Remote Sensing of Environment 113 (2009) 12081219

Contents lists available at ScienceDirect

Remote Sensing of Environment

j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / r s e

Mapping lithology of the Sarfartoq carbonatite complex, southern West Greenland, using HyMap imaging spectrometer data
Enton Bedini
Geological Survey of Denmark and Greenland (GEUS), ster Voldgade 10, 1350 Copenhagen K, Denmark Institute of Geography and Geology (IGG), ster Voldgade 10, 1350 Copenhagen K, Denmark

a r t i c l e

i n f o

a b s t r a c t
The Sarfartoq carbonatite complex occurs in the southern West Greenland in a transition zone between Archaean gneiss complex to the south and a Proterozoic mobile belt to the north. The Sarfartoq carbonatite complex consists of a core zone composed of dolomite carbonatite and minor svite (calcite carbonatite) surrounded by a fenite zone and a marginal zone of gneisses frequently altered due to hydrothermal activity. High spatial and spectral resolution imaging spectrometer data recorded by the HyMap imaging system were used to map lithology of the Sarfartoq carbonatite complex. A careful analysis of the spectral reectance properties of the carbonatite lithology preceded the HyMap data analysis stage. The spectral reectance measurements showed that the various lithologic units including dolomite carbonatite, svite, fenite and the marginal alteration zone have distinct spectral reectance characteristics. The analysis of the HyMap data was based on an unsupervised clustering algorithm, the Self Organizing Maps (SOM), for the mapping of the main lithology and a hierarchical tree for the mapping of sparsely occurring svite rocks. Spectral mixture analysis was applied to map fractional abundances and compare with the SOM results. The resulting lithological map shows the spatial distribution of dolomite carbonatite, svite, fenite with abundant carbonatite dykes (representing the outer core of the carbonatite complex), fenite and hematized gneiss (marginal alteration zone). The results compare well with the eld data collected for the assessment of the mapping accuracy and due to the spatially contiguous nature of the hyperspectral data could be used to better map the outcropping carbonatite lithology. The spectral reectance measurements and the mapping results provide information of petrological importance for the carbonatite core zone. 2009 Elsevier Inc. All rights reserved.

Article history: Received 9 November 2008 Received in revised form 10 February 2009 Accepted 10 February 2009 Keywords: Lithologic mapping Carbonatites Imaging spectrometry Greenland

1. Introduction The reectance spectrum in the 0.45- to 2.5 m wavelength region provides mineralogical information based on analysis of electronic absorption features in transitional metals, especially iron (Clark, 1999; Hunt, 1977), and of molecular absorption features in carbonate, hydrate and hydroxide minerals (Clark, 1999; Hunt, 1977). Imaging spectrometers or hyperspectral sensors measuring hundreds of spectral bands from aircraft and satellite platforms provide unique spatial/spectral datasets for analysis of surface mineralogy (Goetz et al., 1985; Kruse et al., 2003). Imaging spectrometry or hyperspectral remote sensing is a promising technology for assisting the geological mapping and mineral exploration studies of the remote Arctic terrains (e.g., Harris et al., 2005; Rivard & Arvidson, 1992). An extensive and complex suite of alkaline igneous rocks of carbonatitic and kimberlitic afnity is known to occur in the basement rocks of southern West Greenland (Larsen et al., 1983). One of the most important and major carbonatite intrusions is the Sarfartoq

carbonatite (Secher & Larsen, 1980). The Sarfartoq carbonatite complex consists of inner and outer carbonatite core zones, a fenite zone and a marginal alteration zone (Secher & Larsen, 1980). The interesting carbonatite lithology and its remote location in the Arctic latitudes of southern West Greenland make the Sarfartoq carbonatite complex a very good target to evaluate the use of imaging spectrometry for geological mapping of carbonatite complexes and in the Arctic environment of West Greenland. The purpose of this paper is to report on the application of high spatial and spectral resolution airborne imaging spectrometer data recorded by the HyMap imaging system to map lithology of the Sarfartoq carbonatite complex in southern West Greenland. 2. Geological setting The Sarfartoq carbonatite complex was emplaced in Neoproterozoic time in the transition zone between Archaean gneiss complex to the south and the Proterozoic Nagssugtoqidian mobile belt to the north, in southern West Greenland (Fig. 1A), (Larsen & Rex, 1992; Secher et al., in press; Secher, 1986; Secher & Larsen, 1980). The Sarfartoq carbonatite complex was discovered in 1975 as a result of regional airborne gamma-ray spectrometry survey carried out by the

E-mail address: enton_bedini@hotmail.com. 0034-4257/$ see front matter 2009 Elsevier Inc. All rights reserved. doi:10.1016/j.rse.2009.02.007

E. Bedini / Remote Sensing of Environment 113 (2009) 12081219

1209

Fig. 1. (A) Geographic and geologic position of the Sarfartoq carbonatite complex in southern West Greenland (modied from Allaart, 1982). (B) Geological map of the Sarfartoq carbonatite complex (modied from Secher, 1986). Used with permission from the Geological Survey of Denmark and Greenland. The rectangle indicates the approximate spatial extent of the HyMap data analyzed in this study.

Geological Survey of Greenland (Secher, 1976). The Sarfartoq carbonatite is an intrusive conical body with a core of carbonatite sheets surrounded by a marginal zone of hematized gneiss with carbonatite dykes (Secher & Larsen, 1980), (Fig. 1B). The carbonatite core zone consists of an inner core zone (N 50% carbonatite), an outer

core zone (b 50% carbonatite) and a narrow rim of fenite surrounding this (Fig. 1B). The predominating carbonatite type in the core zone is dolomite carbonatite. Svite rocks (calcite carbonatite) occur only sporadically as discrete layers in the inner core zone and in schlieren in the outer core zone (Secher & Larsen, 1980). The main minerals in the

1210

E. Bedini / Remote Sensing of Environment 113 (2009) 12081219

Fig. 2. Photo (facing north) of the outcropping part of the Sarfartoq carbonatite complex, along the Paradise Valley (Arnangarnup Qoorua), southern West Greenland. The height of the slope is ca. 400500 m. Photo by the author.

