INTERNATIONAL JOURNAL OF GEOMATICS AND GEOSCIENCES Volume 2, No 2, 2011

Transcription

1 INTERNATIONAL JOURNAL OF GEOMATICS AND GEOSCIENCES Volume 2, No 2, 2011 Copyright 2010 All rights reserved Integrated Publishing services Research article ISSN Pre processing of Hyperspectral Data: A case study of Henry and Lothian Islands in Sunderban Region, Chakravortty Somdatta 1, Chakrabarti S 2 1 Assistant Professor, Department of Information Technology, Govt. College of Engineering & Ceramic Technology, 73, Abinash Banerjee Lane, Kolkata Associate Professor, Department of Geology & Environmental Systems Management, Presidency University, Kolkata, India csomdatta@rediffmail.com ABSTRACT Hyperspectral data finds wide applicability in species level mapping of forest cover in pure and mixed stands. The Sunderban Biosphere Reserve of West Bengal is an ideal locale where hyperspectral image data may be successfully utilized for accurate mapping of nearly 94 mangrove species that exist here. The present study is the first attempt to use hyperspectral data in the Sunderban eco geographic province to make species level discrimination of mangroves in a mixed stand. However, prior to data classification, several corrections are required to be made for pre processing and meaningful interpretation of data. Atmospheric correction is one such crucial correction and pre processing step which is done to minimize the effect of atmospheric agents that alters the actual radiance data that the sensor should represent. This is followed by geometric correction of the atmospherically corrected data. In this paper the MODTRAN based FLAASH algorithm and scene based QUAC algorithm have been applied on the Hyperion data and a comparative analysis carried out. The transformation techniques such as RST(Rotation Scale Translation)and Polynomial in combination with resampling methods like Nearest Neighbour (NN), Bilinear and Cubic Convolution(CC) have been used for analysis of geometric correction results and the best result selected for the next level of processing. This paper analyses the data processing steps for both atmospheric and geometric correction of hyperspectral data acquired by the Hyperion sensor over the dense mangrove forest cover of the Henry and Lothian Islands of the Sunderban Delta of West Bengal. Keywords: Hyperspectral data, Atmospheric Correction, FLAASH, QUAC, Geometric Correction, Nearest Neighbour resampling, Cubic Convolution, Sunderban 1. Introduction The basic objective of remote sensing has been to characterize the properties of the objects through detection, registration and analysis of the radiant flow emitted or reflected by them. However, the mechanism of acquisition of this radiation is not ideal due to the presence of an extremely dynamic medium the atmosphere between the sensor and the earth surface. This atmosphere interacts with the electromagnetic radiation leading to significant alterations in the incoming radiant flow from the target. Hence, it becomes a necessity to poise a larger emphasis on minimization of atmospheric influences on the acquired imagery through application of atmospheric correction algorithms (Kim et.al., 2002). Along with atmosphere generated errors most commercial sensors have systematic as well as unsystematic geometric errors. The systematic errors are removed in most cases. However, Submitted on November 2011 published on November

2 unless otherwise processed, the unsystematic error remains in the imagery, making it nonplanimetric(i.e. not in proper x,y planimetric position). This paper also focuses on geometric correction procedures used to register two images of like geometry and of the same geographic area to make the corresponding elements of the same ground area appear in the same place on the registered images (Jensen, 1996). The general objective of this study is to demonstrate the importance of atmospheric and geometric correction in hyperspectral image analysis with the dense mangrove forest of Sunderbans as the target area for the study (Vaiphasa, 2005) (Singh et. al., 2006). Needless to mention, it is also the first attempt to focus on the application of hyperspectral imagery and test their efficacy in the Sunderban delta especially for mangrove mapping at species level (Kanniah et. al., 2005) (Gao et. al.,1999). The FLAASH and QUAC algorithms have been applied for the atmospheric correction of hyperspectral images of the Henry and Lothian Islands of Sunderban Region,West Bengal. The transformation techniques such as RST(Rotation Scale Translation)and Polynomial in combination with resampling methods like Nearest Neighbour (NN), Bilinear and Cubic Convolution(CC) have been used for analysis of geometric correction results The dataset used for analysis has been acquired by the Hyperion sensor onboard the EO 1(Earth Observatory 1) satellite launched by NASA. Lothian Island Henry Island 2. Study Area Figure 1: Browse Overlay of Hyperion Image over Sunderban As a case study, the pristine mangrove habitats of Henry island (approximately 10 sq. km. in area, extending between 21 o N to 21 o N latitude and 86 o E to 88 o E 491

