Application of Feature Based Principal Component Analysis (FPCA) technique on Landsat8 OLI multispectral data to map Kimberlite pipes

Objectives: Tomap the kimberlite pipes emplaced in parts of AnantpurDistrict, India using Landsat-8 OLI multispectral data. Kimberlite are considered as the primary host of natural diamond. Kimberlite pipes have very limited exposure and are altered, therefore the indirect surface indicators associated with kimberlite such as ferric iron bearing minerals (hematite, goethite), hydroxyl (clay) and carbonate (calcrete) minerals, were mapped to trace kimberlite pipe. Methods: Feature based Principal Component Analysis (FPCA) was applied over the OLI bands 2, 4, 5 and 6, and 2, 5, 6 and 7 to generate ferric iron (F image) and hydroxyl/carbonate image (H/C images). The color compositewas generated by assigning RGB colours to F, H/C and F+H/C images. Findings: When matched with the pre-explored kimberlite pipe locations, it was observed that the kimberlitic pipes display different colours in the above colour composite. Hence, the Isodata clustering was carried out to segregate the classes, which resulted in 12 unique classes. Of these, the kimberlite pipes fall in 4 classes. However, due to the moderate resolution of OLI, false positive areas were also noted. Further the target area was found to be reduced by incorporating the structural control (lineament) over the emplacement of Kimberlite pipes. Novelty: The present work highlights the usefulness of the moderate resolution multispectral image in mapping the Kimberlite pipes in semiarid region, in absence of a hyperspectral sensor.


Introduction
Kimberlites are rare igneous rocks. They are considered to be geochemists probe to the earth's mantle due to presence of mantle xenocrysts carried by the kimberlite magma originating in the deep mantle. Presence of natural diamond in some of the kimberlite pipes makes them economically important (1)(2)(3) . These unique ultrabasic and ultrapotassic volcanic rocks, essentially having mantle xenolith showing porphyritic

Geological setting of the study area
The Archaean Dharwar Craton of the India is bounded by Bastar Craton towards north east, Proterozoic Eastern Ghats Mobile Belt (EGMB) towards the east, Deccan Volcanic Province in the north west and the Southern Granulite Terrain (SGT) in the south (35) (Figure 1). The Dharwar Craton is made up of granite-green stone belts as well as of tonalite-trondhjemitegranodiorite (TTG) basement gneiss (36) . Numerous Paleo-Mesoproterozoic intracratonic sedimentary basins (e.g. Cuddapah basin) unconformably overlie the granite-greenstone terrain in the eastern end margins. Dharwar craton is divisible into two distinct groups called as Eastern Dharwar Craton (EDC) and Western Dharwar Craton (WDC) (37) based on the history of crustal evolution (38,39) . A prominent shear zone located along the eastern margin of the Chitradurga greenstone belt, the Chitradurga Boundary Fault (CBF), is widely considered to be the boundary between these blocks (37)(38)(39)(40) . A North-south trending Closepet Granite pluton of 2510 Ma age intrude the craton (41) . Widespread mafic dyke swarms are also observed in https://www.indjst.org/ the area (42) . Kimberlite pipes intruding the EDC represent the youngest rocks (~1100 Ma) (42,43) . The kimberlite pipes observed in EDC are grouped under Wajrakarur Kimberlite Field (WKF), Narayanapet Kimberlite Field (NKF) and Raichur Kimberlite Field (RKF) (44) .The Wajrakarur Kimberlite Field (WKF) is situated in the eastern part of the Eastern Dharwar Craton (EDC) of the South Indian Shield. WKF includes four kimbelite clusters (i) Wajrakarur-Latavaram Cluster, (ii) Kalyandurg Cluster, (iii) Timmasamudram Cluster and (iv) Chigicherla Cluster (8) . Each cluster has several pipes along with more than 40 pipes discovered so far from WKF (45,46) , out of which a few are reported to being diamondiferous.
Kimberlite pipe emplacement in the region is highly controlled by regional lineaments (16) and structural dome (8,47) . The lineaments and the dykes show two major trends, NW-SE trending Dharwarian lineaments and NE-SW trending regional faults/lineaments. The probability of kimberlite emplacement increases at the intersection of these two set of lineaments (8) . Morphologically, kimberlite pipes occur in sloping grounds or topographically depressed areas, some of them showing circular outline (16) . The general geology and the spatial distribution of Kimberlite pipes in WKF with the tectonic settings are shown in Figure 1. (a)The location mapof the study area; (b) regional geological map of the Wajrakarur Kimberlite field (after Nayak and Kudari, 1999); and (c) Landsat8 OLI standard False Color Composite covering the study area.