core zone are ferroan dolomite, ankerite, phlogopite, alkali amphibole, magnetite, and apatite. The most frequently found fenite is a light-grey aegirine bearing fenite. The marginal alteration zone consists of gneisses frequently altered due to hydrothermal activity and cataclastic deformation of varying intensity (Secher & Larsen, 1980). Locally large radioactive shear zones 50200 m wide are found oriented tangentially to the core. They consist of strongly limonitized, hematized, jointed and crushed gneisses (Secher, 1986). Most of the rocks of the carbonatite core zone are covered by glaciouvial sediments in the oor of the Paradise Valley (the Greenlandic name for this valley is Arnangarnup Qoorua). The peripheral parts of the carbonatite core are exposed in the slopes of the valley (Fig. 2). In this study the focus is at the outcropping part of the carbonatite complex north of the Paradise Valley. All recognized zones of the Sarfartoq carbonatite complex (except radioactive shear zones) are exposed for about 5 km along the slopes of the valley within the study area. 3. HyMap data The HyMap is an airborne imaging system developed by Integrated Spectronics, Sydney, Australia, and operated by HyVista Corporation. It consists of sensors located on a xed wing aircraft typically own at an altitude of 2.5 km. The sensors collect reected solar radiation in 126 bands covering the 0.452.5 m wavelength range, including the visible to near infrared (VNIR) and short-wave infrared (SWIR) regions of the electromagnetic spectrum (Table 1), (Cocks et al., 1998).

The Sarfartoq HyMap scenes are part of the HyperGreen-2002 campaign of the Geological Survey of Denmark and Greenland (Tukiainen & Thorning, 2005). The Sarfartoq HyMap scenes were recorded on 09.08.2002 in 126 narrow bands, from 0.44 to 2.48 m with 4 m nominal pixel size. This study is based on two HyMap subscenes covering the outcropping part of the Sarfartoq carbonatite complex. The spatial extent of the HyMap subscenes is shown mosaicked in Fig. 3. The Sarfartoq HyMap data were geometrically and atmospherically corrected using the software PARGE and ATCOR4. The parametric geocoding procedure (PARGE) reconstructs the scanning geometry for each image pixel using position, attitude and terrain elevation data (Schlpfer & Richter, 2002). The Airborne Atmospheric and Topographic Correction Model (ATCOR4) performs the combined atmospheric/topographic correction accounting for the angular and elevation dependence of the atmospheric correction functions and calculates surface reectance (solar spectral wavelength region) based on the geocoded and orthorectied imagery (Richter & Schlpfer, 2002). 4. Lithologic composition and spectral reectance properties Spectral reectance measurements were considered important to understand the spectral reectance of the carbonatite rocks and analyze the HyMap data. A collection of rock samples collected in the course of previous geological studies of the Sarfartoq carbonatite complex (Secher, 1986; Secher & Larsen, 1980; Nielsen, 1982) was acquired from the rock samples' archive of the Geological Survey of Denmark and Greenland. This made possible a detailed study of the spectral reectance properties of the carbonatite rocks before the eld activities. Reectance spectra of more than forty rock samples from the Sarfartoq carbonatite complex were recorded using an Analytical Spectral Device (ASD), which records 2151 channels within the 350to 2500 nm wavelength region. 4.1. Carbonatites Carbonatites by denition are igneous rocks that contain more than 50% modal carbonate minerals. Dolomite-rich carbonatites are called dolomite carbonatites whereas the term svite is used for coarse grained

Table 1 HyMap instrument specications (Cocks et al., 1998). Spectral module Wavelength range (m) Bandwidth (nm) Spectral sampling (nm) VIS NIR SWIR1 SWIR2 0.450.89 0.891.35 1.401.80 1.952.48 1516 1516 1516 1820 15 15 13 17

IFOV: 2.5 mrad along track; 2.0 mrad across track FOV: 60 (512 pixels) Swath: 2.3 km at 5 m IFOV; 4.6 km at 10 m IFOV IFOV, instantaneous-eld-of-view; FOV, eld-of-view; mrad, milliradian.

E. Bedini / Remote Sensing of Environment 113 (2009) 12081219

1211

Fig. 3. The spatial extent of the HyMap data used in this study, covering the outcropping part of the Sarfartoq carbonatite complex north of the Paradise Valley in the southern West Greenland. The boundaries of the geological map of the carbonatite complex are shown for reference with Fig. 1. White circles show the locations of the eld stations used to assess the accuracy of the mapping results. The rectangle (white dashed line) indicates the subset of the image analyzed using the spectral mixture analysis.

calcite carbonatites (Winter, 2001). Several studies have considered the spectral properties of carbonate minerals (Gaffey, 1985; Gaffey, 1986, 1987; Hunt,1977; Hunt & Salisbury,1971) in the wavelengths covered by most remote sensing instruments. A study on the spectral reectance of the carbonatite rocks has been reported by Rowan et al. (1986). Spectra of powders of carbonate minerals containing no transitional metal cations are nearly straight lines near unity reectance at wavelengths shorter than 1.6 m (Gaffey, 1986). At wavelengths N 1.6 m there is a series of absorptions due to vibrational processes of the carbonate anion (e.g., Gaffey, 1985). Gaffey (1986) observed that in general absorption bands in dolomite spectra are centered at shorter wavelengths than the equivalent absorption bands in calcite spectra. Reectance spectra of dolomite carbonatites and svites from the Sarfartoq carbonatite core zone are shown in Fig. 4A. The dolomite carbonatites (Fig. 4A, DLC1, DLC2, DLC3) display characteristic 2 absorption centered dolomite absorption features with the main CO 3 near 2.32 m. The reectance spectra DLC1 and DLC2 are characterized by broad ferrous (Fe2+) absorption feature in the 1.01.3 m region (Gaffey, 1986). Increase in intensity of this broad band is positively correlated with increasing Fe2+ content in dolomites or calcites (Gaffey, 1986). However, in the carbonatite samples with limonitic coating the intensity of the Fe2+ absorption feature diminishes (e.g., spectrum DLC3). The svite spectrum (Fig. 4A, SV1) has the main carbonate absorption feature at 2.333 m. In the visible to near infrared (VNIR) wavelength region, in addition to the broad ferrous-iron absorption feature this spectrum exhibits characteristic Rare Earth Elements (REE) absorptions at 0.58 m, 0.74 m and 0.80 m which are

