The Hyperion dataset has to be corrected for abnormal pixels, striping and smiling prior to the atmospheric correction. Pre-processing of hyperspectral images is required not only for removing sensor errors during acquisition but also for display, band selection (to reduce the data dimensionality) and to reduce computational complexity.

The Hyperion sensor covers the visible and near infra-red (VNIR, or 400 nm to 1000 nm) and shortwave infra-red (SWIR, 900 nm to 2500 nm) ranges. The instrument has a single telescope but two spectrometers―one for the VNIR and the other for the SWIR data. The VNIR spectrometer has 70 bands and the SWIR 172 bands providing 242 potential bands which are normally about 10 nm apart in each spectrometer and have spectral response functions with approximately 11 nm full-width at half maximum (FWHM). If every band were operating, the spectral range covered would be from 356 nm to 2577 nm with a common area of overlap between the spectrometers between 852 nm and 1058 nm. In practice a number of the bands were not illuminated and others correspond to areas of low sensitivity of the spectrometer material so that the 242 a subset of 198 bands spectrally sub-setted from the Hyperion data set.

VNIR Range: Bands 8 - 57 (436 - 926 nm)

SWIR Range: Bands 77 - 224 (892 - 2406 nm)

These results in 198 useable channels or bands with 196 unique wavebands Hence the redefined spectral ranges of the Hyperion image set is as follows.

VNIR Range: Bands 8 - 57 (436 - 926nm)

SWIR Range: Bands 79 - 224 (933 - 2406 nm)

The basic 196 unique data channels that was further used for analysis.

A vertical stripe is said to occur where the statistics indicate that the image information is likely to be valid, i.e. pixel is not bad but have significantly modified gain and offset. In a pushbroom sensor, a poorly calibrated detector in either VNIR or SWIR array leaves high frequency errors (“vertical stripes”) on the image bands.

In the present study the bad columns were identified visually to avoid enforcing severe change in the spectra. A total of 36 bad columns were identified in dataset. SWIR bands were found devoid of visible stripes.

Bad columns were identified visually and bad column removal filter was used to target the bad columns in each band. The bad columns were replaced by the 3 × 3 neighbourhood mean, without taking into account the bad column value.

ENVI’s FLAASH (Fast line of sight Atmosphere Analysis of Spectral Hyper cubes) was used for the retrieval of reflectance from radiance. FLAASH starts from a standard equation for spectral radiance at a sensor pixel, L, that applies to the solar wavelength range (thermal emission is neglected) and flat, Lambertian materials or their equivalents. The equation is as follows:

L = ( A ρ / 1 − ρ e S ) + ( B ρ e / 1 − ρ e S ) + L a (1)

After the water retrieval is performed, Equation (1) is solved for the pixel surface reflectance in all of the sensor channels. The solution method involves computing a spatially averaged radiance image is estimated using the approximate equation

L e = ( A + B ) ρ e / ( 1 − ρ e S ) + L a (2)

Once the atmospheric correction model is run, the reflectance is retrieved from the radiance.

The validation of the reflectance spectra has been done with the analytical spectral device Fieldspec to generate the laboratory spectra of the rock samples that have collected from the field. The spectral collection of the rock samples were conducted almost at the same time of the sensor image capturing time. This was done to regenerate the same illumination conditions that persisted at the time of the image capture IN order to make measurements of surface reflectance of geological samples.

To compare the ground-measured spectra collected from ASD with that of atmospherically corrected image spectra, the ASD spectra were re-sampled to the Hyperion FWHM and bandwidth. A spectral library of re-sampled ASD spectra was created. The spectra from the ASD library were used as standard to compare image reflectance spectra extracted using FLAASH. The ground control points collected during the ground truth investigation were used to locate the exact pixel in the image which corresponds to the respective sample location. The image spectra for all the major sampling locations were generated similarly and saved in spectral library.

The minimum noise fraction (MNF) transformation is used to determine the inherent dimensionality of image data, Minimum Noise Fraction images were created for the 196 unique bands of Hyperion data. Around 10 MNF components contained the maximum information. The remaining bands contain maximum noise. Hence the first 10 bands were used for further processing and for selecting pure pixels in the image [4] .