Materials and Methods
The datasets used in the study includes SOI toposheet of the study area of 1:50,000 scale; published geological map, Landsat 8 OLI satellite image and ancillary data about explored kimberlite pipes from published literatures, and limited ground validation. https://www.indjst.org/ The topographical maps and the published Geological maps were georeferenced to geographic coordinate with WGS84 datum and finally stored in the geodatabase.
Cloud free Level 1 Terrain corrected (L1T) Landsat8 OLI multispectral data of the study area LC81440502013103LGN01_WK, acquired on 2013-04-13 was downloaded from Earth Explorer website. Landsat8 has total 11 bands, of which six bands (band 2 to 7) are in visible to shortwave infrared region of EMR with a 30m spatial resolution. Band 8 is panchromatic band with 15 m resolution. Band 11 and 12 are thermal bands with 100 m resolution resampled to 60m spatial resolution. Band 1 and band 9 are Coastal aerosol and Cirrus bands respectively, which has 30m spatial resolution (48) (Table 1). The Level 1 Terrain corrected (L1T) Landsat 8 data provides the terrain corrected data in calibrated Digital Numbers (DNs). Data in each band (Band 2 to Band 7) were first converted to at-sensor-radiance, which was given as input to FLAASH (Fast Line of sight Atmospheric Analysis of Hypercube) atmospheric correction modelling tool in ENVI. FLAASH incorporates the MODTRAN4 radiation transfer code (48)(49)(50) and converts the radiance values to the Top-of-Atmosphere (TOA) reflectance values. The reflectance bands were further used for analysis to target iron oxides, clay/carbonate minerals and also for lineament mapping in the area.
Feature based Principal Component Analysis (FPCA) technique is a modified version of the PCA method developed by Crosta and Moore (28) . It allows application of PCA over selected combination of bands that provide crucial information about a particular target object (29) . The target was delineated based on the eigenvector loading of the most spectrally significant band over the Principal components (29,30) . Field studies as well as literature about kimberlite pipes exposed in the Eastern Dharwar craton and around the globe had revealed that the Kimberlite pipes were marked by presence of red soil mixed with clay or calcrete. Therefore, it was decided to target iron oxide, clay and carbonate minerals as a proxy for locating kimberlite pipes in the study area. Band selection for FPCA was decided based on the spectral signature of the target minerals. From Figure 2, it can be seen that the iron oxide has strong absorption in blue region, whereas it has high reflection in the red region. Similarly, the hydroxyl ion bearing mineral (clay) exhibit diagnostic absorption spectrum around 2.2µm to 2.3µm and depending upon the nature of the clay minerals strong reflection around 1.5µm region. The carbonate minerals exhibit diagnostic absorption around 2.3µm. As the spectral resolution of the Landsat OLI image is moderate, the clay and carbonate absorption would have been recorded in the SWIR2 band (Landsat OLI band 7) itself and could not be identified separately and hence mapped together.
In order to target hydroxyl/carbonate minerals, PCA transformations were applied over an image data that contain Landsat 8 OLI bands 2, 5, 6 and 7. In a similar way, PCA transformation were applied over an image that contained bands 2, 4, 5 and 6 to target iron oxide. The Eigen vector loading of each band on the Principal component were noted and the principal component that has the highest loading was considered to map the target. The clay/carbonate minerals or iron oxides could be distinguished by noting the eigenvector loading of corresponding bands having diagnostic spectral characteristics in the PC component by careful examination of correlation matrix. The hydroxyl/carbonate minerals could be better distinguished by a PC image with high or moderate eigenvector loadings for band 7, either in positive or negative sign; and a high or moderate eigenvector loadings of the opposite sign for band 6. If band 7 loadings are positive, then the clay/hydroxyl/carbonate bearing minerals will be highlighted as dark pixels in the PC image, these pixels can be made brighter by negation of the PC (28)(29)(30)(31)(32)(33)(34) . Similarly, for iron oxide, the magnitude of eigenvector loading of band 2 and 4 in the PCs has to be studied and it should be moderate or strong with opposite signs. https://www.indjst.org/ The colour composite generated by combining the H/C, F components were useful in highlighting the regions rich in these minerals (34) . To create colour composite, a third band was also needed, and hence H/C and F images were added together to produce F+H/C image which highlights the region having both iron oxide and hydroxyl/carbonate minerals. Colour composite was prepared by assigning red colour to H/C image, green colour to F+H/C image and blue colour was assigned to F image (34) . Prior to generation of the colour composite image, the H/C, F and H/C +F images were rescaled to 256 grey levels to normalize the values in each image. The resulting colour composite was compared with the locations of kimberlite pipe previously explored by GSI and other agencies. ISODATA (51) unsupervised classification technique was applied over the colour composite to derive the spectral classes representing kimberlite, hereafter referred to as Kimberlite Classes. In order to narrow down the kimberlite class, it was decided to incorporate structural control in emplacement of kimberlite. In the same context, the lineament buffer zone was generated to mark the zone of influence and incorporated while finalizing the kimberlite probable zones. The results were cross verified with kimberlite pipe positions as well as field observations