generally attributed to electronic transitions in Nd3+ contained in rare earth element-bearing minerals (Rowan et al., 1986). The cause of a similar absorption feature in the same spectrum at 0.62 m has not been determined. The HyMap image spectra of the carbonatites (Fig. 5A) despite the lower spectral resolution are similar to the reectance spectra acquired in laboratory, except for the atmospheric obscuration regions around 1.40 m and 1.90 m. The HyMap image spectrum DLC4 (Fig. 5A) from an erosive area within the core zone of the carbonatite is characteristic for the ferroan dolomite carbonatite, exhibiting the broad Fe2+ absorption feature in the 1.01.3 m wavelength region. The spectrum DLC5 (Fig. 5A) is typical for the inner core zone of the carbonatite representing ferroan dolomite carbonatite with limonitic coating. The spectral reectance features of the limonite (e.g., Hunt, 1977; Hunt and Ashley, 1979) are combined with the spectral reectance features of the ferroan dolomite carbonatite in the VNIR wavelength region. It is to be noted that the limonitic coating (which is ubiquitous at the Sarfartoq carbonatite core zone) does not inuence the SWIR region of the spectrum and the main dolomite absorption at 2.32 m. The HyMap image spectrum SV2 (Fig. 5A) which represents a svite zone, is similar to the spectrum DLC5, except for the main carbonate absorption that is located at the HyMap band recording at 2.34 m, indicative of the presence of calcite. Rare Earth Elements (REE) absorptions have not been noted in the HyMap image spectra of the carbonatites in the study area. This could be related to the level of concentration of the REE, their narrow absorption features and spectral resolution of the hyperspectral data. Absorption features due to very high concentrations of Light Rare Earth Elements in

1212

E. Bedini / Remote Sensing of Environment 113 (2009) 12081219

Fig. 4. (A) Representative reectance spectra of rock samples of dolomite carbonatite (DLC1; DLC2; DLC3) and svite (SV1). (B) Representative reectance spectra of rock samples for fenites (FEN1; FEN2), carbonatite and fenite (SIL-CRB1) and hematized gneiss (HEM + SER). The vertical short arrows show important absorption features in the adjacent spectra.

carbonatites have been mapped in AVIRIS data of the Mountain Pass, California carbonatites (Rowan & Mars, 2003). 4.2. Fenites An almost universal characteristic of carbonatite complexes is the presence of a distinctive metasomatic aureole in which the wall rocks have been converted to aegirine-rich and alkali amphibole-rich rocks and in some cases to K-feldspar rich rocks (Winter, 2001). These metasomatic rocks are commonly called fenites and the process fenitization (Winter, 2001). Rowan et al. (1995) showed reectance spectra of fenites from the Iron Hill carbonatite complex, Colorado, displaying an MgOH doublet absorption feature due to amphibole and weathered fenites displaying only weak AlOH due to illite and/or MgOH absorption features caused by biotite or chlorite. The fenite rocks spectrally analyzed in this study, in the SWIR wavelength region display MgOH absorption features (Fig. 4B; FEN1 and FEN2) that are centered near 2.32 m and 2.38 m. These absorption features are attributed to the alkali amphibole phase present in the fenites according to petrographic descriptions (Nielsen, 1982). In some cases (Fig. 4B; FEN2) the MgOH doublet absorption feature is associated with a shallow AlOH absorption feature near 2.20 m due to sericite. The latter is attributed to the surface weathering of the feldspar. Both reectance spectra (FEN1 and FEN2) display a rapid fall-off in intensity from 2.0 m to the blue due to broad Fe2+ and Fe3+ absorptions (e.g., Hunt & Salisbury, 1970; Hunt et al., 1973). The HyMap image spectra from zones where eld mapping has recorded fenite in the boundary between the carbonatite core zone and the marginal alteration zone show both an MgOH doublet absorption feature and an AlOH absorption feature in the SWIR region (Fig. 5B;

FEN3 and FEN4). The spectra exhibit a rapid fall-off of the spectrum in the VNIR region due to combinations of broad Fe2+ and Fe3+ absorption bands (e.g., Hunt & Salisbury, 1970). However, fenites often show ferric-iron absorption features in the VNIR wavelength region. 4.3. Outer core zone (fenites with carbonatite dykes) The outer core zone of the Sarfartoq carbonatite consists of fenitized basement rocks intersected from abundant carbonatite dykes that make up between 10 and 50% of the rock (Secher & Larsen, 1980). The spectral reectance properties of these rocks represent a mixture of the spectral reectance of the carbonatite and fenite. In the SWIR wavelength region the spectral reectance of the rocks of this 2 absorption features of the carbonate zone is characterized by CO 3 minerals combined with MgOH absorption features of the fenite or fenitized country rocks (Fig. 4B, SIL-CRB1; Fig. 5A, SIL-CRB2). As the abundance of the carbonate phase decreases the MgOH absorption features dominate the reectance spectrum. The increase in the fenite content is also associated with the reduction of the characteristic broad ferrous-iron absorption of the ferroan dolomite in the VNIR region of the reectance spectrum. 4.4. Country rocks The country rocks consist mainly of granitic gneisses. Spectral reflectance properties of these rocks vary from featureless to displaying AlOH features due to muscovite and/or Fe,MgOH features due to biotite and chlorite (Fig. 5B). Field observations also showed that the country rocks are extensively covered from lichen, usually of black

E. Bedini / Remote Sensing of Environment 113 (2009) 12081219

1213

Fig. 5. Representative HyMap image spectra of the main lithologies of the Sarfartoq carbonatite complex. (A) Ferroan dolomite carbonatite displaying very intense ferrous-iron absorption feature (DLC4), ferroan dolomite with limonitic coating (DLC5), svite (SV2), carbonatite and fenite (SIL-CRB2). (B) Fenites (FEN3; FEN4), hematite and sericite (HEM+SER) from the marginal alteration zone, gneiss (GNEISS), carbonate and green vegetation (VEGETATION+CARBONATE). The vertical short arrows show important absorption features in the adjacent spectra.