3 longitude) and Lothian island (approximately 38 sq. km. in area, extending between 21 o N to 21 o N latitude and 88 o E to 88 o E longitude) of the Sunderban Biosphere Reserve of West Bengal have been selected for study. The selection of the study area is based considering the fact that Sunderban harbours a rich and bio diverse mangrove community (Demuro and Chisholm, 2003) (Saxena et. al., 2004) with a wide array of ecologically rare, endangered and endemic mangrove species. 3. Materials and Method 3.1. Data Acquisition Cloud free Hyperion data of Henry and Lothian Islands of Sunderban region was acquired on 27 th May, A browse overlay of the two islands of Sunderban is shown in Figure Data Processing The Hyperion data consists of 242 bands covering the wavelengths from 356 nm to 2577 nm. The VNIR detector collects data in bands 1 to 70 and the SWIR detector collects data from bands 71 to Removing the absorption bands and bands having no information It has been found that some bands are set to zero during Level 1 processing. The zeroed bands are 1 7, 58 76, ( EO1 User Guide, 2003). The remaining 194 bands have been retained for processing. Amongst the 194 unique bands there a number of atmospheric water vapour bands that absorb most of the solar radiation. These have been identified by examining the radiance spectra. While analyzing the hyper spectral images it has been found that the strongest water vapour bands occur between 1346nm(band 120) 1497 nm(band135),1517nm(band 137) 1537nm(band 139),1568nm(band 142) 1578nm(band 143),1598nm(band 145),1669nm(band 152) 1689nm(band 154),1709nm(band 156) 2385nm(band 213). Since these bands contain little or no information about the surface, they have been ignored for further processing(eo 1 User Guide,2003). However some atmospheric correction programs such as ENVI FLAASH do require bands centred near 1380 nm in the strong water vapour wavelengths for masking clouds (especially high altitude clouds). Therefore bands (1376, 1386, and 1396 nm) have been retained in the image.ignoring the above strong water vapour bands along with the zeroed bands leaves subset of about 92 bands for atmospheric correction and further processing. The zeroed bands and strong water vapour bands have been set as bad and ignored in further processing. This reduces the data volume and speeds up processing Removal of Bad Columns and Vertical Stripes The level 1R product consists of the bands affected with vertical stripes which has been minimised by checking each band for the vertical stripe and replacing the DN value of the affected column by the average of the DN values of the adjacent columns. In the present work, the bad columns have been identified visually in order to avoid imposing severe changes in the spectra. These bad columns once identified have been reduced by checking each band for the bad column, and replacing the DN value of the affected column by the average of the DN values of the adjacent columns. 492

4 3.2.3 Atmospheric Correction The atmospheric correction is considered as a critical pre processing step to achieve full spectral information from every pixel especially in case of hyper spectral data.this paper investigates and tries to minimise the effect of atmospheric correction on the hyperspectral bands covering the study area. The time in which the data has been acquired, the monsoon sets in and there are clouds and lot of water vapour in the atmosphere. Also, there is spectral variation in the vegetation of the region with the trees in its full vigour. Atmospheric correction for the Hyperion image is hence considered as a critical step in the present study. For atmospheric correction, FLAASH and QUAC models (Advanced Hyperspectral Analysis) have been used to convert the radiance values in the image to its reflectance values (ENVI User Guide). Quick Atmospheric Correction (QUAC) This approach is based on the radiance values present in the image (i.e. scene), hence it is known as scene based empirical approach. It determines atmospheric compensation parameters directly from the information contained within the scene (observed pixel spectra), without ancillary information. QUAC is based on the empirical finding that the average reflectance of a collection of diverse material spectra, such as the endmember spectra in a scene, is essentially scene independent. Fast Line of Sight Atmospheric Analysis of Spectral Hypercubes (FLAASH) FLAASH is an atmospheric correction tool that corrects wavelengths in the visible through near infrared and shortwave infrared regions, up to 3 μm. Unlike many other atmospheric correction programs that interpolate radiation transfer properties from a pre calculated database of modelling results, FLAASH incorporates the MODTRAN4 radiation transfer code. Water vapor and aerosol retrieval are only possible when the image contains bands in appropriate wavelength positions. In addition, FLAASH can correct images collected in either vertical (nadir) or slant viewing(off nadir) geometries Geometric Correction Selection of Method The Image to Image Registration method for geometric correction has been selected after considering the characteristics of available reference data and distortion in image data. This is a translation and rotation alignment process by which two images of like geometry and of same geographic area are positioned coincident with respect to one another so that the corresponding elements of the same ground area appear in the same place on the registered images.a geo referenced Hyperion image of the study area acquired by USGS has been used as a reference image which will be used to match the unrectified image. The projection for the study area is UTM at Zone 45(North) and Datum: WGS

