E-mail: josee.levesque@ccrs.nrcan.gc.ca

• K. Staenz is with the Canada Centre for Remote Sensing,588 Booth Street, Ottawa, ON, KIA OY7, Canada.

• T. Szeredi is with MacDonald Dettwiler and Associates,13800 Commerce Parkway, Richmond, BC, V6V 213, Canada.

© Government of Canada/Gouvernement du Canada ..

Dow

nloa

ded

by [

Col

orad

o C

olle

ge]

at 1

8:17

29

Oct

ober

201

4

Vol. 26, No.3, .Iune/juin 2000

INTRODUCTION

Canadian liability for acid mine drainage is in the order of$2 billion to $5 billion, depending on the technique used to

dispose of and treat the acidic waste (Feasby and Jones, 1994).In the Sudbury area alone, Inco Ltd. spends an annual $5million to reclaim property (lnco, 1998). Given these enormouscosts, monitoring techniques are needed to evaluate theirefficiency for reclamation of mine tailings sites, e.g.,revegetation (Hossner, 1988). Recent work indicated thatremote sensing, especially imaging spectrometry, has thepotential to playa significant role in this area (Ferrier, 1999;Shang et al., 1999 ; Mueller et al., 1997; Farrand and Harsanyi,1997; Swayze et al., 1996; Singhroy, 1996). In previousstudies, Levesque et al., (1997 and 1999) demonstrated thevegetation monitoring capability of the Compact AirborneSpectrographic Imager (cas i) over the Copper Cliff minetailings area near Sudbury, Ontario, Canada. If long termmonitoring is required, airborne surveys can become quiteexpensive. The question is whether one can achieve similarresults using Ikonos 2 and other forthcoming high spatialresolution satellites such as Orbview-3 and Quickbird (ASPRS,1996; Space Imaging, 1999). Table 1 summarizes the basicinstument charactersitics of these sensors. The spatialresolution of these sensors is similar to that of the casi dataacquired over the Copper Cliff mine tailings site (2.5 m by 4.3m). However, these sensors have fewer spectral bands and theirbands are wider than those of the casi visible-near infrared(VNI R) hyperspectral data set used for this study.

This paper investigates the effect of varying bandwidth andnumber of bands on the ability to monitor mine tailingsrevegetation over a site near Sudbury using a spectralunmixing approach. In addition, the four Ikonos 2 bands andrelated bandwidths were simulated in order to evaluate theirpotential monitoring capability. Since Orbview-3 andQuickbird have basically the same spectral configuration asIkonos 2, the results obtained for Ikonos 2 are assumed to beapplicable to these sensors as well. The casi data set used forthe study presented in this paper was collected in 72 bands in

1996. The data analysis was carried out on the ImagingSpectrometer Data Analysis System (lSDAS) of the CanadaCentre for Remote Sensing (Staenz et al., 1998).

IMAGE DATA USED

The image data was collected with a casi instrument inspectral mode on August 24, 1996 over the Copper Cliff minetailings site near Sudbury, Ontario, Canada. The data setincludes 72 contiguous, 8.7 nm wide, spectral bands coveringa wavelength range in the VNIR from 407 nm to 944 nm(Anger et al., 1996). The instrument was flown at an altitudeof 1905 m resulting in a pixel size of 2.5 m in the across trackdirection and 4.3 m in the along track direction. Table 2summarizes the casi sensor configuration parameters. Onlywavelengths between 429 nm and 913 nm (65 bands) wereused from the casi data set because of the poor data quality ofthe first three bands as well as the last four bands. This ismainly due to the drop-off of the responsivity of the silicondetector at both ends of the casi wavelength range.

ANALYSIS ApPROACH

Analysis of the casi data included removal of the mostsignificant aircraft motion effect, surface retlectanceretrieval, band simulation analysis, constrained linearspectral unmixing, and comparison of the resulting fractionimages achieved with the various simulated data sets versusthe one retrieved from the full casi data set. The dataprocessing scheme is outlined in Figure 1.

Pre-Processing

In order to correct for the most significant aircraft motioneffect, the aircraft roll was calculated using the navigation datato calculate lateral pixel shifts for each line. These shifts werethen applied to the entire image cube on a line-by-line basis.