colour. This is also evident in the image pixel spectra especially by the presence of the characteristic broad cellulose absorption feature of lichens around 2.10 m (e.g., Ager & Milton, 1987). 4.5. Marginal alteration zone The marginal alteration zone is distinguished by the strong hematization/limonitization of the gneisses. In the VNIR region, the reectance spectra of these rocks display intense ferric-iron absorption features due to hematite or limonite (e.g., Fig. 4B, HEM + SER). This spectrum (Fig. 4B, HEM + SER) in the SWIR region exhibits AlOH absorption features at 2.20 m, 2.35 m and 2.45 m due to sericite. Depending on the composition of the gneiss, in the SWIR region spectra from these rocks often display both AlOH absorption features and Fe,MgOH absorption features due to chlorite and biotite. The HyMap image spectra from the marginal alteration zone (Fig. 5B, HEM + SER) display ferric-iron spectral features in the VNIR region and AlOH absorption feature at 2.202.22 m. These are sometimes associated with an absorption feature at 2.25 m which could be attributed to the presence of an Fe,MgOH phase probably chlorite or biotite, based on eld observations. 4.6. Vegetation cover In the plateau (see Fig. 2), tundra vegetation and the presence of lichens constitute the major challenges for the remote sensing applied to lithologic mapping. On the other hand, the high levels of the slope, where is also the best exposure of the carbonatite complex (see Fig. 2), are practically devoid of vegetation cover. In general even in the plateau the lichen cover is minimal on the carbonatites probably due to the chemical composition. Sparse vegetation cover associates debris of carbonatite material and weathered surfaces in the lowest parts of

the carbonatite core zone outcrop. The HyMap reectance spectra from these areas show often carbonate absorption features in the SWIR wavelength region and green vegetation spectral reectance features superimposed on the VNIR wavelength region of the spectrum (Fig. 5B) as indicated by the characteristic chlorophyll absorption at about 0.68 m (Knipling, 1970). 5. Methods of HyMap data analysis 5.1. Unsupervised classication of HyMap data to map lithology of the Sarfartoq carbonatite complex Due to the low sun angle in high geographic latitudes, areas in shadow and presence of water bodies, the HyMap data in several parts of the study area were dominated from noise. To avoid artefacts from dark pixels in the data analysis stage and to focus at the rock outcrop, pixels with low reectance (less than 5%) and pixels with high NDVI-score were masked out. The HyMap data analysis was based in the 0.452.48 m spectral region. The HyMap bands which measure close or in the atmospheric water absorption zones around 1.40 m and 1.90 m were excluded from further analysis. As the study area is covered from two HyMap ights, due to different image statistics the subscenes were analyzed separately. A spectral reduction and data compression was performed using the Minimum Noise Fraction (MNF) transformation (Green et al., 1988). The MNF is a form of principal components analysis but instead of ordering the data in terms of variance the data are ordered based on the signal to noise ratio (Green et al., 1988). The rst 20 MNF bands were used as input to an unsupervised articial neural network, the Kohonen self-organizing maps (SOM). The SOM belongs to a class of unsupervised articial neural networks which are trained by competitive learning (Kohonen, 1982; Kohonen, 2001). A brief description of the SOM given below is based

1214

E. Bedini / Remote Sensing of Environment 113 (2009) 12081219

on a summary by Canty (2006). The network consists of an input layer and an output layer of neurons. The input signal (in our case a pixel spectrum in MNF space) is represented by the vector x = (x1, x2 xn)T where T denotes transpose and n is the number of bands considered. The vectors {x(i) |i = 1 p} (where p is the number of all pixel spectra used in the training process) are used as training data for the neural network. Each neuron is associated with a synaptic weight. The components of the vector wk = (wk1, wk2 wkn)T are thus the synaptic weights of the kth neuron. The initial synaptic weights for each neuron are usually randomly chosen from the data. When a training vector x is presented to the network, the neuron whose weight vector is most similar to x (usually the Euclidian Distance is used as a measure of similarity) is called the winner. Then the weight vector of the winner neuron (k) is shifted in the direction of the training vector: wk i + 1 = wk i + xi wk i 1

5.2. Methods to map the svite zones of the carbonatite An important objective of geological mapping and economic geology studies of carbonatite complexes is the mapping of dolomite carbonatite and svite (calcite carbonatite). The Sarfartoq carbonatite complex is predominantly composed of dolomite carbonatite with minor svite (Secher and Larsen, 1980). As discussed in Section 4.1 positions of carbonate absorption bands in the SWIR region of reectance spectra can be used to distinguish between dolomite and calcite (e.g., Gaffey, 1985). Reectance spectra obtained from hyperspectral imaging sensors allow mapping of the surface spatial distribution of these minerals over large areas, providing information which could not be easily obtained from eld observations. The hierarchical tree approach has been shown successful for the differentiation between dolomite and calcite (e.g., Kruse, 1990; Kruse et al., 1990). The hierarchical tree used here consists of a twostep procedure: (1) mapping of the carbonatite class (obtained from the SOM classication); (2) in the region dened from step one, in the HyMap reectance data for every pixel spectrum the following rule is applied: if the strongest carbonate absorption feature is located at the HyMap band recording at 2.32 m then it is dolomite carbonatite; else if the strongest carbonate absorption feature is located at the HyMap band recording at 2.34 m then it is svite. The second step is similar to one of the nodes of the hierarchical tree used by (Kruse et al., 1990) for the identication of calcite and dolomite.

where wk(i + 1) is the adjusted weight vector of the winner neuron. The parameter is called the learning rate of the network. In order for this method to function it is necessary to allow the learning rate to decrease gradually during the training process. A convenient function for this is:  i = max min max i = p : 2

During the learning phase not only the winner neuron but also the neurons in its neighbourhood are moved in the direction of the training vectors: wk i + 1 = wk i + ik; kxi wk i; k = 1 N M 3

where (k, k) denes a neighbourhood function for the winner neuron on the network of neurons usually a Gauss function of the form: h i 2 2 k; k = exp d k; k = 2 : 4

