Ui

Va

b

a

ARA

KApMMEM

1

sMq

CcMfisRtasi

R

0h

International Journal of Applied Earth Observation and Geoinformation 28 (2014) 28–42

Contents lists available at ScienceDirect

International Journal of Applied Earth Observation andGeoinformation

jo ur nal home p age: www.elsev ier .com/ locate / jag

sing multiple spectral feature analysis for quantitative pH mappingn a mining environment

eronika Kopackováa,b,∗

Czech Geological Survey, Klárov 3, Prague 1 118 21, Czech RepublicCharles University in Prague, Faculty of Science, Department of Applied Geoinformatics and Cartography, Albertov 6, Prague 2 128 43, Czech Republic

r t i c l e i n f o

rticle history:eceived 13 June 2013ccepted 30 October 2013

eywords:cid mine drainage (AMD)H modelingineral spectroscopyining impacts

nvironmental monitoringultiple spectral feature fitting

a b s t r a c t

The pH is one of the major chemical parameters affecting the results of remediation programs carried outat abandoned mines and dumps and one of the major parameters controlling heavy metal mobilizationand speciation. This study is concerned with testing the feasibility of estimating surface pH on the basisof airborne hyperspectral (HS) data (HyMap). The work was carried on the Sokolov lignite mine, asit represents a site with extreme material heterogeneity and high pH gradients. First, a geochemicalconceptual model of the site was defined. Pyrite, jarosite or lignite were the diagnostic minerals of verylow pH (<3.0), jarosite in association with goethite indicated increased pH (3.0–6.5) and goethite alonecharacterized nearly neutral or higher pH (>6.5). It was found that these minerals have absorption featureparameters which are common for both forms, individual minerals as well as parts of the mixtures,while the shift to longer wavelengths of the absorption maximum centered between 0.90 and 1.00 �mis the main parameter that allows differentiation among the ferric minerals. The multi range spectralfeature fitting (MRSFF) technique was employed to map the defined end-members indicating certain pHranges in the HS image datasets. This technique was found to be sensitive enough to assess differences

in the desired spectral parameters (e.g., absorption shape, depth and indirectly maximum absorptionwavelength position). Furthermore, the regression model using the fit images, the results of MRSFF, asinputs was constructed (R2 = 0.61, Rv2 = 0.76) to estimate the surface pH. This study represents one of thefew approaches employing image spectroscopy for quantitative pH modeling in a mining environmentand the achieved results demonstrate the potential application of hyperspectral remote sensing as anefficient method for environmental monitoring.

. Introduction

Mining activities, both underground and open cast mining, are

till associated with many environmental problems such as Acidine Drainage (AMD) (Akcil and Koldas, 2006), generation of large

uantities of toxic substances (Kemper and Sommer, 2002) and

Abbreviations: AMD, acid mine drainage; AOT, atmospheric optical thickness;R, continuum removal; FWHM, full-width half-maximum; HCRF, hemispherical-onical reflectance factor; IS, image spectroscopy; LSU, linear spectral unmixing;NF, minimum noise fraction transformation; MRSFF, multi range spectral feature

tting; MTMF, mixture-tuned matched filtering; PLSR, partial least square regres-ion; PPI, pixel purity index; R2, coefficient of determination (training dataset);v2, coefficient of determination (validation dataset); Rad/Ref, at-sensor radiance-o-reflectance ratio; RRDF, radiance-to-reflectance difference factor; SAM, spectralngle mapper; SFF, spectral feature fitting; SMA, spectral mixture analysis; SVC,upervised vicarious calibration; SWIR, short-wave infrared; VNIR, visible nearnfrared; WV, water vapor content; XRD, X-Ray diffraction.∗ Correspondence to: Czech Geological Survey, Klárov 3, Prague 1 118 21, Czechepublic. Tel.: +420 257089481; fax: +420 257531376.

E-mail address: veronika.kopackova@seznam.cz

303-2434/$ – see front matter © 2013 Elsevier B.V. All rights reserved.ttp://dx.doi.org/10.1016/j.jag.2013.10.008

© 2013 Elsevier B.V. All rights reserved.

consequent release of heavy metals into the environment (Gomesand Favas, 2006). As AMD can severely contaminate surface andgroundwater, as well as soils, these anthropogenic activities canhave serious human health and ecological implications (Grimaltet al., 1999; Grimalt and MacPherson, 1999) if the mines are notmonitored and the necessary environmental treatment is not inplace.

AMD release from mine waste rock, tailings and mine structures,such as pits and underground workings, is primarily a function ofthe mineralogy of local rock material (mainly secondary miner-als associated with sulfide-bearing material) and the availabilityof water and oxygen. The typical AMD pattern leads to accumu-lation of Fe sulfates, oxy-hydroxides and oxides in a spatial andtemporal sequence that represents the buffering of an acidic solu-tion as it moves away from its source (Montero et al., 2005; Swayzeet al., 2000). Therefore, these minerals can serve as pH indicators

(indicative minerals). Because mineralogy and the other factorsaffecting AMD formation are highly variable within a site as wellas from site-to-site, predicting the potential for AMD using con-ventional laboratory analysis can be exceedingly challenging and

arth O

cnede2

bearlt–i(dbt22tK2t2armt

-

-

-

-

2

2

ptowKaott

VFKqSapa

V. Kopacková / International Journal of Applied E

ostly. However, modern remote sensing has become a novel tool,ot only for detecting and quantifying geological materials (Plazat al., 2009; Van der Meer et al., 2012), but also for monitoringynamic processes and induced changes in physical/chemical prop-rties (Ben-Dor et al., 2009; Chabrillat et al., 2002; Escribano et al.,010; Haubrock et al., 2008; Kokaly et al., 2003).

In a mining environment, the use of multispectral imagery haseen effectively used to monitor environmental impacts (De Moraist al., 2012; He et al., 2009; Khalifa and Arnous, 2012; Matejíceknd Kopacková, 2010) as well as to detect AMD generating mate-ial (Kopacková et al., 2012a; Robbins et al., 2000). However, theow spectral resolution of multispectral imagery is a major limi-ation. On the other hand, data with very high spectral resolution

hereafter referred to as imaging spectroscopy (IS) data, whichs also known in the remote sensing community as hyperspectralHS) data, has been successfully used in earlier studies to detectiverse mining environmental factors. Reflectance spectroscopy,oth ground and image-based methods, has been successfully usedo locate acid-generating minerals at mine sites (Kopacková et al.,012b; Montero et al., 2005; Quental et al., 2013; Riaza et al., 2011a,011b; Richter et al., 2008, 2009; Swayze et al., 2000, 2006) ando determine heavy metal concentrations (Choe et al., 2008, 2009;emper and Sommer, 2002; Kopacková et al., 2011; Pandit et al.,010). However, very few studies have been published on quanti-ative pH mapping in a mining environment (e.g., Ong and Cudahy,002; Zabcic et al., 2009). Particularly the extreme heterogeneitynd the fact that the material is present in the form of mixturesather than pure minerals (Montero et al., 2005; Riaza et al., 2011a)ake quantitative pH mapping challenging. Therefore, the objec-

ives of this paper are to:

link the mineral, geochemical and spectral properties of the mate-rial at abandoned lignite mines and dumps;

find spectral parameters reflecting the pH conditions whichremain even if the minerals are present in the form of mixtures;

employ a spectral mapping method that allows identification ofthe indicative minerals (based on the above considerations) evenif present in mixtures and enable mapping of the pH spatial pat-terns using airborne multi-flight line hyperspectral data;

build a pH model and validate the estimated pH using groundtruth data.

. Material and methods

.1. Test site

The study was performed in the Sokolov basin in the westernart of the Czech Republic (Fig. 1), in a region affected by long-erm extensive lignite mining. The Sokolov basin, containing rocksf Oligocene to Miocene age, is 8–9 km wide and up to 36 km long,ith a total area of about 200 km2. The basin is limited by therusné Hory Fault (NNE-SSW trending) and is also characterized by

system of minor parallel faults, forming a significant tectonic zonef lithospheric extent (Ziegler, 1990). Another significant fault sys-em of the Ohre Rift consists in the faults running in the NNW–SSEo NW–SE direction (Rajchl et al., 2009).

The basement of the Sokolov Basin is formed of Variscan and pre-ariscan metamorphic complexes of the Eger, Erzgebirge, Slavkovorest, Thuring-Vogtland Crystalline Units and granitoids of thearlovy Vary Pluton. The upper portions of these rocks are fre-uently weathered to kaolinitic residue. The basal late Eocene

taré-Sedlo-Formation is formed of well-sorted fluvial sandstonesnd conglomerates and is overlain by a volcano-sedimentary com-lex up to 350 m thick, which contains three lignite seams with

variable sulfur (S) content: the Josef seam (up to 20 m thick,

bservation and Geoinformation 28 (2014) 28–42 29

4.58% S), the Anezka seam (5–12 m thick, 1.57% S) and Antonínseam (20–30 m thick, reaching up to 62 m, 0.91% S) (Rojík, 2004;Murad and Rojík, 2005). The brown coal (lignite) belongs amongcoal seams enriched in As (Yudovich and Ketris, 2005) and otherheavy metals, such as Cd, Ni, Cu, Zn, Pb (Bouska and Pesek, 1999).