5 Figure 2: Ground Control Points selected for Geo rectification Setting the Ground Control Points (GCPs) The GCPs were identified on the reference as well as the original imagery for mathematical modeling to remove the geometric distortion present in the unrectified image. Table 1.0 shows the GCPs of the reference image(base image), GCPs of the image to be geocorrected(warp) and the corresponding RMS error. Figure 2 shows the GCPs plotted on the imagery of Henry Island and Lothian Island. Determination of Parameters Correction method and correction formula has been decided by judging the characteristics of geometric distortion and the data available for correction. Rotation Scale Translation (RST) and Polynomial equations have been used to convert the source coordinates to rectified coordinates. Accuracy Check The Root Mean Square (RMS) error is calculated for all GCPs to check which GCPs exibit the highest error and to determine the sum of RMS error. Table 1.0 shows the RMS error corresponding to the GCPs of the reference image (base image) and GCPs of the image to be geo corrected (warp). 494

6 Table 1: Image to Image GCP Table Base x Base y Warp x Warp y Predict x Predict y Error x Error y RMS Error Interpolation and Resampling This process involves the extraction of a brightness value from a position in the original (distorted) input image and its relocation in the appropriate coordinate location in the rectified output image. The process of resampling namely, Nearest Neighbour, Bilinear Interpolation and Cubic Convolution has been applied for brightness value interpolation. Nearest Neighbour(NN) Resampling In zero order or nearest neighbour interpolation, the brightness values closest to the x, y coordinate specified is assigned to the output x,y coordinate. Bilinear Interpolation (BIL) First order or bilinear interpolation assigns output pixel values by interpolating brightness values in two orthogonal directions in the input image. It basically fits a plane to the four pixel values nearest to the desired point in the input image and computes a new brightness value based on weighted distances to these points. Cubic Convolution (CC) Resampling This method assigns values to output pixels in the same manner as bilinear interpolation except that the weighted values of 16 input pixels surrounding the located of the desired pixel are used to determine the value of the output pixel. 4. Result and Discussion 4.1 Atmospheric correction results After the image bands have been resized to 92, the FLAASH and QUAC model is run on the images. The scene center location, sensor type, flight date, sensor altitude, average ground elevation of the scene, flight time has been used as input for processing of the radiance data. For more accurate correction, a tropical atmospheric model and a maritime aerosol model has 495

7 been chosen and the water vapour content information extracted from Hyperion water absorption bands. (a) b) (c) Figure 3: (a) Original Band(1030nm) (b) Band after FLAASH Correction (c) Band after QUAC Correction (a) (b) 496

8 (c) Figure 4: Spectral Profile of the mangrove forest area (a) Original data (b) after FLAASH Correction(c) after QUAC Correction In Figure 3 we can trace out visually, the difference in the features seen before and after running the atmospheric correction model. After running FLAASH and QUAC model, the haziness in the image is minimised to a certain level and the features are sharpened with increased brightness. This can be interpreted statistically also, by observing the spectral profile in Figure 4 (graphs showing the variation of wavelength vs reflectance value) of the feature before and after running FLAASH and QUAC. As our main area of focus are mangroves, we have taken the spectral profile of a forested segment. It has been observed that in the visible portion of the spectrum, the Chlorophyll in the plants absorbs the blue and red wavelengths more strongly than green, producing a characteristic small reflectance peak within the green wavelength range. This may be seen in Figure 4(b) after FLAASH correction. The reflectance then rises sharply across the boundary between red and near infrared wavelengths which is primarily due to interactions with the internal cellular structure of leaves (Kumar et. al.,2001). From the above profile, one can observe the enhancement in the vegetation feature class after running both the models (Microimages Inc.) (Prasad et. al., 1992). We can well observe that the dips present in the profile between the wavelengths 625 to 750nm (approximated) and also at range are reduced after the FLAASH atmospheric correction. Instead, we can see a rise in the value starting from the blue region and a steep slope in the atmospherically corrected profile. Also, we can observe in the profile of the atmospherically corrected image, the presence of a number of narrow contiguous peaks in the wavelength range of nm range. After application of QUAC model there has been an enhancement in the vegetation feature class as compared with the original Hyperion data. From the spectral profiles generated it is found that the correction achieved in the near infrared region seems to be better after FLAASH correction than that of QUAC. 4.2 Geometric Correction results The output of RST (Rotation Scaling Transformation) NN and Polynomial NN is shown in Figure.5.0 (d) and (f). The results of RST CC, RST BIL and Polynomial CC is shown in Figure 5 (b), (c) & (e). 497

9 Figure 5(a) Atmospherically Corrected Image Figure 5(b) RST CC geo rectified image Figure 5(c) RST BIL geo rectified image Figure 5(d) RST NN geo rectified image 498