The d2(k,k) is the square of the distance between neurons k and k. The extent of the neighbourhood is allowed to shrink steadily:  = max min max i = p : 5

Typically the neighbourhood is initially the entire network and toward the end of the training is very localized. Despite the design of the SOM as an algorithm for the visualization of similarities in a multi-dimensional space, successful applications of the SOM as a classication tool of remote sensing imagery have been shown (Gonalves et al., 2008; Ji, 2000; Penn & Livo 2001; Villman et al., 2003; among others). In this study the SOM was based on two dimensional eight-by-eight hexagonally oriented units, implemented using the kohonen package (Wehrens & Buydens, 2007) in the R environment for statistical computing (R Development Core Team, 2007). A key point in the application of the SOM was its implementation in the MNF space (e.g., Penn & Livo, 2001). All the image pixels that were not masked in the pre-processing step were used as training data (input layer) for the SOM. The start value for the learning rate was max = 0.05 and the value to stop min = 0.01. The number of iterations (i.e. the number of times the data set is presented to the SOM) was 100. In this implementation (kohonen package in the R) the size of the neighbourhood decreases linearly during training such that after onethird of the iterations only the winning unit is being adapted (Wehrens & Buydens, 2007). The resulting clusters were labelled (using ENVI) based on their mean spectrum (in the original hyperspectral space) interpreted in accordance with the spectroscopic knowledge gained from the reectance spectroscopy stage. Spectrally similar classes were merged together.

Fig. 6. HyMap image spectra (mean normalized) used as reference spectra in the spectral mixture analysis. Image spectra MAZ-1 and MAZ-2 represent the marginal alteration zone. Gaps near 1.40 and 1.90 m due to deletion of channels affected by atmospheric absorption. Note that the MNF transformed spectra were used in the spectral mixture analysis.

E. Bedini / Remote Sensing of Environment 113 (2009) 12081219

1215

5.3. Spectral mixture analysis (SMA) In order to obtain fraction abundances of the classes of interest and compare the SOM thematic mapping results with the spectral unmixing approach, the spectral mixture analysis (Settle & Drake, 1993) was applied to analyze a part of the study area (see Fig. 3) noted for the lithologic transition carbonatite inner core zone carbonatite outer core zone fenite marginal alteration zone. The spectral mixture analysis (SMA) was applied as follows. Pixels with low reectance (less than 5%) and pixels with high NDVI-score were masked out. The HyMap reectance data in the 0.452.48 m region were mean normalized (i.e. each pixel spectrum was divided by its mean). This form of normalization eliminates the effects of different albedos in the spectral unmixing results (e.g., Berman et al., 2004). Minimum Noise Fraction (MNF) transformation (Green et al., 1988) was applied to the mean normalized data. Image derived reference spectra (Fig. 6) representing carbonatite, fenite and the marginal alteration zone (two reference spectra were selected for the latter) were used as input to the spectral mixture analysis. The SMA was applied (using ENVI linear spectral unmixing) in the 20 rst MNF bands space. The application of the SMA in a subset of MNF bands is of advantage, as noise isolated in the excluded MNF bands does not inuence the spectral unmixing (e.g., Nielsen, 2001). The sum of the fractions was not constrained. The fractions

produced from the spectral mixture analysis were ltered using a 3 3 median lter. 6. Results and discussion 6.1. Mapping lithology of the Sarfartoq carbonatite complex The result of the SOM-based unsupervised classication of the HyMap data was a classied image in seven classes which represent the following lithologic units and rockvegetation mixed spectral classes: carbonatite, carbonatite outer core zone (fenite with carbonatite dykes), fenite, marginal alteration zone (hematized gneiss), gneiss, CLASS-1, and CLASS-2 (Fig. 7). The SOM clusters that mapped pixels dominated by the spectral response of lichen or by the combined spectral response of lichen and green vegetation (that had not been masked in the pre-processing step) are not shown in Fig. 7. The carbonatite class mapped areas with a clear carbonate (overwhelmingly dolomite) spectral response in the SWIR wavelengths. In the VNIR region the image spectra from these areas display ferric-iron absorption features of the limonitic coating combined with the broad ferroan-iron absorption at 1.01.3 m, similar to the spectra shown in the section of reectance spectroscopy. A class that mapped erosive areas without limonitic coating in the carbonatite inner core zone (mainly

Fig. 7. Classication results of the HyMap data for the main lithologic units of the Sarfartoq carbonatite complex. The boundaries of the geological map of the carbonatite complex are shown for reference with Fig. 1. White circles show the locations of the eld stations used to assess the accuracy of the mapping results. The rectangle (white dashed line) indicates the spatial extent of Fig. 9.

1216

E. Bedini / Remote Sensing of Environment 113 (2009) 12081219

along the gullies) exhibiting typical ferroan dolomite absorption features (similar to spectrum DLC4 in Fig. 5A) is merged in this class as well. The results show that carbonatite rocks dominate the central part of the carbonatite core zone outcrop but extend somewhat more (up to the top of the slope) than previously mapped (the inner core boundary in Fig.1B). The results for the fenite class (distinguished by the MgOH doublet absorption feature often combined with AlOH absorption at around 2.20 m) are in agreement with the map of the carbonatite complex (Fig. 1B) in the south-western part of the image where there is also the best outcrop of the fenite zone. The rocks of the outer core zone have the spectral characteristics of a mixture between carbonatite and the MgOH silicate phase in fenites. Classication results show that these rocks outcrop in the western and eastern part of the carbonatite core zone as well as in several areas in the north within the carbonatite outer core zone of (Fig. 1B). Field observations, and by interactively checking pixel spectra from the HyMap image cube showed a very high accuracy of the classication for the delineation of the spatial extent of the outcropping carbonatite inner core and outer core zones. Such a detailed mapping of the carbonatite rocks outcrop is a direct contribution of the remote sensing approach to the study of the carbonatite. The marginal alteration zone is spectrally distinguished by the ferric-iron spectral features in the VNIR region and AlOH absorptions at 2.202.20 m in the SWIR region, due to sericite. The marginal alteration zone is well mapped at the outcropping part along the valley slope (Fig. 2) and numerous altered spots have been mapped in the plateau. Gneiss often mixed with lichen was mapped in other parts of the study area. In the lowest parts of the carbonatite core zone vegetation is more abundant as well as loose weathered material. These areas are mapped in CLASS-1. A miscellaneous spectral class (CLASS-2) is recognized at the top of the outcropping carbonatite. In general it has strong ferric oxide/hydroxide spectral reectance features and exhibits often MgOH absorption features. Few