In the next pre-processing stage, surface retlectances werecomputed from calibrated (at-sensor radiance) data. Theprocedure applied to the data uses a five dimensional look-up

Table 1.Spectral band configuration and spatial resolution for the Quickbird, Orbvicw-3 and Ikonos sensors.*

Sensor Ikonos 2 Quickbird Orbview-3

Company Space Imaging Earthwatch Orbital Sciences

Spectral 450 - 520 450 - 520 450 - 520

bands 520 - 600 520 - 600 520 - 600

(nm) 630 - 690 630 - 690 625 - 695

760 - 900 760 - 890 760 - 900

Spatial

resolution (m) 4 4 4

as of September 1999

Dow

nloa

ded

by [

Col

orad

o C

olle

ge]

at 1

8:17

29

Oct

ober

201

4

( anadian .luurnul III Remnte Scnsing/.lllnrnal canadien tic trlctlrtrctilln

casi RadianceData Cube

EndmemberFractions

SelectEndmembers

Figure I.Data processing scheme (SRP = Spectral Response Profile).

Surface Reflectance

Assess/removeband-to-band

errors

Roll Corrected Data

Band Simulations

The number of bands was reduced by skipping every otherband of the 65-band data cube resulting in a 33-band data set.This procedure was repeated to produce data sets containing17, 8, and 4 bands, respectively. This systematic way to pickthe bands, especially concerning the data sets with 8 and 4bands (Table 4), is not ideal from a physical point-of-view. Thebands are not properly positioned for the application focusedon in this study. In order to overcome these difficulties, six casibands were selected based on the spectral characteristics ofvegetation for inclusion in the analysis (Staenz, 1996). Thisdata is referred to in the paper as the geobotany data set. Bandpositions of this data set are listed in Table 4.

In order to test the effect of varying the bandwidth on theunmixing results, the original casi data were fit with a cubicspline and interpolated to a specific wavelength grid (Presset al., 1992). The grid size is determined by dividing thespectral response profile of the bands to be simulated into 20intervals. This approach is appropriate since subtle spectralchanges do not occur for the target type considered in this

table (LUT) approach with tuneable breakpoints to provideadditive and multiplicative coefficients for removal ofscattering and absorption effects (Staenz and Williams,1997). The LUT variables are: wavelength, pixel position,atmospheric water vapour, aerosol optical depth, and terrainelevation. This procedure has the advantage of reducingsignificantly the number of radiative transfer (RT) coderuns, thereby saving the time that would be required to runsuch a code on a pixel-by-pixel basis.

For the LUT generation for the different data sets, theMODTRAN3 radiative transfer code was run for the inputparameters as listed in Table 3 for a low-level (5 %) andhigh-level flat reflectance spectrum (60%). This produces aLUT for each reflectance level. These LUTs were producedfor five pixel locations equally spaced across the swath,including nadir and swath edges, and for single values ofaerosol optical depth (horizontal visibility), terrain elevation,and water vapour content. The final step involved in the LUTgeneration is the convolution of the model output radianceswith the relative spectral response profiles of the sensor.casi's response profiles were approximated by a trapezoidal­shaped line spread function.

For the retrieval of the surface reflectance, the LUTs wereadjusted only for the pixel position using a bi-linear interpolationroutine (Press et al., 1992) since single values for the other LUTparameters were used for the entire cube. The water vapourcontent was determined by applying the atmospheric correctionprocess iteratively to sample scene spectra, while adjusting thisatmospheric parameter, and selecting the value which bestremoved the water vapour absorption features on average. Thesurface reflectance of each pixel could then be computed asdescribed in Staenz and Williams (1997). Sample reflectancespectra are displayed in Figure 2.

An assessment of the retrieved spectra revealed only minorirregularities in the reflectance data that may have originated inthe sensor, or that may have resulted from the approximationsmade in the atmospheric modelling and the selection ofparameters. These minor band-to-band errors were notcorrected.

Table 2.cas; sensor configurations.

Spectral coverage 407-944 run

Number ofbands 72

Spectral sampling interval 7.6 TIm

Bandwidth at FWHM* 8.7 run

Sensor altitude above ground 1905 m

Ground resolution: across track 2.5 m

along track 4.3 III

Swath 406 pixels

* FWHM: Full Width at Half Maximum

..