Mining activities in this region have been documented since1642. However large-scale underground and surface mining opera-tions began only after 1870. In 2009 the Miocene Antonín seam wasmined in the deeper part of the basin in two opencast pits, Druzbaand Jirí, but only the Jiri mine is still active at the present time. Long-term open cast mining required the removal of up to 180 m thickoverburden (Cypris clays) which was stockpiled and replaced afterthe lignite was extracted. At the dumps, material consists mostlyof Cypris clays, which can be characterized as well-laminated clayswith different varieties of mineralogical composition: kaolinite,montmorillonite, illite with admixtures of Ca–Mg–Fe carbonates,sulphates, sulphides, analcite, Mg-micas and bitumen (Rojík, 2004).Due to the presence of S in the coal, the lignite mines both still activeand abandoned, are largely affected by acid mine drainage (AMD)(Kopacková et al., 2011, 2012a).

2.2. Data

2.2.1. Aerial HS image datasetsThe hyperspectral image data was acquired in 2009 (July 27)

during the HyEUROPE 2009 flight campaign using the HyMap(HyVista Corp., Australia) airborne imaging spectrometer. TheHyMap sensor records image data in 126 narrow spectral bandscovering the entire spectral interval between 0.450 and 2.480 �mspectral range with Full Width Half Maximum (FWHM) of 15 nmand ground field of view of 4 m. The resulting ground pixel res-olution of the image datasets was 5 m. In order to successfullypre-process the hyperspectral data, a supportive calibration andvalidation ground campaigns were organized simultaneously withthe HyMap data acquisition in 2009 and 2010. At the selectedhomogenous targets the ground measurements were acquired bythe ASD FieldSpec-3 spectroradiometer to properly calibrate aswell as validate the image data and to enable: (i) atmospheric cor-rection of the airborne hyperspectral images and (ii) retrieving atsurface reflectance values for the further verification. The selectedtargets met the following conditions: (i) spatial homogeneity fora minimum area of 5 × 5 image pixels and (ii) natural or artifi-cial nearly Lambertian ground surfaces. The hemispherical-conicalreflectance factor (HCRF) (Schaepman-Strub et al., 2006) was mea-sured for each reference target. Raw spectroradiometric data weretransformed into the HCRF using the calibrated white spectralonpanel. In addition, Microtops II Sunphotometer (Solar Light Comp.,USA) measurements were taken approximately every 30 s duringthe HyMap data acquisition. Data acquired by the Sunphotometerwere used for estimation of the actual atmospheric conditions (AOT– aerosol optical thickness; WV – water vapor content).

Nine individual HyMap stripes covered the entire area of theSokolov lignite basin (Fig. 1). The orientation and geometry ofthe HyMap strips followed the SW–NE orientation of the lignitebasin. This setting represented an optimal solution from the eco-nomic point of view; on the other hand, this setting (relativesolar azimuth at the acquisition hour was about 73◦) caused thatthe data suffered from strong cross-track illumination and bi-directional reflectance distribution function effects (Verrelst et al.,2008). Therefore, prior to atmospheric correction, the data had to bepreprocessed to minimize these effects. The specific pre-processingfocused on correcting the cross track illumination effect via (i)

for each image separately, calculating the distribution of gaseslocated in different spectral regions (O2: 760 nm, H2O: 930 and1140 nm, CO2: 2015 and 2060 nm) employing continuum removal(CR) and extraction of differences in gas absorption features across

30 V. Kopacková / International Journal of Applied Earth Observation and Geoinformation 28 (2014) 28–42

F orrectt

t(pficAsccsaft(rfbaB

sateccPtcc(

ig. 1. Test site: sampling/measuring points displayed on the HyMap 2009 data (chis figure legend, the reader is referred to the web version of the article.)

he image, (ii) calculating the polynomial coefficients for each gaseseach spectral region), (iii) interpolating between the calculatedolynomial coefficients for all the wavelengths (full spectral con-guration), and (iv) using interpolated polynomial coefficients toorrect differences across the image (for each image separately).fter the preprocessing described above, radiometric rectificationuggested by Brook and Ben Dor (2011) was employed. This pro-edure consisted of (i) selecting a master image for which theal-val data were acquired (image number 5), (ii) simulating “at-ensor-radiance” using MODTRAN for known ground targets (e.g.,sphalt, beach sand, concrete), (iii) averaging the image radianceor the same ground targets, (iv) comparing the modeled radianceo the image radiance and calculating the radiometric indicatorsat-sensor radiance-to-reflectance ratio (Rad/Ref) and radiance-to-eflectance difference factor (RRDF); Brook and Ben Dor, 2011), (i)or the master image, calculating the coefficients for radiance recali-ration and then applying these coefficients to the slave images (see

method called supervised vicarious calibration (SVC); Brook anden Dor, 2011).

Final atmospheric correction was performed in the ATCOR-4oftware package (Richter, 2009). This software was designed fortmospheric correction of airborne hyperspectral image data usinghe MODTRAN 4 physical model of the atmosphere (Adler-Goldent al., 1999). The data obtained during the supportive groundampaign were used to improve the results of the atmosphericorrection. Direct ortho-georectification was performed using theARGE software package (Schläpfer, 1998). Finally, the hyperspec-

ral image data was georeferenced to the UTM 33N (WGS-84)oordinate system. To assess the final accuracy, the product wasompared to the very high spatial resolution aerial orthophotospixel size = 0.5 m) and a resulting standard positional accuracy of

ed reflectance, true color coding). (For interpretation of the references to color in

3.7 m was defined. Prior to image analysis, a vegetation mask wascomputed via thresholding the NDVI (Fig. 2), while the exposedsurfaces were mapped as the NDVI values falling below 0.4. Thisenabled excluding most of the surfaces covered by the vegetation.

2.2.2. Field data: material sampling and analysisOver 170 points (Fig. 1) distributed in still-active (Jirí, Medard

and Druzba) and abandoned (Lítov, Lomnice, Sylvestr and PVS– Podkrusnohorská vysypka) open-pit mines were documentedin the field during the 2008 and 2009 field campaigns. Particu-lar attention was paid to the abandoned mines with significantAMD-affected areas. At the sampling points spectroradiometricmeasurements were collected in natural illumination conditionsusing an ASD FieldSpec 3® portable spectroradiometer (in situ spec-troscopy measurements). The ASD instrument uses three separatespectrometers to measure the radiance between wavelengths of0.350 �m and 2.500 �m. Radiance spectra were normalized againsta 99% Spectralon® white reference to produce relative reflectancespectra for each measurement.

Samples of the surface material (0–2 cm depth) were collectedat 80 selected points. They were dried and sieved to <2 mm andthe abundance of trace elements including major heavy metalswas measured using a portable Innov-x Alpha RFA spectrometer.Furthermore, the samples were subjected to determination of lab-oratory sulphur (S total wt%), and total organic carbon (TOC %).The phase analysis of the studied samples was based on thepowder X-ray diffraction patterns (whole-rock random powder

samples and oriented clay fraction specimens) and voltammetry.The X-ray powder diffraction patterns were obtained on the PhilipsXıPert diffractometer using CuK� radiation and graphite secondarymonochromator. The whole-rock random patterns were collected

V. Kopacková / International Journal of Applied Earth Observation and Geoinformation 28 (2014) 28–42 31

were

iObaagamla1m(

bamsawu

aids

2

w

Fig. 2. The NDVI image, the exposed surfaces

n the angular range from 2◦ to 70◦ 2� with steps of 0.05◦ 2�.riented clay fraction specimens (fraction <2 �m) were preparedy the conventional sedimentation method according to Tannernd Jackson (1948). The oriented clay specimens were analyzeds dried in the air and after saturation for 10 h with ethylenelycol vapor at 60 ◦C. Their diffraction data were acquired in thengular range 2–50◦ 2� with steps of 0.05◦ 2�. Identification ofixed-layered minerals was performed by comparison of the ana-

yzed XRD patterns of ethylene glycolated orientated clay fractionnd modeled XRD patterns obtained by NEWMOD code (Reynolds,985). Goethite and pyrite minerals were identified using voltam-etry according to the procedure of Grygar (1996), Grygar et al.

2002), and Grygar and van Oorschot (2002).Attention was paid to the pH, especially to the relationships

etween the actual and laboratory pH and between the surfacend subsurface pH. The pH was measured in situ (23 measure-ents) using a pH-212 Voltcraft field pH meter. In addition, the

urface samples (0–2 cm) were collected in the field (60 samples)nd 15 subsurface samples (10 cm depth) were also taken. Theseere sieved (<2 mm) and the pH was determined in the laboratorysing an ion-selective electrode in a 1 M KCl solution.