pixels within this area were mapped in the carbonatite outer core zone class. Field observations showed that this area represents a weathered crust over the carbonatite outcrop combined with not in-situ material and it is not relevant for the study of the carbonatite lithology. 6.2. Mapping the svite zones The map of svite and dolomite carbonatite is shown in Fig. 8. The thematic mapping results indicate limited presence of svite in the area mapped as the carbonatite inner core and outer core zones. This is in agreement with previous studies (Secher & Larsen, 1980; Secher, 1986) that report dolomite carbonatite predominance at the outcropping part of the Sarfartoq carbonatite core zone. However, there is a tendency for more abundant presence of svite in the lowest parts of the carbonatite inner core zone. Especially the carbonatite outcrops at locations A and B (Fig. 8) are rich in svite. 6.3. Results of spectral mixture analysis The results of spectral mixture analysis (SMA) for carbonatite, fenite, marginal alteration zone (hematized gneiss) as well as a colour composite of these are shown in Fig. 9. The fraction image of the marginal alteration zone (Fig. 9) is the sum of the fraction abundances produced from the SMA for the reference spectra MAZ-1 and MAZ-2 (Fig. 6). The SMA produced very good results for the carbonatite class. The fenite zone and the marginal alteration zone are well mapped within the exposed part of the carbonatite complex along the valley. The most important result of the SMA is the mapping of the outer core zone of the carbonatite consisting of fenitized country rock and carbonatite dykes. This lithology is distinguished by the image analysis as a mixture of fenite and carbonatite (yellow in the colour

Fig. 8. Thematic map showing the distribution of dolomite carbonatite and svite at the outcropping inner and outer core zones of the Sarfartoq carbonatite complex, based on the analysis of the hyperspectral data.

E. Bedini / Remote Sensing of Environment 113 (2009) 12081219

1217

Fig. 9. Fraction abundances produced from the SMA for (A) carbonatite, (B) fenite, (C) marginal alteration zone and (D) a colour composite (carbonatite = red, fenite = green, marginal alteration zone = blue). Note the outer core zone of the carbonatite complex in yellow as a mixture of fenite and carbonatite. The boundaries of the geological map are superimposed on the colour composite. The image covers an area of 2.56 2.56 km. (For interpretation of the references to colour in this gure legend, the reader is referred to the web version of this article.)

composite). It is to be noted the high similarity of the results of Spectral Mixture Analysis and the SOM classication. 6.4. Accuracy assessment Field work at the Sarfartoq carbonatite complex to evaluate the accuracy of the lithological mapping results was carried out in August, 2008. Eighty-one reference sites representative of carbonatite (dolomite carbonatite and svite), outer core zone (fenite with abundant carbonatite dykes), fenite, marginal alteration zone (hematized gneiss) and gneiss were eld checked (Fig. 7). The location of each of these reference sites was determined using GPS measurements. Rock samples were collected from almost one-third of the reference sites, especially from the sites mapped as svite from the remote sensing image analysis. The relationship between ground reference information and the thematic classication map (Fig. 7) of the hyperspectral data using the SOM algorithm is summarized in the confusion matrix (Table 2). The confusion matrix (sometimes referred to as the error matrix) provides the basis on which to both describe classication accuracy and characterize the errors (Congalton & Green, 1999). The overall

accuracy is 87.6%. The confusion matrix shows that carbonatite (undifferentiated between dolomite carbonatite and svite) is the best mapped class with some confusion with the outer core zone class. The fenite class has error of commission with the outer core zone and gneiss classes (i.e. outer core zone and non-fenitized gneiss rocks wrongly have been mapped as fenite). The marginal alteration zone has no error of commission but the error of omission in which three sites that belong to the marginal alteration zones have been included in the gneiss class indicates that this zone is more widespread than mapped from the hyperspectral data. Twelve sites mapped from the analysis of the HyMap data as svite (calcite carbonatite) were eld checked and samples were collected for further analysis. One of the concerns before the eld work stage was that the spectral response of calcite could be due to monohydrocalcite (a weathering mineral in this Arctic environment); in this way the mapping of the calcite mineral would have no petrological signicance. However, in all the eld-checked sites, which were mapped as svite from the hyperspectral data analysis, was found a white variety of carbonatite (Fig. 10). Rock samples collected from these twelve sites showed strong effervescence with dilute hydrochloric acid (HCl), a characteristic of the calcite mineral as

1218

E. Bedini / Remote Sensing of Environment 113 (2009) 12081219

Table 2 Confusion matrix for the eld observations versus image classication results. Reference data Carbonatite Outer core zone (fenite + carbonatite dykes) MAP Carbonatite Outer core zone (fenite + carbonatite dykes) Fenite Marginal alteration zone Gneiss Column total 28 1 2 18 Fenite Marginal Gneiss Row alteration total zone 0 0 0 0 0 0 30 19