Dow

nloa

ded

by [

Col

orad

o C

olle

ge]

at 1

8:17

29

Oct

ober

201

4

Vol. 26, No.3, .lune/juin 2000

SPECTRAL UNMIXING

Figure 2.casi surface reflectance spectra retrieved from single pixels.

(1 )

830730630

. .... Oxidized Tailings

--Water

--Vegetation

Wavelength (nm)

530

50

40

~~

ell 30uc..'0ell 20'iii0::

10

0430

2 data, the off-nadir viewing direction should be limited to lessthat 15° (casi data set: +20°/-15°) when acquiring the data.

Constrained linear spectral unmixing was performed on all thesimulated casi data cubes using an algorithm implemented inISDAS (Szeredi et al., 2000 ; Boardman, 1989 and 1990). Themethod decomposes the image spectraS in terms of endmemberspectra Si:

,\

where 0 :s; J; :s; I, If, = I andr, is the fraction of endmember i

contributing to th~' image spectrum S, N is the total number ofendmembers, and r is the error term. The result ofthe unmixingprocedure is a set ofN fraction images that show the fractionalabundance of the endmembers. Endmember spectra wereselected from the 65-band image cube using the first threeprincipal components (PCs) which account for 77%, 21% andI% of the variability, respectively. Endmembers are the purestpixel spectra in the data set and are often located at theextremities of the scatter plot which results when the spectraldata are plotted in PC space. Five endmembers were selectedfrom the 65-band data as shown in Figure 3. Theseendmembers were identified as lime, green vegetation,

study. The interpolated data is assumed to be the underlyingdata set for the subsequent simulations. The interpolatedvalues were then convolved to Gaussian spectral responseprofiles with full width at half maximum (FWHM)bandwidths, ranging from 8.5 nm to 76.5 nm in increments of8.5 nm. The profiles were centred at the six band positions ofthe geobotany data set as Iisted in Table 4.

In order to consider a more realistic case with a few wellselected broad bands, the spectral band characteristics of thefour-band instrument onboard Ikonos 2 were simulated withthe same technique as described in the previous paragraph fromthe casi 65-band data set using Gaussian spectral responseprofiles. The centre wavelengths and associated bandwidths aresummarized in Table 4. The fourth band centred at 830 nmwith a bandwidth of 140 nm FWHM could not be simulatedproperly with casi, since the 65-band data set with an upperwavelength limit of 913 nm does not entirely cover the righttail of the response profile. Therefore, a narrower bandwidth of83 nm FWHM was used to approximate the fourth lkonosband. To ensure that our results are compatible with reallkonos

Table 3.Input parameters for MODTRAN3 code runs.

Atmospheric model Mid-latitude Summer

Aerosol model Continental

Date ofoverflight August 24, 1996

Solar zenith angle 31.5°

Solar azimuth angle 176°

Sensor zenith angle variable

Sensor azimuth angle variable

Terrain elevation 0.3 km

Sensor altitude above sea level 2.21 km

Water vapour content 2.35 g/cm'

Ozone column as per model

CO2 mixing ratio as per model

Horizontal visibility 50 km

Table 4.Spectral characteristics of the simulated data sets.

Data Set Centre Wavelength (nm) Bandwidth (nm)at FWHM

8-band 458.7,518.2,578.2,638.6,699.3,760.28,821.4,882.7 8.7

4-band 503.2, 608.3, 714.5, 821.4 8.7

Geobotany (6 bands) 480.9, 548.1,608.3,676.5, 745.0, 829.1 8.5 to 76.5in increments of8.5

Ikonos 2 485,560,660,830 80,80,60,140

Dow

nloa

ded

by [

Col

orad

o C

olle

ge]

at 1

8:17

29

Oct

ober

201

4

Water 2

PC2 Lime

,.