FPCA Analysis to target Kimberlite pipe locations
FPCA facilitates targeting a particular object by observing the eigenvector loading of the most spectrally significant band over the Principal components (33) . Iron oxides and Clay/Carbonate can be distinguished by noting the eigenvector loading of corresponding bands having diagnostic spectral characteristics in the PC component by careful examination of correlation matrix. Iron oxide was targeted by applying FPCA transformations over OLI bands2, 4, 5 and 6. The eigenvector loading of each band on the Principal component are given in Table 2. Iron oxide shows higher reflectance in red band (band 4) and less reflectance in blue band (band 2). Therefore, the magnitude of eigenvector loading of band 2 and 4 in the PCs were studied. The PC band that have eigenvector loading of these bands in moderate or strong with opposite signs was considered for targeting Iron oxide. Table 2 revealed that the eigenvector loading for the first principle component image (PC1) of bands 2, 4, 5 and 6 has positive values for all corresponding bands. The high positive values represent the albedo component and hence neglected (29) . Eigenvector loadings for PC2 and PC3 are high for Band 5 compared to other bands (0.575 and 0.613) and hence the bright pixels in PC2 and PC3 were considered as Vegetation. Eigenvector loading for the PC4 show that band 4 has high negative loading (-0.862), whereas band 2 has moderate loading with opposite sign (+0.369), therefore the iron oxide would be represented as dark pixels. To highlight these pixels as bright pixels, the PC4 was negated and the resulting image corresponds to "Ferric iron https://www.indjst.org/ image" (Figure 3 A). Clay/Carbonate minerals were targeted by applying FPCA transformations over OLI bands 2, 5, 6 and 7 and the eigenvector loading of each band on the Principal component is given in Table 3. Clay/Carbonate minerals exhibit strong reflection in SWIR band 1 (band 6) and strong absorption in SWIR band 2 (bands 7). Therefore, clay/carbonate minerals can be better distinguished by a PC image with high or moderate eigenvector loadings for band 7, either in positive or negative sign; and a high or moderate eigenvector loadings of the opposite sign for band 6. If band 7 loadings were positive, then the clay/hydroxyl/carbonate bearing minerals would be highlighted as dark pixels in the PC image. These pixels could be highlighted by negation of the PC. It was observed that the eigenvector loadings of band 6 and band 7 in PC4 were high compared to other bands, and in opposite signs (-0.6679 and +0.6903 respectively), and highlights hydroxyl/carbonate minerals as dark pixels. Since band 7 has +ve loading, PC4 image was negated to derive Hydroxyl/carbonate minerals as a Bright pixel and represents "Hydroxyl/Carbonate (H/C) image" (Figure 3B). H/C image highlights regions that have both clay rich region as well as carbonate rich region in the study area.  Figure 3C) highlights area that have higher Iron oxide, hydroxyl bearing or carbonate minerals. The colour composite image prepared by assigning RGB colours to H/C, F+H/C and F image bands respectively ( Figure 3D) exhibits various colours ranging from white, dark blue, cyan, red and orange colour depending on the concentration of dominant factor. For example, pixels having white colour indicate region of these pixels that were enriched in both iron oxide and clay/carbonate minerals. Pixels with dark red to orange colour represent areas with higher abundance of clay/carbonate minerals than iron oxide, dark blue pixel represents iron oxide dominant areas, cyan to bluish tone indicate varying amount of iron oxide and clay/carbonates.
While correlating the colour composite image with the geological map and known Kimberlite pipe locations, it was noted that the pixel corresponding to the black soil show orange colour, granites show dark blue colour, whereas the area covering Kimberlite pipes exhibit light blue to cyan colours. From this, it can be inferred that pixels corresponding to Kimberlite pipes have both iron oxide as well as clay/carbonate minerals. Spectrally, iron oxide and hydroxides are dominant than the clay/carbonate minerals (34) and hence the actual concentration of iron oxide might be less than what it appears. However, it was difficult to attribute one single colour to the kimberlite pipe as there was a varying shades of blue, corresponding to varying amount of the carbonate/ clay and iron oxides covering the kimberlite pipe. The observation matches with the field observation also, where the calcrete was seen mixed with red soil and the concentration of iron oxide and clay may vary from place to place.