Further spectra were obtained by measuring the sieved sampless well as the samples of all the minerals and facies encounteredn the Sokolov basin (Rojík, 2004) in artificial illumination con-itions, using the spectroradiometer’s contact probe (laboratorypectroscopy measurements).

.3. Methods

Firstly, the geochemical and mineral properties were linkedith the reflectance spectra acquired in the field and laboratory

mapped as the NDVI values falling below 0.4.

as well as with the reflectance of the corresponding HyMap pixels.The obtained spectra were normalized employing the continuumremoval (CR) method (Kruse et al., 1993). This method is a standardtransformation in the field of spectroscopy (Van der Meer, 2006), asit removes the continuum contribution from the reflectance spec-trum. Multiple absorption features within the VIS/NIR/SWIR regionwere enhanced after the normalization. Additionally, spectra wereconvolved to the HyMap spectral resolution using a Gaussian con-volution and the FWHM value for each band. The effect of mineralmixtures and heterogeneity on the spectral properties was studied.It was then possible to identify the spectral ranges and parame-ters of the indicative minerals which remain the same either at thelevel of pure minerals (laboratory spectra) or at the level of mineralmixtures (field and image spectra).

Prior to material mapping using HS images, it is necessary toderive the end member spectra for the fundamental physical com-ponents (mineral/organic constituents). Pure end-member spectracan be extracted either directly from the image pixels or fromspectral libraries measured in the field. The use of reflectanceend-members from spectral libraries can be problematic becausethey can suffer mainly from spatial and temporal variability in thereflectance properties of the cover types (Asner and Heidebrecht,2002). On the other hand, using end-members from spectrallibraries also has some benefits. Their main advantage is that theyfulfill the purity criterion better than the end-members derivedfrom an image, in other words more “pure” end-members thanthose measured in situ cannot be found within an image. Further-

more, their composition is known and, in addition, the standardprocedure used for deriving image end-members consists of severaltime consuming steps: the minimum noise fraction transforma-tion (MNF) (Boardman and Kruse, 1994; Green et al., 1988) and

32 V. Kopacková / International Journal of Applied Earth Observation and Geoinformation 28 (2014) 28–42

F and

c

plosf

pat((afwAt(Micatsfrfi

a

ig. 3. Homogenous targets used for training and validation. Both the calibrationharacterizing the studied sites.

ixel purity index (PPI) calculation (Boardman, 1995). Nonethe-ess, success still depends on individual skills and the experiencef the expert. For the reasons described above, it was preferred toelect representative field spectra end-members and to use theseor further mapping using HyMap image data.

There are many commonly used analytical techniques for map-ing the target material in hyperspectral images such as: spectralngle mapping (SAM) (Kruse et al., 1993), spectral feature fit-ing (SFF) (Clark et al., 1991), spectral mixture analysis (SMA)Adams et al., 1995) or mixture-tuned matched filtering (MTMF)Boardman, 1998). Each of these techniques has some advantagesnd disadvantages over the others and this aspect is discussedurther in this section. However, in order to model the pH, itas necessary to identify the specific minerals typical for theMD patterns (mainly Fe sulfates and iron oxy-hydroxides). All

hese minerals exhibit diagnostic absorptions before 1.000 �mKopacková et al., 2012a; Montero et al., 2005; Murphy and

onteiro, 2013; Swayze et al., 2000, 2006). The wavelengths andntensities of the absorption features depend on the nature of therystal field around the Fe atom and on the nature of the bondsround it, because the nature of the magnetic coupling betweenhe Fe3+ ions facilitates the transition of electrons between energytates (Sherman and Waite, 1985). Thus, in Fe3+ minerals, subtle dif-erences in the shapes and wavelengths of the absorption features

eflect the crystal structure of the minerals and allow their identi-cation, such parameters should be used for the spectral mapping.

The HS image data differ from the field spectral measurements,s they have lower spectral resolution; moreover, the measured

validation targets were selected in the way to cover the high mineral variability

spatial domain is also different. Pixel reflectance in heterogeneousenvironment has a significant mixing problem as it is a result ofspectral reflectance of different materials present within the pixel;on the other hand, filed spectra tend to represent rather pointmeasurements. Therefore, it was necessary to select a mappingtechnique/method that is scalable, sensitive to subtle changes inabsorption feature parameters (e.g., absorption shape and wave-length position) and also enables setting of specific spectral rangeswhere the characteristic absorptions of target minerals are exhib-ited.

The spectral feature fitting (SFF) technique is a method that wassuccessfully used to map minerals in multispectral/hyperspectralimage data (Debba et al., 2005; Haest et al., 2012; Mars and Rowan,2010). The advantage of SFF over other methods, such as spectralangle mapping (SAM), is that it is sensitive to subtle absorptionfeatures (Tangestani et al., 2011; Van der Meer, 2004) and also mini-mizes the influence of the effects of variations in mineral grain sizeand illumination (Kruse et al., 1993). The linear spectral unmixing(LSU, Settle and Drake, 1993) method takes into account the mixedcharacter of a pixel and allows identification of sub-componentsof the spectrum and determination of the abundance of differentmaterials for each pixel. However, in this specific case, when subtlechanges in diagnostic absorptions need to be detected, SFF also hasan advantage as the LSU method is sensitive primarily to albedo

and color differences (e.g., De Jong et al., 2011). The SFF method isavailable in ENVI software and compares the fit of the image spec-tra to the reference spectra using a least-squares technique (Clarket al., 1991). The reference spectra (whether the laboratory or field

V. Kopacková / International Journal of Applied Earth Observation and Geoinformation 28 (2014) 28–42 33

mmdeamfwt2sanm

spaNfemii

gadsiscddta

3

3

t

Fig. 4. Correlation between the pH measured in situ and in the laboratory.

easurements or extracted directly from the image) are scaled toatch the image spectra after the continuum is removed from both

ata sets. A least-squares fit is calculated band-by-band betweenach reference (end-member) spectra and the unknown spectra ofn image pixel (Clark et al., 1991; Van der Meer, 2004). The total rootean square error (RMSE) is used to generate an RMS error image

or each end-member. The output is represented by a “fit” image,hich is a measure of how well the unknown spectrum matches

he reference spectrum on a pixel-by-pixel basis (Van der Meer,004). This is a technique that allows assessing a whole absorptionhape. In case when the changing position of absorption maximumlso changes the shape on both sites of the absorption, this tech-ique can indirectly detect changes in the position of the absorptionaximum.An improved multiple feature-based technique, multi range

pectral feature fitting (MRSFF), was employed in this study. MRSFFrovides a promising classification technique as yielded higherccuracy than SAM (Judd and Steinberg, 2007) or than SAM andDVI (Pan et al., 2013). The user can choose the Multi Range SFF

unction to define multiple and different wavelength ranges aroundach end-member’s absorption features, which is very useful forineral identification. From this point of view, such an approach

s highly suitable for pH mapping, as specific mineral associationsndicating certain pH ranges exhibit multiple absorptions.

The ground truth data, pH measured for the homogenous tar-ets (3 × 3 pixel size: 15 m × 15 m in extent), were used to buildnd validate a pH model. The data were divided into two differentatasets (Fig. 3): (i) training (12 samples) and (ii) validation (14amples). Both the calibration and validation targets were selectedn the way to cover the high mineral variability characterizing thetudied sites. To estimate the pH, a multiple regression model wasonstructed between the end-member fit images and the trainingataset. The results were then statistically assessed using 14 vali-ation targets/samples and the coefficients of determination (Rv2)ogether with the Std. Error of the Estimate were defined (Fig. 10nd Table 3).

. Results

.1. The pH issue

The pH was measured in both ways, in situ and in the labora-ory. A strong correlation was found (R2 = 0.971) between these two

Fig. 5. Correlation between the surface (0–2 cm) and subsurface pH (10-cm depth).

datasets (Fig. 4). Only small deviations characterize the lower pHvalues, while samples with pH > 7 exhibited larger differences (>0.5on the pH scale) between the in situ and laboratory pH.

In addition, the relationship between the surface and sub-surface pH was studied on the example of the 15 collectedsurface/subsurface samples, and also statistically significant corre-lation was found (R2 = 0.760, Fig. 5), however the pH values higherthan 5 are limited to only one measurement in this dataset.

3.2. Linking the pH with mineral and spectral properties

Spectroscopic AMD approaches are based on mapping of thoseminerals that occur on the surface of waste-rock piles and theirsurroundings, focusing on minerals that serve as indicators of sub-aerial oxidation of pyrite (‘hot spots’) and the subsequent formationof AMD (Fe sulfates, oxy-hydroxides, and oxides accumulating in aspatial and temporal sequences, Montero et al., 2005; Swayze et al.,2000). However, the concept is more complex at lignite open-pitmines, as low pH values also characterize organic material repre-sented by lignite and its weathering products (Kopacková et al.,2012a).