PC1

E)Green Vegetation

PC2

Green Vegetation

Lime

PC3

t'i~ure 3.Scarter pl"t~ " f prin cipal compo"ent (rC) I versu s PC2 and PC3 \ e rsus PC2 "f the {',,~i h5-band dab.

oxidized tailings, water 1, and water 2 using field referenceinform ation in combina tion with ground-based spectra lmeasurements collected with a GER3700 fieldspectroradiometer. Water 2 is distinct from water 1 because ofits high content of sewage, ta ilings, and lime. The spectra ofthese endmembcrs are displayed in Figure 4. According to thePC scalier plots, the same five endmembers were identified forthe different simulated data cubes. The endmcmbcrs for thedifferent simulations were then generated from theendmcmbers of the 65-band data the same way as the simulateddata sets discussed in the previous section.

Assessment of Results

ta ilings rehabilitation purposes. The fractions vary between 0and I with blue representing a low value and red a highvalue. The lower part of Figure Sa is an active tai lings areawhere fresh tailings are being deposited.

An examination of the error images associated with spectralunmixing (Equat ion I ) of the 65-band data revealed nosignificant errors and hence, one can be confident all majorcndrncmbers have been found and the endmembers span thedata space . In addit ion, the unmixing results have beenvalidated in the field as reponed by Stacnz et al., (1999).Therefore, the fraction images derived from this data cuberepresent an ideal reference for this study.

The fraction images retrieved with the spectral unmixingprocedure from the simu lated data sets were then compared tothose extracted from the 65-band data. The absolute values ofthe differences between the fractions found using the 65-banddata, and that found using the sub-sampled multispectral datasets, were computed on a pixel-by-pixel basis for eachendmembcr. Subsequently, the mean and standard deviation ofthese absolute differences were calculated.

Rand Reduction

Figure 6 shows the effect ofdecreas ing the number of castbands on the spec tra l unmixing results. The mean abso lutedifferences remain unde r 0.02 (standard dev iation : ::!: 0.03)

50

Green Vegetation Ume• • • · Oxidized TaIlings --water 1.. . .•. INaIer2

Figu re 4.Endmcmb<' r spectra retrieved from the hS-ba nd {'Q~i dal a set.

'30630 730Wavelength (nnl)

/

530

I\s an example of the output of the const rained unmixing, thefraction images of vegetation, lime, and oxidized ta ilingsretrieved from the 65-band data cube are shown in Figure 5together with a colour composite providing an ove rview ofthe study site. The upper part of Figure Sa is the inactivetaili ngs where most of the revegetation work is being done.Some of the typical areas of the tailings arc identified in thecolour composite and depicted in the adjacent photograph s.The three end members shown are those most important for

RESULTS

Dow

nloa

ded

by [

Col

orad

o C

olle

ge]

at 1

8:17

29

Oct

ober

201

4

' nl. 211. ' 0.3, JunelJ,"n 2UllO

a b c d

1.'o"E I

Fraction1.'

Fra ctiono~ '

. ,0.0 1.0

Frac tion

t

i•

Inacti veTailings

Area

Act iveTaili ngs

Area

.·i~u~ 5.('olour cemposkc lind IIs\ociatcd ptcrurcs sho...inl:talan evervie... of the ta i l in~s area ...ilh fraction i mll~(" of lhe tbree endmcmhers (h) ~rt'e n vCll,clatlon,(e) linu.-, and (d) o\ idi/.cd lailinJ.:«. The celeur eoml,,,silc remeved fr"m the n/,i liS-hand dalll w i was ~ennllled .. ilh hand, 19 t540nm) In Ihe hlne, 37(11711 nm) in Ihe green. and 45(737 nm) in the red, rt'sl't't'liwly.

ses but rema ins below 0,04 ( :!:O.O5) for greention and lime and helow 0.05 ( :!:0.06) for oxidized oreos. respectively. These errors dec rease to the same III O.ldoze<l Ta'~"Q$

S for the 33, 17. and 8 hand cases when the hands arc • 0 125 DWater1U DWa le,]0

.d according to physical spectral properties, as is the e • G'O'en Vell"lal"'"

~oteo m~

01 the 6-hand gcob otan y data set. The two watermhers consistently display a higher mean absolute e

0.075

' nee than the oth er endmembcrs. This can he ~snood from Equat ion I . The wate r endmcmbcr spectra, • 0 .050