ISODATA classification
In order to further narrow down the regions of possible occurrence of Kimberlite pipe, ISODATA (51) unsupervised classification of the colour composite was carried out. The Isodata classification resulted in generation of 16 classes. It was observed that the known Kimberlite pipes do not fall in one single class; instead they are found to occur in class8, 9, 11and 12. All the above four classes that coincide with the known Kimberlite location were retained to produce a Kimberlite classification map representing the regions with chance of occurrence of Kimberlite pipes (Figure 4). The classification map shows a large area as kimberlite, which was over representation and could be attributed to the false positive identification. The situation might have raised due to lower spatial and spectral resolution of Landsat image as well as occurrence of surface material which has similarity to kimberlite weathering products. In such case, the mapping needs to be refined by taking help of other factors such as structural control or morphology and hence, influence of lineament over the emplacement of kimberlite pipes was considered to narrow down the Kimberlite classification map.

Lineaments Mapping
Lineaments in the study area were generated from Landsat 8 OLI False Colour Composite. Lineaments were interpreted by visual interpretation technique and onscreen digitization was carried out in ArcGIS. Image enhancement techniques such as application of directional filters were applied to highlight the invisible lineaments. To eliminate the cultural lineaments such as roads, railway network etc. the data was compared with Survey of India toposheets and high resolution Google Earth images. Lineament map of Wajrakarur Kimberlite Field is shown in Figure 5. Many of the lineament are marked by intrusion of mafic dykes. Kimberlite pipes in the study area are generally found to be emplaced along regional scale weak zones, particularly intersection of the E-W Lineaments with N-S trending lineaments are more probable locations for the Kimberlite emplacement (17,18) . Kimberlite pipes are also observed to fall in between two set of parallel E-W trending lineaments and intersection of N-S lineament.