The results of the chemical/mineral analysis were studied indetail and a site-specific conceptual model describing the relation-ship between the mineral composition and the pH is presentedin Fig. 6. The pH of the studied samples varied between 2.3 and8.5. Low pH (<3.0) characterized the material containing pyrite,jarosite or lignite, which were present either individually or asa part of mixtures. The coal seams in Sokolov vary in S con-tent (from rather low to high) and pyrite is a common sulphidemineral which can be found in fresh coal, as well as in its weath-ered products (e.g., secondary oxyhumolites) or coal/clay mixtures.Therefore, this mineral association diagnoses an early stage of sul-phide weathering, which includes a chain of chemical processessuch as oxidation and subsequent hydrolysis (Morris et al., 1985).Under still very acidic conditions (pH < 3), jarosite is formed ashydrated Fe2+ sulphates are exposed to the atmosphere in the sur-face material (Murad and Rojík, 2005).

When the pH increased (3.0 < pH < 6.5), jarosite was alwayspresent in association with goethite. Unlike reported by Monteroet al. (2005), goethite was present throughout a wide pH range

(pH between 3.0 and 8.5); however if together with jarosite, thiscorresponded to acid to nearly neutral pH (pH between 3.0 and6.5). Jarosite is a metastable form transforming to goethite withinmonths to years depending on the changing physical chemical

34 V. Kopacková / International Journal of Applied Earth O

cTtt

tTit(kdia

mowctpdmto

daohcu

h

Fig. 6. Mineral conceptual model: Sokolov case study.

onditions as well as the local mineralogy (Murad and Rojík, 2005).hus, this mineral association represents a transition stage whenhe metastable mineral is being transformed to its stable form overime as the pH increases.

Once the pH exceeded 6.5, jarosite was no longer present inhe material, and goethite was the most common ferric mineral.his characterizes an environment with stable physical and chem-cal conditions that allowed all the metastable Fe3+ sulphates to beransformed over time into the final stable form of goethite. Claysmainly kaolinite) were the most frequent minerals present in allind of abundances and mixtures throughout the whole pH andid not indicate any specific pH conditions. Based on these find-

ngs, further investigations were focused on minerals or mineralssociations described above.

The most frequent mineral constituents typical for diverseaterial type sorted by pH are summarized in Table 1. Not all

f the minerals identified by XRD exhibit diagnostic absorptionithin the VNIR-SWIR spectral region (e.g., feldspar, quarts) and

an be mapped using optical data. Nevertheless, the fundamen-al pH-indicative minerals can be identified using their reflectanceroperties, except for pyrite. Pyrite is not stable and quickly oxi-izes when it reaches the surface, where it is replaced by secondaryinerals (e.g., hydroxysulfates and oxy-hydroxides). Consequently,

his mineral would not be detectable by optical methods as theynly allow analysis of the surface.

The spectral properties of selected mineral constituents areepicted in Fig. 7. Clearly, pure minerals measured in the lab (Fig. 7And B) exhibit multiple diagnostic absorption features through-ut the whole spectral range (VNIR/SWIR). These absorptions haveigher separability compared to the spectra of the same mineral

onstituents present as mixtures (field spectra acquired in the fieldnder natural conditions) (Fig. 7C and D).

Secondary minerals with Fe3+ (hydroxysulfates and oxy-ydroxides) exhibit absorption features around 0.45 �m and before

bservation and Geoinformation 28 (2014) 28–42

1.00 �m (Clark et al., 1990; Montero et al., 2005; Murphy andMonteiro, 2013). If present as pure minerals (Fig. 7A), goethite showa strong and rather wide absorption centered at around 0.500 �m,on the other hand the diagnostic absorption for jarosite is narrowand centered at slightly shorter wavelength (around 0.450 �m).The absorption feature centered between 0.900 and 1.000 �m dif-fers for these two minerals as well. The absorption maximum isshifted, reflecting differences in crystal structure (Sherman andWaite, 1985), goethite absorption maximum is centered at longerwavelengths (closer to 1.000 �m) while jarosite exibit maximumabsorption at shorter wavelengths (closer to 0.900 �m). The sametrend in the shift of the absorption maximum wavelength remainsfor jarosite and goethite even if they are present as mixtures(Fig. 7C). This shift also reflect the changes in quantities betweenthese two minerals as, with increasing goethite content, the absorp-tion wavelength shifts to longer wavelengths (Fig. 7C). In addition,the absorptions of these minerals, in either pure or mixed form,not only differ in the positions of the absorption maximum, butalso shapes of the absorption feature centered between 0.900 and1.000 �m are different for jarosite and goethite. Jarosite, if presentin a pure form, exhibits an additional absorption feature around2.260 �m in the SWIR region (Fig. 7B). Material with high lig-nite content (>5%) exhibited a characteristic, very small slope ofthe spectral curve between 0.800 �m and 1.200 �m (Fig. 7A) withan absorption maximum at approximately 0.600 �m. Additionally,absorption at 1.700 and 2.309 �m indicating humic acid (Ben-Doret al., 1997) can be identified (Fig. 7B).

In terms of the predominant absorption in SWIR for mineralmixtures (Fig. 7D), the doublet between 2.100 and 2.200 �m causedby the vibration of the Al–OH molecule (Clark et al., 1990) indi-cates the presence of kaolinite. The jarosite absorption at 2.260 �mis overprinted by the kaolinite absorption if the mixtures containjarosite. The absorptions at 1.700 and 2.309 typical for the lignite-rich material remained; however, the absorption depth is smallerand the shape is simplified.

The field spectra of the mineral constituents described abovewere compared to the reflectance spectra of the corresponding pix-els of the HyMap images (Fig. 8). The same trend in the shift of theabsorption maximum wavelengths before 1.000 �m could be seeneven with the HyMap decreased spectral and spatial resolution;where the maximum is shifted even to the longer wavelengths.

3.3. Employing multi range spectral feature fitting

To select the end-members for spectral mapping, the fieldspectral libraries were assessed together with the results of XRDanalysis and grouped into (Fig. 9) (i) the spectral libraries of thefresh lignite which characterize the deep absorption at 0.640 �m(end-member 1); this material is characterized by very low pH val-ues (<3.0), (ii) the spectral libraries of clays rich in lignite whichstill exhibit the typical absorption at 0.640 �m (end-member 2);this material is also characterized by very low pH values (<3.0),(iii) the spectra of the material where high abundance of jarositewas identified and the absorption maximum of the absorption fea-ture centered between 0.900 and 1.000 �m was located at theshorter wavelengths (end-member 3); this material is character-ized by very low pH (<3.0), (iv) the libraries of materials where ahigh abundance of goethite was identified and the maximum ofthe absorption feature centered between 0.900 and 1.000 �m waslocated at the longer wavelengths (end-member 4); this materialis characterized by pH > 6.5, Aside these, additional group was cre-ated containing (v) the material where the secondary Fe minerals

represented a minimal fraction and where muscovite and chloritewere the major minerals present in the samples (end-member 5).

The spectral libraries were averaged to generate representa-tive spectra for each mineral group defined above (Fig. 6). These

V. Kopacková / International Journal of Applied Earth Observation and Geoinformation 28 (2014) 28–42 35

Table 1Most frequent mineral constituents typical for diverse material type sorted by pH (minerals detectable by the means of optical spectroscopy are in bold).

Minerals Quartz Muscovite Kaolinite Smektite Siderite Anatas Pyroxene k-Feldspar Pyrite Jarosite Goethite Lignite

Material: lignite-rich, verylow pH (<3)

× × × × × × × × × ×

Material: without lignite,very low pH (<3)

× × × × × × × × ×

Material: without lignite,pH between 3.0 and 6.5

× × × × × × ×

Material: without lignite,nearly neutral or higherpH (6.5–8.5)

× × × × × × × ×

Fig. 7. Spectral plots for the typical mineral constituents (A, B: pure minerals; C, D: depicted minerals present in mixtures; arrows pointing at the absorption maximumwavelengths which are characteristic for the depicted minerals).

Table 2Different scenarios tested under the MRSFF analysis (scenario 6 achieving the best result is in bold).

End-member EM1 EM2 EM3 EM4 EM5

Range 0.460–1.200 �mScenario 1 × × ×Range 0.460–1.200 �m; 2.080–2.400 �mScenario 2 × × ×Range 0.460–1.200 �mScenario 3 × × × × ×Range 0.460–1.200 �m; 2.080–2.400 �mScenario 4 × × × × ×Range 0.460–0.780 �m; 0.780–1.200 �m; 2.08–2.400 �mScenario 5 × × × × ×

Range 0.460–0.780 �mScenario 6 × ×

; 0.780–1.200 �m; 1.600–1.790 �m; 2.080–2.400 �m× × ×

36 V. Kopacková / International Journal of Applied Earth Observation and Geoinformation 28 (2014) 28–42

Fig. 8. Trend in the shift of the absorption maximum, the field spectra were compared to the reflectance spectra of the corresponding pixels of the HyMap images.