S"_,, relat ive ly dark. the spectra l0

d are henc e ••tudes I'~"I I and IS-we! <I re small compared to the other • 0025 .nm Dh lmbers . Due 10 this fact the fractions /.'1and .I~.! can 0.000hy 0 relat ively large amount withou t chang ing 33 " e , ecantl y the ;0 Equat ion I. The

GooOOI. nysum spectrum NumberofBands D.", 501

ined unmixing takes advantage of this fact and. .·i~"re II.the fract ions .1:. 1and .t:,! vary more than the fraction s Mean ahsol"te difference betw een '" 6S-hand nnmi\inl: res"l..

other endmembers. (e"dnlenlber fractions) and "nmhinl: res,,_" "r Ihe hand red"e' i"nanalysis.

for all the cndmernbers using 33, 17. and 8 bands. Whenconsiderin g fo ur bands. the overall mean absolute differenceincrcavcgctatai linglevel asclcc rccase fendmediffereund crs5.-, anrnagrucndmcvarys ignificonstrahence,of the

Dow

nloa

ded

by [

Col

orad

o C

olle

ge]

at 1

8:17

29

Oct

ober

201

4

( .1ll,HII.' " ,,,,, , 11,,1 " I I{ , ", ,,1 , 'en, ,,,!: I" ", " .,1, ..., "t"·,, d, 1"I,d,I "Ilnn

Ikonos Simulation

D Standa<d Devoa1"", of t"el\MoIul~ ~,~!\CC

Endmembe,

Wa,... l

JL0000

coso• 0070,e coer

~ 0050

S 0040a1: 0030•: 0020

::I 0010

lkonos bands are better positioned to enhance differencesbetween spect ra. The mean absolute difference for allendmembcrs do not exceed 0.014 (~0.07). As pointed outbefore, the two water endmembers. followed by the oxidizedtailings, the green vegetation and the lime endmembers. showan inverse relationship between their spectral rellectancemagnitude and their mean absolute difference. Similarly, thestandard deviation of the mean absolute difference is related tothe magnitude of the endmember spectra, which indicates thatmore variation is expec ted when using low rellectanccendmembers such as water I and water 2.

As an example. the spatial representation of the lkonossimulation ana lysis is presented in Figure 9 for the greenvegetation endmcmber. Figure 9a and Figure 9b depict thefractions retrieved from the 65-band and the simulatedlkonos data, respectively. The absolute difference map ofthese fraction images is shown in Figure 9c. The bluishportion of the difference map indicates lower differenceswhile the red parts show higher differences. Of the data,98.8% are in the absolute difference range of 0.00 to 0.05.

Although the resu lts indicate genera lly a reasonab leagreement between the endrnember fraction images retrievedfrom the simulated data sets versus the 65·band data, severalpoints have to be considered .

The separation and identification of endmembers becomesmore difficult if one moves from hyperspectral data towardmultispectral broad-band data. The limited spectral resolutioncombined with a limited number of bands makes it difficult 10separate spectra lly subtle differences in endmember spectraand, subsequently, can be expected to lead generally to poorunmixing results. However, since the endmembers in thisstudy are spectrally very different, the unmixing of thesimulated broad-band Ikonos 2 data produced fract ionssimilar to those resulting from unmixing of the casi 65-banddata. This result is consistent with the results of thebandwidth simulation ana lysis using well-de fined bandposit ions for the application under consideration. However,

Fig" re 8.Mean and standa rd deviation or the abs olute difference between th e 65­blind unm;,,;inlt res"lt$ lInd the Ikonos 2 simula ti" n unmixinl: re.ulls foreach endmember.

DISCUSSION

• _ A.bsoIu1e O"""ence

• MftM Absolute 0<11........,.

D Standa'll~ 01!he~"Ch~~enoo

Water 1

0"'"

002{1•,i 0015

o

r:II J...-l-l-.-.l-L• • -'-,..~ I .~~ Jl85 11 25. 5 :M 42.5 51 !>9S 68 76.S

Bandw ld1tl (n m)

Green Vegetation

,•e Oa.O

~o ""0~! 0020

i:: IL• .• ~ . n n.Jl85 11 25S 3olo 425 51 595 &8 165

B.ndwMfth l nm)

In Figure 8 the mean absolute difference is shown for eachcndmcmbcr for the lkonos 2 simulation unmlxlng resultsagainst the 65-band unmixing results. Unlike for the four-handcase used in Figure 6, the four Ikonos 2 bands display lowermean absolute differences. This was expected since the four

Figu re 7.Mean and standa rd dn- iation of the ab solute diffe rence of the g reenH'gela tton and .... ter t end membe r fr actions deri~ed flllm th e oS-bandCa.' ; tlata anti tbl' different bandwid th s;muilltion$ o~;ng the geooo lan y(6-band ) data cub e.