7. Conclusions The results of this study show that analysis of high spatial and spectral resolution HyMap imaging spectrometer data can provide detailed maps of surface lithology/mineralogy for outcropping carbonatites and their associated rocks. Through an image processing strategy based on a careful analysis of rock spectral reectance measurements, the inner carbonatite core zone, the outer carbonatite core zone, the fenite zone and the marginal alteration zone of the Sarfartoq carbonatite complex in southern West Greenland were mapped. The mapping of the spatial distribution of the svite rocks within a predominantly dolomite carbonatite core zone has igneous petrology signicance, indicating svite-rich zones in the lower levels of the inner core zone of the Sarfartoq carbonatite complex. The results obtained from the remote sensing approach especially for the carbonatite inner and outer core zones have the potential to improve the geological mapping of the carbonatite complex, by providing spatially contiguous mineralogic and lithologic information, which for inaccessible areas cannot be effectively obtained in any other way. Acknowledgments This study was funded by the Geological Survey of Denmark and Greenland (GEUS). I thank Leif Thorning, Karsten Secher, Tapani Tukiainen, Thorkild M. Rasmussen (all with GEUS) and Birger Ulf Hansen (Institute of Geography and Geology, University of Copenhagen) for helpful discussions and their continuous support for this study. Hudson Resources Inc. is thanked for the hospitality at their Garnet Lake eld camp in southern West Greenland and for the logistic support of the eld work. I thank Karsten Secher (GEUS) for the geological eld guide to the Sarfartoq carbonatite complex. I thank Llewellyn Pilbeam (GEUS) for his assistance during the eld work at the Sarfartoq carbonatite complex. The HyMap data were geometrically corrected by Tapani Tukiainen (GEUS). The spectral reectance measurements of rock samples were carried out at the spectral laboratory of the Geological Survey of Finland (GTK). I thank Leah

0 0

2 0

8 0

0 10

1 0

11 10

0 29

0 22

1 9

3 13

7 8

11 81

Overall accuracy = 71/81 = 87.6% Khat coefcient of agreement = 83.6% Users' accuracy Carbonatite = 28/30 = 93.3% Outer core zone = 18/19 = 94.7% Fenite = 8/11 = 72.7% Marginal alteration zone = 10/ 10 = 100% Gneiss = 7/11 = 63.6%

Producer's accuracy Carbonatite = 28/29 = 96.5% Outer core zone = 18/22 = 81.8% Fenite = 8/9 = 88.9 Marginal alteration zone = 10/ 13 = 76.9% Gneiss = 7/8 = 87.5%

opposed to dolomite. Spectral reectance measurements on these samples showed that the carbonate main absorption feature varies between 2.33 and 2.34 m. X-ray diffraction (XRD) analyses carried out on ve of these samples indicated that the calcite is the predominant mineral in four of them, although dolomite is also present in all of these samples. The XRD analysis of one sample indicated an almost equal abundance of calcite and dolomite. As the majority of the eld-checked sites are located at the area A (Fig. 8) more eld checking is needed for the zones mapped as svite in other parts of the carbonatite.

Fig. 10. Field photograph of svite rocks mapped from the hyperspectral data at the area A (see Fig. 8).

E. Bedini / Remote Sensing of Environment 113 (2009) 12081219

1219

Roach (with Planetary Geosciences Group of Brown University, USA) and three anonymous reviewers for their helpful comments and suggestions on the manuscript.

References
Ager, C. M., & Milton, N. M. (1987). Spectral reectance of lichens and their effects on the reectance of rock substrates. Geophysics, 52, 898906. Allaart, J. H. (1982). Geological map of Greenland 1:500 000. Sheet 2, Frederikshb Isblink Sndre Strmfjord. Copenhagen, Denmark: Geological Survey of Greenland. Berman, M., Kiiveri, H., Lagerstrom, R., Ernst, A., Dunne, R., & Huntington, J. (2004). ICE: A statistical approach to identifying endmembers. IEEE Transactions on Geoscience and Remote Sensing, 42, 20852095. Canty, M. J. (2006). Image Analysis, Classication and Change Detection in Remote Sensing, with Algorithms for ENVI/IDL. 348 p. Taylor & Francis, CRC Press. Clark, R. N. (1999). Spectroscopy of rocks and minerals, and principles of spectroscopy. In A. Rencz (Ed.), Manual of Remote Sensing, Vol. 3 (pp. 358). New York: Wiley and Sons Inc. Cocks, T., Jenssen, R., Stewart, W. I., & Shields, T. (1998). The HyMap airborne hyperspectral sensor: The system, calibration and performance. In M. Schaepman, D. Schlpfer, & K. I. Itten (Eds.), Proceedings of the 1st EARSeL Workshop on Imaging Spectroscopy, 68 October 1998, Zurich (pp. 3743). Paris: EARSeL. Congalton, R. G., & Green, K. (1999). Assessing the accuracy of remotely sensed data: Principles and practices.Boca Raton FL: CRC/Lewis Press 137 pp. Gaffey, S. J. (1985). Reectance spectroscopy in the visible and near-infrared (0.352.55 m): Applications in carbonate petrology. Geology, 13, 270273. Gaffey, S. J. (1986). Spectral reectance of-carbonate minerals in the visible and near infrared (0.352.55 microns): Calcite, aragonite, and dolomite. American Mineralogist, 71, 151162. Gaffey, S. J. (1987). Spectral reectance of carbonate minerals in the visible and nearinfrared (0.352.55 m): Anhydrous carbonate minerals. Journal of Geophysical Research, 92, 14291440. Goetz, A. F. H., Vane, G., Solomon, G. J., & Rock, B. N. (1985). Imaging spectrometry for Earth remote sensing. Science, 228, 11471153. Gonalves, M. L., Netto, M. L. A., Costa, J. A. F., & ZulloJunior, J. (2008). An unsupervised method of classifying remotely sensed images using Kohonen self-organizing maps and agglomerative hierarchical clustering methods. International Journal of Remote Sensing, 29, 31713207. Green, A. A., Berman, M., Switzer, P., & Craig, M. D. (1988). A transformation for ordering multispectral data in terms of image quality with implications for noise removal. IEEE Transactions on Geoscience and Remote Sensing, 26, 6574. Harris, J. R., Rogge, D., Hitchcock, R., Iljewliw, O., & Wright, D. (2005). Mapping lithology in Canada's Arctic: Application of hyperspectral data using the minimum noise fraction transformation and matched ltering. Canadian Journal of Earth Sciences, 42, 21732193. Hunt, G. R. (1977). Spectral signatures of particulate minerals in the visible and near infrared. Geophysics, 42, 501513. Hunt, G. R., & Ashley, R. P. (1979). Spectra of altered rocks in the visible and near infrared. Economic Geology, 74, 16131629. Hunt, G. R., & Salisbury, J. W. (1970). Visible and near-infrared spectra of minerals and rocks, I. Silicate minerals. Modern Geology, 1, 283300. Hunt, G. R., & Salisbury, J. W. (1971). Visible and near-infrared spectra of minerals and rocks, II. Carbonates. Modern Geology, 2, 2330. Hunt, G. R., Salisbury, J. W., & Lenhoff, C. J. (1973). Visible and near-infrared spectra of minerals and rocks: VI. Additional silicates. Modern Geology, 4, 85106. Ji, C. Y. (2000). Land-use classication of remotely sensed data using Kohonen self organizing feature map neural networks. Photogrammetric Engineering & Remote Sensing, 66, 14511460. Knipling, E. B. (1970). Physical and physiological basis for the reectance of visible and near-infrared radiation from vegetation. Remote Sensing of Environment, 1, 155159. Kohonen, T. (1982). Self-organized formation of topologically correct feature maps. Biological Cybernetics, 43, 5969. Kohonen, T. (2001). Self-organizing maps, Third Edition Springer series in information sciences, vol. 30 (pp. ) Berlin Heidelberg: Springer-Verlag 501 pp.