Fig. 9. The a representative spectra (end-members, resampled to the HyMap dataspectral resolution) used for multi range spectral feature fitting (MRSFF): the freshlignite which characterize the deep absorption at 0.640 �m (end-member 1); claysrich in lignite which still exhibit the typical absorption at 0.640 �m (end-member2); the material where high abundance of jarosite was identified and the absorptionmaximum of the absorption feature centered between 0.900 and 1.000 �m waslocated at the shorter wavelengths (end-member 3), the materials where a highabundance of goethite was identified and the maximum of the absorption featurecentered between 0.900 and 1.000 �m was located at the longer wavelengths (end-member 4); the material where the secondary Fe minerals represented a minimalfs

rpfmwa

iwa

Table 3Prediction statistics for the scenario 6 (add Table 2).

R R2 Adjusted R2 Std. error of the estimate Sig.

Training0.779 0.606 0.567 1.140 0.003

raction and where muscovite and chlorite were the major minerals present in theamples (end-member 5).

epresented the end/members further employed for spectral map-ing using MRSFF. The end-member fit images were derived andurther statistically assessed to test whether acceptable regression

odels can be obtained to model the surface pH. Different scenariosere tested including different numbers of end-members as well

s setting different spectral ranges (Table 2).The best result in identification of selected minerals as well as

n predicting the surface pH (R2 = 0.63, Rv2 = 0.77) was achievedhen all five end-members were included (scenario 6, Fig. 10)

nd when the spectral ranges were defined separately for the

Validation0.873 0.763 0.744 1.138 0.000

diagnostic absorptions between 0.460–0.780 �m, 0.780–1.200 �m,1.600–1.790 �m and 2.080–2.400 �m (Table 2). As the pH canpotentially vary between 0 and 14, the unrealistic pH values (outof the 0–14 range) were masked out in the final map. This issue isfurther elaborated in Section 4.

4. Discussion

In this study, the pH is quantitatively estimated using hyper-spectral image data (HyMap). Good performance of hyperspectralimage analysis depends on accurate atmospheric correction (Brookand Ben Dor, 2011; Pan et al., 2013), which has a strong influ-ence on the mineralogical spectral diagnoses (Riaza et al., 2011a,2011b). Especially the shorter wavelengths are directly affected byscattering of small particles in the atmosphere. The HS image dataused in this study were corrected for the BRDF and atmosphericeffects, and radiometric rectification suggested by Brook and BenDor (2011) was also implemented. The ground truth data (spectro-scopic in situ measurements of the known targets) were utilized(i) to model the at-sensor radiance (calibration dataset) in orderto facilitate the radiometric rectification, and (ii) to check the cor-rected image reflectance (validation dataset). As a result, the imagereflectance differed from the surface reflectance by values from2.5% (dark reference targets approx.10%) to 9.0% (bright referencetargets approx. 70%). From this point of view, the data were foundto be suitable for further analysis. In addition, as it was described forthe pH prediction, different scenarios were tested while different

spectral ranges were included in the analysis. The best result wasachieved when the diverse spectral regions were set including theshort-wavelength regions (e.g., 0.460–0.780 �m, 0.780–1.200 �m);

V. Kopacková / International Journal of Applied Earth Observation and Geoinformation 28 (2014) 28–42 37

Fig. 10. pH training/validation regression models.

Fig. 11. (A) Color infrared (CIR) image (HyMap 2009), (B) mask used for the further image analysis: the exposed surfaces were mapped as the NDVI values falling below0.4, (C) the end-member spectra are compared to the typical spectra of a vegetation mixed pixel, which exhibited an NDVI value below 0.4 but some minor vegetationfraction/cover was still present, (D) estimated pH.

3 arth O

tc

NeFt0Cstsfef

8 V. Kopacková / International Journal of Applied E

hus this also indirectly indicates that the data were successfullyorrected for the atmospheric effect.

In the pre-processing stage, vegetation was masked out using anDVI mask (Fig. 2); however, some mixed pixels with minor veg-tation abundances may have remained in the analyzed area. Inig. 11 the end-member spectra are compared to the typical spec-ra of such a mixed pixel, which exhibited an NDVI value below.4 where some minor vegetation fraction/cover was still present.learly, the mixed pixel spectrum differs significantly from thepectrum of the end-members used for spectral mapping and fur-her image analysis. Significantly lower fit values were obtained foruch pixels. Consequently, the estimated pH values for these pixels

ell below 0 in most cases; thus the majority of these pixels wereliminated when pH values outside the 0–14 range were excludedrom the final pH map. In general, this also happened with the other

Fig. 12. Example of the thematic output (A: end-member fit images were thresho

bservation and Geoinformation 28 (2014) 28–42

surfaces whose spectra significantly differ from the end-memberspectra (e.g., water bodies, diverse artificial ponds, wet areas orlithologies with high silica content). This could be an advantageof the described method at one site, but at the other site it needsto be ensured that all the desired fundamental end-members areincluded in the analysis.

It is important to emphasize that the optical data such as theHyMap are limited to the surface analysis only as they measureapproximately the upper 50 �m of the ground surface (Buckinghamand Sommer, 1983). The pH measured in situ was probably themost representative real-time pH for the thin surface layer whichis detected by optical methods. The in situ pH measurements were

compared to the pH determined in the laboratory where analyticalprecision was achieved; however, in this case, the measured sur-face samples were taken from a depth of 0–2 cm. In addition, the pH

lded to depict the pixels with the closest spectral match; B: estimated pH).

arth O

fdc2epactpIfiwmawAcev

ttpss0uitatwaa

stssr

V. Kopacková / International Journal of Applied E

or a limited number of subsurface samples (10-cm depth) was alsoetermined in the laboratory. In general, the in situ measurementsorresponded very well to the laboratory values for the pH range.3–5.0 (Fig. 4). For pH higher than 7.0, the in situ measurementsxhibit higher values (approximately higher about 1 unit on theH scale) than those measured in the laboratory. Feldspar, quartz,nd pyrite are generally higher in the coarser fraction, while totallay minerals, gypsum, and jarosite are higher in the finer frac-ion (Morkeh and Mclemore, 2012; Murad and Rojík, 2005). TheH values decrease from the coarse fraction to the fine fraction.

n general, the coarser fractions are less acid-generating than thener fractions (Morkeh and Mclemore, 2012). This could explainhy the sieved samples exhibited lower pH compared to the pHeasured in situ. A correlation was also found between the surface

nd subsurface pH values (Fig. 5); however only a limited datasetas available, especially samples representing pH > 5 were missing.lthough a correlation between the surface and subsurface samplesan be expected as the covering material was dumped in thick lay-rs at Sokolov, the match between the surface and subsurface pHalues still needs to be further studied.

Based on the observations in this study, the minerals that reflecthe specific site conditions and indicate a certain pH can be iden-ified by assessing the subtle differences in absorption featurearameters (e.g., absorption shape, maximum wavelength position,ymmetry, depth). For jarosite and goethite the same trend in thehift of the absorption maximum (the feature centered between.900 and 1.000 �m) is visible whether they are present individ-ally or as mixtures. The systematic shift also reflects changes

n quantities between these two minerals (Fig. 7C) and, in addi-ion, the shapes of the absorption feature centered between 0.900nd 1.000 �m are different. Lignite-rich material has a characteris-ic small slope of the spectral curve between 0.800 and 1.200 �mith an absorption maximum around 0.640 �m, and also exhibits

bsorptions centered at 1.700 and 2.309 �m. These characteristicslso remain the same even if lignite is present in mixtures.

Using hyperspectral data for mapping purposes, the key issuetill remains in selection of proper end-members. The field spec-ral libraries were used to map the indicative minerals. The field

pectra of five fundamental mineral groups defined under theite-specific conceptual model were averaged to derive the rep-esentative end-members, which were further utilized for the

Fig. 13. The Sokolov lignite

bservation and Geoinformation 28 (2014) 28–42 39

MRSFF. To succeed on this, it was necessary to acquire field spectrallibraries which reflect the all major mineral varieties of the site.Alternatively, the end-members could be extracted from hyper-spectral images. Although diverse techniques for this extractionexist, success still depends on the proper identification of pure pix-els (Zortea and Plaza, 2009). The spatial resolution of the HyMapdata was 5 m × 5 m and in such a heterogeneous environment, itcan be a difficult task to derive real pure pixels. The presentedapproach is built on systematic field work and field data collection,so the representative field spectra can be used for image spec-tral mapping. In such a case no procedure to extract pure pixels isrequested, moreover, the same end-members can be used if multi-date hyperspectral data are available, and thus this approach maybe more universal than, for instance, partial least squares regression(PLSR).