In the present study, number of iterations with different threshold limit was carried out interactively to isolate the position of most pure pixels in the image [5] . A threshold of two was fixed for the identification of pure pixels in the image which can be explained as, all the pixels having 2 DN values (maximum limit) greater than the extreme pixel is considered as being pure. Different sets of iterations 50, 100, 500 and 1000 were carried out on this dataset with keeping the threshold at 2. It was observed that more the number of iterations more the number of extreme pixels found with more variability in the dataset.

In the study area, all the relevant topographic features are clustered in the N-Dimensional visualizer and demarcated as separate region of interests and saved. These ROI’s were used for the further mapping using various techniques. The endmembers clustered in the study area are as follows. 1) Al rich Laterite/Bauxite 2) Khondalite 3) Weathered soil with agriculture 4) Weathered soil 5) Vegetation

Linear Spectral Unmixing is a means of determining the relative abundances of materials depicted in multispectral imagery based on the materials spectral characteristics. The reflectance at each pixel of the image is assumed to be a linear combination of the reflectance of each material (or endmember) present within the pixel.

A unit-sum constraint in the linear Mixing algorithm was applied in this study where the score varies between zero and one. This was implemented to allow user-defined weighting of a sum-to-unity constraint on the abundance fractions. It also permits proper unmixing of MNF-transform data, with zero-mean bands. The weight factor, a default value of one is used for the extra constraint equation. This weighted unit-sum constraint is then added to the system of simultaneous equations in the unmixing inversion process [6] .

Mixture tuned matched filtering (MTMF) performs a partial unmixing―finding the abundances of user defined endmembers.

Mixture Tuned Matched Filtering technique combines the best part of the linear spectral unmixing model and the statistical matched filter model while avoiding the drawbacks of each model. From matched filtering it inherits the advantage of its ability to map a single known target without knowing the other background endmember signatures, unlike traditional spectral mixing models. From spectral mixture modeling it inherits the leverage arising from the mixed pixel model, the constraint on feasibility including the unit sum and positive requirements unlike the Matched Filter which does not employ these fundamental facts [7] . Hence this method can out perform either method in case of subtle, sub-pixel occurrences.

MNF transformation was applied on the Hyperion data. Several trials of Minimum Noise Fraction transformation have been run to obtain the optimum results. The first 10 eigenvectors (Figure 2) were chosen for both the datasets and the rest of them were more noise prone hence discarded from further spectral analysis. The FCC using the first 3 bands of MNF Transformation [8] .

The pixel purity index image was derived from the MNF image (Figure 4). The number of iterations was chosen as 10000 with 250 iterations per block, given a threshold of 2. The threshold were changed and tested on various MNF images of the study area. It was found that the threshold value of 2 has given the best

results and hence been chosen for further selection of the region of interests (ROIs).

The endmember needs to be selected for the abundance mapping using various mapping methods such as Linear Spectral Unmixing (LSU), Mixed Tuned Matched Filtering (MTMF) and Matched Filtering etc. The various endmembers are chosen with respect to their relevance in the study area and the geologically relevant features are chosen as endmembers (Figure 3).

Linear spectral unmixing has been performed using the final endmembers of the study area. It is obtained by giving endmember region of interest as the reference spectra and their abundance in the image are given in the resultant final Linear spectral unmixing image.

The linear spectral unmixing result of the study area clearly demarcates the Bauxite ore. This is basically due to the mineral composition of the mineral. It has the characteristic absorption depth at 2200 nm that enables these minerals to be distinguished in the image, the abundance has also been successfully mapped using the linear spectral unmixing model.

Image endmembers are derived from the hyperion satellite imagery using the interactive analysis technique of ENVI. All the final image endmembers mostly

describing the image were considered for creating a set of score and infeasibility image.

The interpretation of mixture tuned matched filtering mapping method involves the integration of close analysis of both the score and infeasibility images of the individual endmembers. The Al rich laterite/Bauxite and Vegetation endmember’s score as well as infeasibility image are given in (Figure 4). The score images of Bauxite exhibits their abundance in the area with high pixel values and they can be considered only valid if the same area has low pixel values in the infeasibility image (Figure 5).

Maximum Likelihood Classifier was applied to the MNF transformed components, using the ROI’s of selected endmembers. In order to compare the results with those from the Maximum likelihood, the classification thresholds are adjusted to 2. The ground truth makes it easy to make absolute performance comparisons. Classified Image was examined for its validity by using the error matrix of ground truth ROI’s which were collected during the Field Visit of the Study area and the sample spectra that were collected.