Varying Bandwidth

The comparison between the fractions retrieved from the 65­band data and the data sets of different bandwidth simulatedfrom the geobotany image cube revealed mean absolutedifferences below 0.008 ( ~ 0.02 1) for the vegetation, lime, andoxidized tailings endmembcrs. The maximum mean absolutedifference values occured at the 8.5 nm bandwidth. A similartrend was found for the endmembcrs water I and water 2, butwith a higher mean absolute difference of up to 0.02 1( ::t:: 0.051). The standard deviation of the absolute differencedecreases for all endmembcrs as the bandwidth increases sincethe local variations (in wavelength) in the spectrum aresmoothed over, This results in less fractional variations and asmaller standard deviation. As an example, Figure 7 shows themean and standard deviation of the absolute difference for thegreen vegetation and water I endmembers.

Dow

nloa

ded

by [

Col

orad

o C

olle

ge]

at 1

8:17

29

Oct

ober

201

4

\01. 26, ' 0, J . June/Jilin 2000

a b c

0.0Fraction

1.0 0.0Fraction

1.0 0.0Fraction

0.05

Fi~u r" 9.Fraction imag,,~ r"lri"H'<1 [rum (al Ihe 6S-hand ...t,,1 dala and (h) Ih.. , i", " la l"tl l l;.,m", 2 tI:.l a .. Ith (cl a..od aled ab,nll' l" tlilTcrcnn i"' '' I:'-o....r Ih" ('ulll,erCliff mine lailin gs a rea in Sud bnry.

Dow

nloa

ded

by [

Col

orad

o C

olle

ge]

at 1

8:17

29

Oct

ober

201

4

the unmixing results are strongly affected if the bands are notpositioned properly for a given application as shown inFigure 6 with the band reduction analysis.

The endrnembers for a specific simulated data set werecomputed the same way as the image data set itself, e.g., forthe Ikonos simulation the endrnember spectra of the 65-banddata were convolved to the Ikonos four-band characteristics.In order to evaluate the impact of this approach on thefraction images, the endrnembers were also extracted fromthe simulated data sets themselves. The latter procedure isusually the preferred approach since scaling issues betweenimage data and endrnernbers of ground (library) spectra canbe avoided. However, if spectra from a library are used asendmernbers, then the band characteristics of the sensorunder consideration have to be simulated from those spectra.No significant differences were found between fractionsretrieved with endrnembers selected from the simulated datasets directly and via the 65-band data.

The unmixing procedure requires N-I bands to unmix Nendrnembers. With five endrnembers retrieved from the datasets under consideration, the number of bands was limited tofour in this study. Complicated scenes containing moreendrnernbers might need more bands to perform unrnixingthan provided by the high spatial resolution sensors such ason board Ikonos 2. Accordingly, the unmixing procedurecannot be applied in such cases using this type of sensor.Other methods such as traditional classification approaches(e.g., maximum likelihood) could be used, but they are notable to provide the same detailed target information asspectral unmixing.

CONCLUSIONS

The influence of spectral band characteristics (number ofbands, bandwidth) on spectral unmixing have beeninvestigated for monitoring the rehabilitation status of minetailings. For this purpose, different band sets, including thefour-band multispectral sensor onboard Ikonos 2, have beensimulated from hyperspectral VNIR 65-band casi surfacereflectance data. The resulting endmernber fractions for greenvegetation, lime, oxidized tailings, water I, and water 2,retrieved from the simulated data sets, have been comparedagainst those obtained from the 65-band data.