The multi range spectral feature fitting (MRSFF) techniquewas tested to identify the selected minerals or their associationsdefined under the conceptual model. For instance, this tech-nique was successfully used to map diverse vegetation types inYellowstone National Park (Kokaly et al., 2003). The vegetationalso differed in the shapes and depths of absorptions present inthe spectra and these were the key characteristics that enabledspecies mapping. Montero et al. (2005) used spectral librariesand the spectral feature technique for identification of Fe-bearingminerals, sheet and other silicates to study patterns representingthe evolution of acid solutions discharged from the pyritic wastepiles and the subsequent accumulation of secondary precipitates.Haest et al. (2012) employed the Multiple spectral feature fittingtechnique to identify and quantify minerals (iron(oxyhydr-)oxideand clay contents) in drill cores, achieving promising results(RMSE between 3.9 and 9.1 wt%). De Jong et al. (2011) employedthe Spectral Feature Fitting (SFF) and Linear Spectra Unmixing(LSU) techniques to map soil surface crusts and compared theresults. They judged that spectral unmixing was superior to spec-tral feature fitting. However, the main differences between crustedand non-crusted soils were in the overall albedo, brightness andshape and depth of the absorption feature at 2.200 �m. Unlike SFF,LSU is a method that is sensitive to albedo and color differences

and this may explain why better results were obtained whenthis technique was employed. The wavelength of the absorptionmaximum centered between 0.900 and 1.000 �m and its shift

basin: estimated pH.

4 arth O

wttm

nsm0aitTefiampsimmatsopc

hCatSdmtbmt≥rbsp

5

gmsrdtuitoml0tia

0 V. Kopacková / International Journal of Applied E

ere the key factors for mapping the indicative ferric minerals;his approach employing MRSFF, a method that is sensitive tohe absorption feature parameters (shape, depth and absorption

aximum wavelength), seems to be highly suitable.The MRSFF technique was successfully employed; this tech-

ique enables assessment of the whole absorption shape, for ferricecondary minerals, a changing position of the absorption maxi-um also changes the shape of the absorption centered between

.900 and 1.000 �m (Fig. 7A and C). Therefore, this technique canlso indirectly detect changes in the position of the absorption max-mum. These key spectral feature parameters enabled mapping ofhe indicative minerals of the mineral stack defined in this study.he pH was then estimated via application of the regression mod-ling to the fit images resulting from MRSFF analysis seems to beeasible (R2 = 0.63, Rv2 = 0.77). Fit images, where the pixel valuesndicate the closeness of the match between the pixel spectrumnd the end-member library spectrum, could be combined into the-atic maps (Fig. 12A) to identify, for instance, material with low

H. However, the quantitative pH map (Figs. 12B and 13) clearly hasignificant advantages over simple material identification. The pHs one of the most important chemical parameters governing heavy

etal mobility. Heavy metals are most mobile in an acid environ-ent and become least mobile when the pH approaches neutral. In

ddition, the pH is the most important factor that can cause vege-ation stress and senescence and thus it can be an impediment touccess in reclaiming work. In the future, this kind of map could bef potential use in detecting and monitoring mines or other anthro-ogenic surfaces and their dynamics (e.g., on-going physical andhemical processes).

Too few quantitative image-spectroscopy based approachesave been made in estimating the pH at mining sites. Ong andudahy (2002) employed PLSR to estimate the surface pH at thebandoned Brukunga pyrite mine, South Australia, using multi-emporal HyMap datasets acquired between 1998 and 2001.imilarly, Zabcic et al. (2009) used Hymap airborne hyperspectralata to generate predictive pH maps of AMD for the Sotiel-Migollasine complex, Southwest Spain, using PLSR analysis. Validation of

he model for an independent data set results in an R2 value of 0.71etween the actual and predicted pH values. Quental et al. (2013)apped material associated with acid AMD using HyMap data and

he final map displayed the mineralogical assemblage correlations0.8 of variable pH indicators, particularly pinpointing a low-pH

elationship to the contamination in the area. The limited num-er of such studies demonstrates that quantitative pH mapping istill a challenging task and the presented approach seems to beromising.

. Conclusions

The mining environment is characteristic for its high hetero-eneity and complexity. Therefore, Acid Mine Drainage (AMD)apping should be tailored to the specifics of the tested mining

ite. In this study, a conceptual model depicting the minerals thateflect the specific site conditions and indicate a certain pH was firstefined. It was found that pH < 3.0 characterized the material withhe presence of pyrite, jarosite or lignite whether present individ-ally or in mixtures. Jarosite in association with goethite indicated

ncreased pH (3.0–6.5), while goethite alone indicated nearly neu-ral or higher pH (>6.5). The spectral properties of these mineralsr their mineral associations were further analyzed and com-on absorption feature parameters were identified. The shift to

onger wavelengths of the absorption maximum centered between

.900 and 1.000 �m is the main parameter allowing differentia-ion among Fe3+ secondary minerals and this trend is still visiblef the minerals are part of mixtures. Lignite-rich material exhibits

characteristic small slope of the spectral curve between 0.800

bservation and Geoinformation 28 (2014) 28–42

and 1.200 �m with absorption maximum around 0.640 �m, andadditional minor absorptions at 1.700 and 2.309 �m.

Although the mapping concept presented in this study wasadjusted to the conditions of the local mining site, the conceptbased on detecting the subtle changes in the absorption parame-ters of the pH-indicating minerals (absorption shape and indirectlya shift in the absorption maximum) is generally transferable. Thespectra of mining site surfaces are too complex and there are anunlimited number of combinations of material mixtures and frac-tions. Furthermore, other influences such as the presence of crusts,vegetation or the atmosphere also alter the surface spectral prop-erties. Therefore, it is essential to simplify the concept and to focuson the parameters they remain the same at different spectral andspatial domains.

The multi range spectral feature fitting (MRSFF) technique,allowing setting of different spectral ranges for multiple diagnos-tic absorptions, was successfully employed to identify and map theindicative minerals specified above using the hyperspectral imagesand proved to be a sufficiently sensitive method for assessingthe desired spectral parameters (e.g., absorption shape, depth andindirectly maximum absorption wavelength position). Potentially,other technique that allows assessing the whole absorption shapeand maximum position could be used.

A multiple regression model using the fit images, i.e. the resultsof MRSFF, as inputs was constructed to estimate the surface pH(R2 = 0.61, Rv2 = 0.76). This study still represents one of the fewapproaches employing image spectroscopy for quantitative pHmodeling in a mining environment. As the results seem to bepromising, further testing and validation using multi-temporalhyperspectral data is planned.

Acknowledgements

The present research is being undertaken within the frameworkof Grants No. 205/09/1989 (HypSo: grant funded by the Czech Sci-ence Foundation) and No. 244242 (EO-MINERS: FP7 grant fundedby the EC). Many thanks are due to Dr. Petr Rojík (Sokolovskáuhelná a.s.) for his substantial assistance with the field campaign.I would also like to express my great thanks to my colleagues atthe Remote Sensing Laboratory at TAU, namely to Eyal Ben-Dor,Anna Brook, Gila Notesko and Ido Livne, for the HyMap 2009 datapre-processing.

References

Adams, J.B., Sabol, D.E., Kapos, V., Almeida, R., Roberts, D.A., Smith, M.O., Gille-spie, A.R., 1995. Classification of multispectral images based on fractionsof endmembers—application to land-cover change in the Brazilian Amazon.Remote Sensing of Environment 52 (2), 137–154.

Adler-Golden, S.M., Matthew, M.W., Bernstein, L.S., Levine, R.Y., Berk, A., Richtsmeier,S.C., Acharya, P.K., Anderson, G.P., Felde, G., Gardner, J., Hoke, M., Jeong, L.S.,Pukall, B., Ratkowski, A., Burke, H.H., 1999. Atmospheric correction for short-wave spectral imagery based on MODTRAN4. Imaging Spectrometry V 3753,61–69.

Akcil, A., Koldas, S., 2006. Acid mine drainage (AMD): causes, treatment and casestudies. Journal of Cleaner Production 14 (12–13), 1139–1145.

Asner, G.P., Heidebrecht, K.B., 2002. Spectral unmixing of vegetation, soil and drycarbon cover in arid regions: comparing multispectral and hyperspectral obser-vations. International Journal of Remote Sensing 23 (19), 3939–3958.

Ben-Dor, E., Chabrillat, S., Dematte, J.A.M., Taylor, G.R., Hill, J., Whiting, M.L., Sommer,S., 2009. Using imaging spectroscopy to study soil properties. Remote Sensingof Environment 113, S38–S55.