10 Figure 5(e) Polynomial CC georectified image Figure 5(f) Polynomial NN georectified image The Nearest Neighbour resampling method is a more computationally efficient procedure and does not alter the pixel value brightness values during resampling. The drawback observed is a stair case or jagged effect in the NN corrected images relative to the original unrectified data. The other interpolation techniques like Bilinear Interpolation and Cubic Convolution use averages to compute the output intensity which often leads to removal of valuable spectral information from the images. The bilinear method acts as a spatial moving filter that subdues extremes in brightness value thoughout the output image. This method is quite computationally demanding. It is observed that the stair case effect caused by the NN approach is reduced and looks smooth. The cubic convolution corrected image has a dramatic smoothing effect as compared to the other two methods because of averaging more number of pixels than the bilinear method. 5. Conclusion After running FLAASH and QUAC model, the haziness in the image is minimised to a certain extent and the features are sharpened with increased brightness. It is evident from the spectral profiles after correction that the strong absorption bands near the VNIR and SWIR regions of the original spectra have been compensated and corrected to a large extent. It has been observed that there is slight difference in the spectral profile obtained after QUAC and FLAASH corrections. FLAASH correction shows better correction results than QUAC as it incorporates some knowledge of the atmospheric conditions of the study area at the time of acquisition which is not considered in case of QUAC. QUAC performs a good approximate atmospheric correction to FLAASH. In this study FLAASH has been found to give better correction results than QUAC. The execution time of QUAC is however faster than FLAASH 499

11 as it needs less data for processing and performs well for atmospheric corrections of areas whose details are not known. The atmospheric corrected reflectance image can now be used for geo referencing and further classification. After analyzing the results obtained from geometric correction, the geo rectified images of NN methods have been selected for further processing because of its capability to retain the original brightness values. This is important as very subtle changes in the brightness values make a difference when discriminating mangrove species from each other (Ramsey and Jensen, 1996) (Sulong et. al.,2002). The results of RST NN show a more jagged output as compared to Polynomial NN. Hence the Polynomial NN method has been chosen for the next level of processing. Acknowledgement The corresponding author expresses her sincerest thanks to DST for extending financial support in the form of a Major Reaserch Project on the above study problem. 6. References 1. Demuro, M., Chisholm, L (2003), assessment of Hyperion for characterizing mangrove communities. In: Proceedings of the International Conference the AVIRIS 2003 Workshop, pp EO 1 User Guide (2003),Version ENVI User Guide: Advanced Hyperspectral Analysis 4. Gao, J (1999), a comparative study on spatial and spectral resolutions of satellite data in mapping mangrove forests. International Journal of Remote Sensing, 20, pp Jensen (1996), Introductory Digital Image Processing : ARemote Sensing Perspective, Second Edition 6. Kanniah. K. D., Ng, S. W., Lau, A.M.S. and Rasib, A. W (2005), linear Mixture Modelling applied to Ikonos data for Mangrove Mapping. In: Proceedings of the 26th Asian Conf. on Rem. Sens., ACRS 2005, 7 11 Nov. 2005, Hanoi, Vietnam. 7. Kumar, L., Schmidt, K., Dury, S., Skidmore, A.K (2001), imaging spectrometry and vegetation science. In: van de Meer, F., de Jong, S.M. (Eds.), Imaging Spectrometry. Kluwer Academic Press, Dordrecht, pp Kim, Shin, Yoo, Lee (2002), Effect of atmospheric correction for the land cover classification using hyperspectral data 9. Microimages Inc., Introduction to Hyperspectral Imaging 10. Prasad, P. R. C., Reddy, C. S., Rajasekhar, G., Dutt, C.B.S (1992), mapping and Analyzing Vegetation Types of North Andaman Islands, India. GIS development > Geospatial Application Papers > Natural Resource Management > Overview. 3 (1), pp. 500

12 1 3. [available online]. /application/ nrm/overview/ over004_1.htm. 11. Ramsey, E.W., Jensen, J.R (1996), remote sensing of mangrove wetlands: relating canopy spectra to site specific data. Photogrammetric Engineering and Remote Sensing 62, Saxena, A., Rawat, J.K., Singh, S.K (2004), survey and Mapping of Mangrove Cover using Remote Sensing A Case Study of Sundarbans. Map Asia (Geospatial Application Papers > Natural Resource Management > Coastal Zone Management > Management & Monitoring [available online]. /application/nrm/coastal/mnm/index.htm 13. Singh, S.K., Kushwaha, S.P.S., Joshi, P.K (2006), mangrove Mapping and Monitoring. GIM International 20(3), international.com/issues/ articles/ id634 Mangrove_Mapping_and_ Monitoring. html 14. Sulong, I., Mohd Lokman, H., Mohd Tarmizi, K., Ismail, A (2002), mangrove mapping using Landsat imagery and aerial photographs: Kemaman District, Terengganu. Malaysia Environment, Development and Sustainability, 4, pp Vaiphasa (2005), Hyperspectral Data for Tropical Mangrove Species Discrimination. 501