The unmixing results obtained for the different endmernbersby decreasing the number of bands produces mean absolutedifferences that do not exceed 0.02 (standard deviation:±0.03) for 33-band, 17-band and 8-band data sets. The 4-banddata set yields larger differences, up to 0.14 (±0.22).However, Ikonos 2 simulation indicates that a mean absolutedifference of 0.014 is achievable if the bands are selectedaccording to physical spectral properties. The largest meanabsolute differences between the fraction images occur forwater 1 and water 2. This is also true for the bandwidthsimulation study. It revealed mean absolute differences of amagnitude similar to those resulting from the band reductionstudy with the exception of the 4-band simulation. In this case,

( anadian ,1,,",",11 "I Remote S~lslllgl.I"I,"alcanadien d~ tel~dl'l~rtlOlI

the mean absolute difference is significantly lower for thebandwidth cases, about 7 to 10 times depending on theendmernber, Furthermore, the bandwidth simulation resultsindicate that as bandwidth increases, the standard deviation ofthe mean absolute difference decreases.

The study shows that especially for the non-waterendrnembers, which are most important for revegetationmonitoring, similar unmixing results were obtained using thesimulated data as retrieved with casi hyperspectral data as longas the bands are positioned according to physical spectralproperties. This is even true for the four broad-band Ikonos 2simulation. However, for sensors with similar bandcharacteristics as Ikonos 2, the results cannot be readilytransferred to more complicated scenes and hence, containingmore endmernbers than the scene used in this study. This is dueto the required minimum number of bands, N-l, for unmixingN endrnernbers. It is also more difficult to identifyendmembers, especially when subtle differences occur, if onlya few spectral bands are available. Nevertheless, the studydemonstrated that high spatial broad-band sensors such asIkonos 2 and Quickbird have the potential for monitoring minetailings rehabilitation using spectral unmixing.

ACKNOWLEDGEMENTS

The authors thank The Centre for Research in Earth andSpace Technology (CRESTech) for the acquisition of the casidata and related technical support. The support of the 1996field campaign by R, McGregor (Atomic Energy of CanadaLtd.) and A. Dumouchel (Canada Centre for Remote Sensing)and the technical assistance of J. Lefebvre (MacDonaldDettwiler and Associated) is gratefully acknowledged. Ourthanks to P. Yearwood of Inco Limited for providing access tothe mine tailings site and for logistical support.

REFERENCES

Anger, C.D., S. Achal, T. Ivanco, S. Mah, R. Price, and J. Busler. 1996."Extended Operational Capabilities of cast" Proceedings of the 2ndInternational Airborne Remote Sensing Conference, San Francisco, California,pp.124-133.

ASPRS. 1996. "Land Satellite Information in the Next Decade - The WorldUnder a Microscope", Executive Summary, American Society forPhotogrammetry and Remote Sensing, Bethesda, Maryland, U.S.A., 72 pages.

Boardman, J.W. 1989. "Inversion of Imaging Spectrometry Data UsingSingular Value Decomposition", Proceedings of the /989 InternationalGeoscience and Remote Sensing Symposium (IGARSS'S9j and the 12'"Canadian Symposium on Remote Sensing, Vancouver, British Columbia,Canada, Vol. 4, pp. 2069-2072.

Boardman, J.w' 1990. "Inversion of High Spectral Resolution Data",Proceedings of SPIE Conference on Imaging Spectrometry of the TerrestrialEnvironment, Orlando, Florida, U.S.A., Vol. 1298, pp. 222-233.

Farrand, w'H., and C. Harsanyi. 1997. "Mapping of Mine Tailings in theCoeur d' Alene River Valley, Idaho, Through the Use of Constrained EnergyMinimization Technique", Remote Sensing ofEnvironment, Vol. 59, pp. 64-76.

Feasby, G., and R.K. Jones. 1994. "Report of Results of a Workshop on MineReclamation", IGWG - Industry Task Force on Mine Reclamation Workshop,Toronto, Ontario, March 10-11, 18 pages.

Dow

nloa

ded

by [

Col

orad

o C

olle

ge]

at 1

8:17

29

Oct

ober

201

4

\01. 26, "0. J, .lune/juin 2000

Ferrier, G, 1999. "Application of Imaging Spectrometer Data in IdentifyingEnvironmental Pollution Caused by Mining at Rodaquilar, Spain", RemoteSensing ofEnvironment, 68: 125-1 J 7.