Ben-Dor, E., Inbar, Y., Chen, Y., 1997. The reflectance spectra of organic matter inthe visible near-infrared and short wave infrared region (400–2500 nm) duringa controlled decomposition process. Remote Sensing of Environment 61 (1),1–15.

Boardman, J.W., 1995. Analysis, understanding and visualization of hyperspectral

data as convex sets in n-space. Imaging Spectrometry 2480, 14–22.

Boardman, J.W., Kruse, F.A., 1994. Automated spectral analysis—a geological exam-ple using AVIRIS data, North Grapevine Mountains, Nevada. In: Proceedingsof the Tenth Thematic Conference on Geologic Remote Sensing—Exploration,Environment, and Engineering, vol. I, pp. I407–I418.

arth O

B

B

B

BC

C

C

C

C

D

D

D

E

G

G

G

G

G

G

G

H

H

H

J

K

K

V. Kopacková / International Journal of Applied E

oardman, J., 1998. Leveraging the high dimensionality of AVIRIS data forimproved sub-pixel target unmixing and rejection of false positives: mix-ture tuned matched filtering. In: Summaries of the Seventh Annual JPLAirborneGeoscience Workshop. Jet Propulsion Laboratory Publication,Pasadena, California, pp. 55.

ouska, V., Pesek, J., 1999. Quality parameters of lignite of the North Bohemian Basinin the Czech Republic in comparison with the world average lignite. Interna-tional Journal of Coal Geology 40 (2–3), 211–235.

rook, A., Ben Dor, E., 2011. Supervised vicarious calibration (SVC) of hyper-spectral remote-sensing data. Remote Sensing of Environment 115 (6),1543–1555.

uckingham, W.F., Sommer, S.E., 1983. Economic Geology 78, 664–674.habrillat, S., Goetz, A.F.H., Krosley, L., Olsen, H.W., 2002. Use of hyperspectral images

in the identification and mapping of expansive clay soils and the role of spatialresolution. Remote Sensing of Environment 82 (2–3), 431–445.

hoe, E., Kim, K.W., Bang, S., Yoon, I.H., Lee, K.Y., 2009. Qualitative analysis and map-ping of heavy metals in an abandoned Au–Ag mine area using NIR spectroscopy.Environmental Geology 58 (3), 477–482.

hoe, E., van der Meer, F., van Ruitenbeek, F., van der Werff, H., de Smeth, B., Kim,Y.W., 2008. Mapping of heavy metal pollution in stream sediments using com-bined geochemistry, field spectroscopy, and hyperspectral remote sensing: acase study of the Rodalquilar mining area, SE Spain. Remote Sensing of Environ-ment 112 (7), 3222–3233.

lark, R.N., King, T.V.V., Klejwa, M., Swayze, G.A., Vergo, N., 1990. High spectral res-olution reflectance spectroscopy of minerals. Journal of Geophysical Research –Solid Earth and Planets 95 (B8), 12653–12680.

lark, R., Swayze, G., Gallagher, A., Gorelick, N., Kruse, F.,1991. Mapping withimaging spectrometer data using the complete band shape least-squares algo-rithm simultaneously fit to multiple spectral features from multiple materials.In: Proceedings, 2nd Airborne Visible/Infrared Imaging Spectrometer (AVIRIS)Workshop. Jet Propulsion Laboratory Publication, Pasadena, California, pp.176–187.

ebba, P., van Ruitenbeek, F.J.A., van der Meer, F.D., Carranza, E.J.M., Stein, A., 2005.Optimal field sampling for targeting minerals using hyperspectral data. RemoteSensing of Environment 99 (4), 373–386.

e Jong, S.M., Addink, E.A., van Beek, L.P.H., Duijsings, D., 2011. Physical character-ization, spectral response and remotely sensed mapping of Mediterranean soilsurface crusts. Catena 86 (1), 24–35.

e Morais, M.C., Martins, P.P., Paradella, W.R., 2012. Multi-scale approach usingremote sensing images to characterize the iron deposit N1 influence areas inCarajas Mineral Province (Brazilian Amazon). Environmental Earth Sciences 66(7), 2085–2096.

scribano, P., Palacios-Orueta, A., Oyonarte, C., Chabrillat, S., 2010. Spec-tral properties and sources of variability of ecosystem components in aMediterranean semiarid environment. Journal of Arid Environments 74 (9),1041–1051.

omes, M.E.P., Favas, P.J.C., 2006. Mineralogical controls on mine drainage of theabandoned Ervedosa tin mine in north-eastern Portugal. Applied Geochemistry21 (8), 1322–1334.

reen, A.A., Berman, M., Switzer, P., Craig, M.D., 1988. A transformation for orderingmultispectral data in terms of image quality with implications for noise removal.IEEE Transactions on Geoscience and Remote Sensing 26 (1), 65–74.

rimalt, J.O., Ferrer, M., Macpherson, E., 1999. The mine tailing accident in Aznal-collar. Science of the Total Environment 242 (1–3), 3–11.

rygar, T., 1996. Electrochemical dissolution of iron(III) hydroxyoxides: moreinformation about the particles. Collection of Czechoslovak Chemical Commu-nications 61, 93–106.

rygar, T., Dedecek, J., Hradil, D., 2002. Analysis of low concentration of free ferricoxides in clays vis diffuse reflectance spectroscopy and voltametry. GeologicaCarpathica 53 (2), 71–77.

rygar, T., van Oorschot, I.H.M., 2002. Voltammetric identification of pedogenic ironoxides in paleosol and loess. Electroanalysis 14 (3), 339–344.

rimalt, J.O., MacPherson, E., 1999. Special issue: the environmental impact of themine tailing accident in Aznalcollar (south-west Spain). Science of the TotalEnvironment 242 (1–3), 1–2.

aest, M., Cudahy, T., Laukamp, C., Gregory, S., 2012. Quantitative mineralogy frominfrared spectroscopic data. I. Validation of mineral abundance and composi-tion scripts at the Rocklea channel iron deposit in Western Australia. EconomicGeology 107 (2), 209–228.

aubrock, S.N., Chabrillat, S., Kuhnert, M., Hostert, P., Kaufmann, H., 2008. Surfacesoil moisture quantification and validation based on hyperspectral data and fieldmeasurements. Journal of Applied Remote Sensing 2, 023552.

e, B., Oki, K., Wang, Y., Oki, T., 2009. Using remotely sensed imagery to estimatepotential annual pollutant loads in river basins. Water Science and Technology60 (8), 2009–2015.

udd, C., Steinberg, S., 2007. Mapping submerged macrophytes: using multi-rangespectral feature fitting to map submerged eelgrass in a turbid estuary. In:American Society for Photogrammetry and Remote Sensing—ASPRS Annual Con-ference, pp. 326–331.

emper, T., Sommer, S., 2002. Estimate of heavy metal contamination in soils after amining accident using reflectance spectroscopy. Environmental Science, Tech-

nology 36 (12), 2742–2747.

halifa, I.H., Arnous, M.O., 2012. Assessment of hazardous mine waste transport inwest central Sinai, using remote sensing and GIS approaches: a case study of UmBogma area, Egypt. Arabian Journal of Geosciences 5 (3), 407–420.

bservation and Geoinformation 28 (2014) 28–42 41

Kokaly, R.F., Despain, D.G., Clark, R.N., Livo, K.E., 2003. Mapping vegetation in Yel-lowstone National Park using spectral feature analysis of AVIRIS data. RemoteSensing of Environment 84 (3), 437–456.

Kopacková, V., Chevrel, S., Bourguignon, A., Rojik, P., 2012a. Application of highaltitude and ground-based spectroradiometry to mapping hazardous low-pHmaterial derived from the Sokolov open-pit mine. Journal of Maps 8 (3),220–230.

Kopacková, V., Chevrel, S., Bourguignon, A., Rojik, P., 2012b. Mapping hazardouslow-pH material in mining environment: multispectral and hyperspectralapproaches. In: 2012 IEEE International Geoscience and Remote Sensing Sym-posium (IGARSS), pp. 2695–2698.

Kopacková, V., Chevrel, S., Bourguignon, A., 2011. Spectroscopy as a tool for geo-chemical modeling. Proceedings of SPIE 8181, 818106.

Kruse, F.A., Lefkoff, A.B., Boardman, J.W., Heidebrecht, K.B., Shapiro, A.T., Barloon, P.J.,Goetz, A.F.H., 1993. The spectral image-processing system (SIPS)—interactivevisualization and analysis of imaging spectrometer data. Remote Sensing ofEnvironment 44 (2–3), 145–163.

Mars, J.C., Rowan, L.C., 2010. Spectral assessment of new ASTER SWIR surfacereflectance data products for spectroscopic mapping of rocks and minerals.Remote Sensing of Environment 114 (9), 2011–2025.

Matejícek, L., Kopacková, V., 2010. Changes in croplands as a result of large scalemining and the associated impact on food security studied using time-serieslandsat images. Remote Sensing 2 (6), 1463–1480.