GIS analysis and field validation
The present study attempts to locate the probable locations of kimberlite pipes occurrence in the Wajrakarur Kimberlite Field (WKF) using multispectral remote sensing image. Field observation had revealed that fresh kimberlite outcrops were rare and pipes in the study area were of smaller dimension. Since the spatial extent of Landsat 8 OLI is 30 m, the pipes were usually covered https://www.indjst.org/ by a few pixels and hence it was difficult to discern the smaller dimension kimberlite pipes. Most of the kimberlite pipes were deeply weathered and often covered by few meter thick calcrete or transported soil. The calcrete which was developed over the kimberlite couldnot be spectrally resolved from the pedogenic calcrete developed over granitoid/granitic gneiss, thus resulted in false positives. The problem was compounded by the large spatial extent of the pixel which results in the mixed pixel. Based on the analysis of ASTER image of Narayanapeta kimberlite field, Guha et al. (12) concluded that calcrete cannot be detected from a mixed pixel of granite/granodiorite and calcrete with the later making up less than 10 weight percentage of the mixture. Hence results of the spectral based mapping technique to target kimberlite needs to be viewed with caution. Constraining the image analysis results by taking inputs from tectonic and geomorphic setting of kimberlite emplacement might reduce the false positives and enhance the results. (12) Kimberlite emplacement in the study area was found to be structurally controlled, as majority of the kimberlites pipes were found to be in the vicinity of deep-rooted regional scale lineament, particularly where the NE-SW/E-W trending lineament intersects with NW-SE/N-S trending lineaments (8) . Hence, the result of image analysis was refined by considering the lineament control over kimberlite emplacement. The zone of influence of the lineament on the kimberlite pipe emplacement was noted by overlaying the pre-explored pipe locations on the lineament map in the GIS environment. Most of the kimberlite pipes, except a few were found to occur at distance of few meters to 1000 m from the lineament. Accordingly, a buffer zone of about 1000 m around the lineament was considered to narrow down the Kimberlite emplacement zone.
Four Kimberlite Classes derived from Isodata Classification were merged together to a single class, which was later converted to vector format and stored in GIS geodatabase along with lineament buffer zone. Integration overlay analysis was carried out in ARCGIS to retain the Kimberlite class that occur within the buffer zone of lineaments and were considered as high probabilistic of Kimberlite emplacement zones ( Figure 6). The results of GIS analysis were validated by comparing them with the geological map, and field observation. From Figure 6, it was noted that incorporation of the structural control over the kimberlite mapping brought down the kimberlite class area drastically; thus reducing the false positive. Overlaying the known kimberlite pipes over the refined kimberlite classification map, showed that most of the existing kimberlite pipes fall within refined Kimberlite classification. During field validation, previously known kimberlite pipe locations [ Figure 7] and six new locations were visited. The preexplored kimberlite pipe locations matched with kimberlite class derived from Landsat OLI and GIS analysis. Of the other six places visited, four locations had development of calcrete cover. Further detailed study such as rock sample analysis, geophysical exploration would be needed to verify the results.  Figure 6) B. Calcrete mixed with red soil in a field observed at Chiggicherla(marked as 2 in Figure 6) C. Clay and calcrete observed in a pit where Kimberlite was dug out at Lattavaram ( marked as 3 in Figure 6 )

Conclusion
This study attest to feasibility of using Landsat 8 OLI data to locate the weathered kimberlite pipe exposed in a semi-arid region. Through FPCA technique, it was possible to locate the iron oxides and the clay or carbonate minerals, the pathfinder minerals for targeting the kimberlite pipe. Though lower spatial and spectral resolution of Landsat data resulted in few false positive signature, constraining with the known kimberlite locations and structural control (lineaments) was observed to enhance its efficiency in mapping of kimberlite. The refined map may be given as input in further GIS analysis which will aid in preparation of Kimberlite prediction map in any given area. https://www.indjst.org/