Hossner, I..R. 1988. "Reclamation of Surface-Mined Lands", CRC PressInc., Boca Raton, Florida, U.S.A., Vol. I, 189 pages.

Levesque, J., Szeredi, '1'., Staenz, K., Singhroy, V., and Bolton, D. 1997."Spectral Unmixing for Monitoring Mine Tailings Site Rehabilitation, CopperCliff Mine, Sudbury, Ontario", Proceedings ofthe Twelfth InternationalConference and Workshops on Applied Geological Remote Sensing, Denver,Colorado, U.S.A., Vol. I, pp.340-347.

Levesque, J., K. Staenz, J. Shang, R.A. Neville, P. Yearwood, and VSinghroy. 1999. "Temporal Monitoring of Mine Tailings Revegetation UsingHyperspectral Data, Sudbury, Ontario", Proceedings of the ThirteenthInternational Conference on Applied Geologic Remote Sensing, Vancouver,British Columbia, Canada, Vol. 2, pp. 21-28.

Inco, 1998: http://www.incoltd.com/invest/annrpt/96-envlg.htm.

Mueller, A., P. Strobl, F. Lehmann, and G Reinhaeckel, 1997. "Case Studiesof Airborne Remote Sensing for the Assessment of Mining Impacts",Proceedings ofthe Third International Airborne Remote Sensing Conferenceand Exhibition. Copenhagen, Denmark, pp. 1-257-226.

Press, W.lI., S.A. Teukolsky, W.T. Vetterling, and B.P. Flannery. 1992."Numerical Recipes in C", Cambridge University Press, Cambridge, England,994 pages.

Shang, J., J. Levesque, PJ. Howarth, B. Morris, K. Staenz, and P. Yearwood.1999. "Preliminary Investigation of Acid Mine Drainage Detection Using cas;Data, Copper Cliff, Ontario, Canada", Proceedings ofthe Fourth InternationalAirborne Remote Sensing Conference and the 21st Canadian Symposium onRemote Sensing, Ottawa, Ontario, Canada, pp. 1-771-778.

Singhroy, VII. 1996. "Environmental and Geological Site Characterizationin Vegetated Areas: Image Enhancement Guidelines", Remote Sensing and GIS[or Site Characterization: Applications and Standards, eds. VH. Singhroy, D.P.Nebert, and A.1. Johnson, ASTM STP 1279, West Coushocken, Pennsylvania,U.S.A., pp. 5-16.

Space Imaging, 1999: http://www.spaceimaging.com.

Staenz, K. 1996. "Classification of a Hyperspectral Agricultural Data Setusing Band Moments for Reduction of the Spectral Dimensionality", CanadianJournal ofRemote Sensing, 22(3): 248-257.

Staenz, K., and DJ. Williams. 1997. "Retrieval of Surface Reflectance fromHyperspectral Data Using a Look-Up Table Approach", Canadian Journal (JfRemote Sensing, 23(4): 354-368.

Staenz, K., 'I'. Szeredi, and J. Schwarz. 1998. "ISDAS - A System forProcessing/Analyzing Hyperspectral Data", Canadian Journal ol RemoteSensing, 24(2): 99-113.

Staenz, K., Neville, R.A., Levesque, J., Szeredi, '1'., Singhroy, V, Borstad,GA., and Hauff, P. 1999. "Evaluation of cas; and SFSI Hyperspectral Data forGeological Applications .- Two Case Studies", Canadian Journal of RemoteSensing. Vol. 25, No.5, pp. 311-322.

Swayze, CiA., R.N. Clark, R.M. Pearson, and E. Livo. 1996. "MappingAcid-Generating Minerals at the California Gulch Superfund Site in Leadville,Colorado using Imaging Spectroscopy", Summaries ofthe Sixth Annual JPLAirborne Earth Science Workshop, JPL Publication 96-4, Pasadena, California,U.S.A., Vol. I, pp. 231-234,

Szeredi, '1'., Staenz, K., and Neville, R.A. 2000. "Automated EndmemberSelection: Par! I Theory", Remote Sensing of Environment (submitted).

Dow

nloa

ded

by [

Col

orad

o C

olle

ge]

at 1

8:17

29

Oct

ober

201

4