Montero, I.C., Brimhall, G.H., Alpers, C.N., Swayze, G.A., 2005. Characterization ofwaste rock associated with acid drainage at the Penn Mine, California, by ground-based visible to short-wave infrared reflectance spectroscopy assisted by digitalmapping. Chemical Geology 215 (1–4), 453–472.

Morkeh, J., Mclemore, V.T., 2012. The effect of particle size fractions on chem-istry, mineralogy, and acid potential of the Questa Rock Piles, Taos County, NewMexico. Open-file Report 545, 1–96.

Murad, E., Rojik, P., 2005. Iron mineralogy of mine-drainage precipitates as environ-mental indicators: review of current concepts and a case study from the SokolovBasin, Czech Republic. Clay Minerals 40 (4), 427–440.

Murphy, R.J., Monteiro, S.T., 2013. Mapping the distribution of ferric iron miner-als on a vertical mine face using derivative analysis of hyperspectral imagery(430–970 nm). ISPRS Journal of Photogrammetry and Remote Sensing 75,29–39.

Morris, R., Lauer, H., Lawson, C., Gibson Jr., E., Nace, G., Stewart, C., 1985. Spectral andother physicochemical properties of submicron powders of hematite (a-Fe2O3),maghemite (g-Fe2O3), magnetite (Fe3O4), goethite (a-FeOOH), and lepidocrocite(g-FeOOH). Journal of Geophysical Research 90, 3126–3144.

Ong, C., Cudahy, T., 2002. Deriving quantitative monitoring data related toacid drainage using multitemporal hyperspectral data, Downloaded fromhttp://ees.elsevier.com/jag/automail query.asp (20.8.13).

Pan, Z., Huang, J., Wang, F., 2013. Multi range spectral feature fitting for hyperspectralimagery in extracting oilseed rape planting area. International Journal of AppliedEarth Observation and Geoinformation 25, 21–29.

Pandit, C.M., Filippelli, G.M., Li, L., 2010. Estimation of heavy-metal contamination insoil using reflectance spectroscopy and partial least-squares regression. Inter-national Journal of Remote Sensing 31 (15), 4111–4123.

Plaza, A., Benediktsson, J.A., Boardman, J.W., Brazile, J., Bruzzone, L., Camps-Valls,G., Chanussot, J., Fauvel, M., Gamba, P., Gualtieri, A., Marconcini, M., Tilton,J.C., Trianni, G., 2009. Recent advances in techniques for hyperspectral imageprocessing. Remote Sensing of Environment 113, S110–S122.

Quental, L., Sousa, A.J., Marsh, S., Abreu, M.M., 2013. Identification of materi-als related to acid mine drainage using multi-source spectra at S. DomingosMine, southeast Portugal. International Journal of Remote Sensing 34 (6),1928–1948.

Rajchl, M., Ulicny, D., Grygar, R., Mach, K., 2009. Evolution of basin architecturein an incipient continental rift: the Cenozoic Most Basin, Eger Graben (CentralEurope). Basin Research 21 (3), 269–294.

Reynolds Jr., R.C., 1985. NEWMOD©, a computer program for the calculation of one-dimensional diffraction patterns of mixed-layered clays. R. C. Reynolds, Jr., 8Brook Dr., Hanover, NH.

Riaza, A., Buzzi, J., Garcia-Melendez, E., Carrere, V., Muller, A., 2011a. Monitoring theextent of contamination from acid mine drainage in the Iberian Pyrite Belt (SWSpain) using hyperspectral imagery. Remote Sensing 3 (10), 2166–2186.

Riaza, A., Garcia-Melendez, E., Mueller, A., 2011b. Spectral identification of pyritemud weathering products: a field and laboratory evaluation. International Jour-nal of Remote Sensing 32 (1), 185–208.

Richter, N., Jarmer, T., Chabrillat, S., Oyonarte, C., Hostert, P., Kaufmann, H., 2009.Free iron oxide determination in Mediterranean soils using diffuse reflectancespectroscopy. Soil Science Society of America Journal 73 (1), 72–81.

Richter, N., Staenz, K., Kaufmann, H., 2008. Spectral unmixing of airborne hyper-spectral data for baseline mapping of mine tailings areas. International Journalof Remote Sensing 29 (13), 3937–3956.

Richter, R., 2009. Atmospheric/Topographic Correction for Airborne Imagery. DLR-German Aerospace Centre, Wessling, Germany.

Robbins, E.I., Anderson, J.E., Cravotta, C.A., Nord, G.L., Slonecker, E.T., 2000.Remotely-sensed multispectral reflectance variations in acidic versus near-neutral contaminated coal mine drainage in Pennsylvania. In: ICARD 2000,

Proceedings, vols. I and II, pp. 1551–1561.

Rojík, P., 2004. New stratigraphic subdivision of the tertiary in the Sokolov Basinin Northwestern Bohemia. Journal of the Czech Geological Society 49 (3–4),173–185.

4 arth O

S

S

S

S

S

S

T

T

154–157.

2 V. Kopacková / International Journal of Applied E

chaepman-Strub, G., Schaepman, M.E., Painter, T.H., Dangel, S., Martonchik, J.V.,2006. Reflectance quantities in optical remote sensing-definitions and case stud-ies. Remote Sensing of Environment 103 (1), 27–42.

chläpfer, D., 1998. Parametric Geocoding, PARGE Using Guide, Version 2.3. ReSeApplications Schläpfer, Remote Sensing Laboratories, University of Zurich.

ettle, J.J., Drake, N.A., 1993. Linear mixing and the estimation of ground coverproportions. International Journal of Remote Sensing 14 (6), 1159–1177.

herman, D.M., Waite, T.D., 1985. Electronic-spectra of Fe-3+ oxides and oxidehydroxides in the near Ir to near Uv. American Mineralogist 70 (11–12),1262–1269.

wayze, G.A., Livo, K.E., Clark, R.N., Hoefen, T.M., Higgins, C.T., Ong, C., Kokaly,R.F., Kruse, F.A., 2006. Evaluating minerals of environmental concern usingspectroscopy. In: 2006 IEEE International Geoscience and Remote Sensing Sym-posium, vols. 1–8, pp. 1990–1991.

wayze, G.A., Smith, K.S., Clark, R.N., Sutley, S.J., Pearson, R.M., Vance, J.S., Hageman,P.L., Briggs, P.H., Meier, A.L., Singleton, M.J., Roth, S., 2000. Using imaging spec-troscopy to map acidic mine waste. Environmental Science, Technology 34 (1),47–54.

anner, C.B., Jackson, M.L., 1948. Nomographs of sedimentation times for soil parti-cles under gravity or centrifugal acceleration. Soil Science Society of America –

Proceedings 12, 60–65.

angestani, M.H., Jaffari, L., Vincent, R.K., Sridhar, B.B.M., 2011. Spectral character-ization and ASTER-based lithological mapping of an ophiolite complex: a casestudy from Neyriz ophiolite, SW Iran. Remote Sensing of Environment 115 (9),2243–2254.

bservation and Geoinformation 28 (2014) 28–42

Van der Meer, F.D., Van der Werff, H.M.A., Van Ruitenbeek, F.J.A., Hecker, C.A.,Bakker, W.H., Noomen, M.F., Van der Meijde, M., Carranza, E.J.M., de Smeth, J.B.,Woldai, T., 2012. Multi- and hyperspectral geologic remote sensing: a review.International Journal of Applied Earth Observation and Geoinformation 14 (1),112–128.

Van der Meer, F., 2006. The effectiveness of spectral similarity measures for theanalysis of hyperspectral imagery. International Journal of Applied Earth Obser-vation and Geoinformation 8 (1), 3–17.

Van der Meer, F., 2004. Analysis of spectral absorption features in hyperspectralimagery. International Journal of Applied Earth Observation and Geoinformation5, 55–68.

Verrelst, J., Schaepman, M.E., Koetz, B., Kneubuhler, M., 2008. Angular sensitivityanalysis of vegetation indices derived from CHRIS/PROBA data. Remote Sensingof Environment 112 (5), 2341–2353.

Yudovich, Y.E., Ketris, M.P., 2005. Arsenic in coal: a review. International Journal ofCoal Geology 61 (3–4), 141–196.

Zabcic, N., Rivard, B., Ong, C., Muller, A., 2009. Using airborne hyperspectral data tocharacterize the surface Ph of pyrite mine tailings. In: 2009 First Workshop onHyperspectral Image and Signal Processing: Evolution in Remote Sensing, pp.

Ziegler, P.A., 1990. Collision related intra-plate compression deformations in West-ern and Central-Europe. Journal of Geodynamics 11 (4), 357–388.

Zortea, M., Plaza, A., 2009. Spatial preprocessing for endmember extraction. IEEETransactions on Geoscience and Remote Sensing 47 (8), 2679–2693.