Kruse, F. A. (1990). Articial intelligence for analysis of imaging spectrometer data. Proceedings, IPRS Commission VII, Working Group 2: Analysis of High Spectral Resolution Imaging Data, Victoria, B.C., Canada, 1721 September, 1990 (pp. 5968). Kruse, F. A., Boardman, J. W., & Huntington, J. F. (2003). Evaluation and validation of EO-1 Hyperion for mineral mapping. IEEE Transactions on Geoscience & Remote Sensing, 41, 13881400. Kruse, F. A., Seznec, O., & Krotkov, P. M. (1990). An expert system for geologic mapping with imaging spectrometers. Proceedings, Applications of Articial Intelligence VIIISociety of Photo-Optical Instrumentation Engineers (SPIE), vol. 1293 (pp. 904917). Larsen, L. M., & Rex, D. C. (1992). A review of 2500 Ma span of alkalineultramac, potassic and carbonatite magmatism in West Greenland. Lithos, 28, 367402. Larsen, L. M., Rex, D. C., & Secher, K. (1983). The age of carbonatites, kimberlites and lamprophyres from southern West Greenland: Recurrent alkaline magmatism during 2500 million years. Lithos, 16, 215221. Nielsen, A. A. (2001). Spectral mixture analysis: Linear and semi-parametric full and iterated partial unmixing in multi- and hyperspectral data. International Journal of Computer Vision, 42, 1737. Nielsen, L. W. (1982). Geokimiske og malmgeologiske undersgelser i og omkring karbonatitkomplekset Sarfartq, Vestgrnland, med srlig henblik p CuPbZnminerliseringer. Dissertation. University of Copenhagen, 167 pp. Penn, B. S., & Livo, E. K. (2001). Using AVIRIS data to map geologic signatures of copper at porphyry copper deposit, Hillsboro, New Mexico. Summaries JPL Airborne Earth Science Workshop. JPL Publication 02-1, Pasadena, California. R Development Core Team. (2007). R: A language and environment for statistical computing.Vienna, Austria: R Foundation for Statistical Computing3-900051-07-0 URL http://www.R-project.org/. Richter, R., & Schlpfer, D. (2002). Geo-atmospheric processing of airborne imaging spectrometry data. Part 2: Atmospheric/topographic correction. International Journal of Remote Sensing, 23, 26312649. Rivard, B., & Arvidson, R. E. (1992). Utility of imaging spectrometry for lithologic mapping in Greenland. Photogrammetric Engineering & Remote Sensing, 58, 945949. Rowan, L. C., Bowers, T. L., Crowley, J. K., Anton-Pacheco, C., Gumiel, P., & Kingston, M. J. (1995). Analysis of airborne visible-infrared imaging spectrometer (AVIRIS) data of the Iron Hill, Colorado, carbonate-alkalic igneous rock complex. Economic Geology, 90, 19661982. Rowan, L. C., Kingston, M. J., & Crowley, J. K. (1986). Spectral reectance of carbonatites and related alkalic igneous rocks: Selected samples from four North American localities. Economic Geology, 81, 857871. Rowan, L. C., & Mars, J. C. (2003). Lithologic mapping in the mountain pass, California area using Advanced Spaceborne Thermal Emission and Reection Radiometer (ASTER) data. Remote Sensing of Environment, 84, 350366. Schlpfer, D., & Richter, R. (2002). Geo-atmospheric processing of airborne imaging spectrometry data. Part 1: Parametric orthorectication. International Journal of Remote Sensing, 23, 26092630. Secher, K. (1976). Airborne radiometric survey between 66 and 69N, southern and central West Greenland. Rapport Grnlands Geologiske Undersgelse, 80, 6567. Secher, K. (1986). Exploration of the Sarfartoq carbonatite complex southern West Greenland. Rapport Grnlands Geologiske Undersgelse, 128, 89101. Secher, K., Heaman, L. M., Nielsen, T. F. D., Jensen, S. M., Schjth, F., & Creaser, R. A. (in press). Timing of kimberlite, carbonatite, and ultramac lamprophyre emplacement in the alkaline province located 6467N in southern West Greenland. Lithos (Special 9IKC- volume). Secher, K., & Larsen, L. M. (1980). Geology and mineralogy of Sarfartoq carbonatite complex, southern West Greenland. Lithos, 13, 199212. Settle, J. J., & Drake, N. A. (1993). Linear mixing and the estimation of ground cover proportions. International Journal of Remote Sensing, 14, 11591177. Tukiainen, T., & Thorning, L. (2005). Detection of kimberlitic rocks in West Greenland using airborne hyperspectral data: The HyperGreen 2002 project. Geological Survey of Denmark and Greenland Bulletin, 7, 6972. Villman, Th., Merney, E., & Hammer, B. (2003). Neural maps in remote sensing image analysis. Neural Networks, 16, 389403. Wehrens, R., & Buydens, L. M. C. (2007, October). Self and Super-organizing maps in R: The Kohonen package.Journal of Statistical Software, 21(5) 19 pp. Winter, J. D. (2001). An introduction to igneous and metamorphic petrology.: PrenticeHall Inc. 695 pp.