Filtering High Resolution Hyperspectral Imagery and Analyzing it for Quantification of Water Quality Parameters and Aquatic Vegetation.

Material Information

Filtering High Resolution Hyperspectral Imagery and Analyzing it for Quantification of Water Quality Parameters and Aquatic Vegetation.
Pande-Chhetri, Roshan B
Place of Publication:
[Gainesville, Fla.]
University of Florida
Publication Date:
Physical Description:
1 online resource (173 p.)

Thesis/Dissertation Information

Doctorate ( Ph.D.)
Degree Grantor:
University of Florida
Degree Disciplines:
Forest Resources and Conservation
Committee Chair:
Abd-Elrahman, Amr H.
Committee Members:
Smith, Scot E
Dewitt, Bon A
Toor, Gurpal Singh
Jacoby, Charles A
Graduation Date:


Subjects / Keywords:
Image classification ( jstor )
Image filters ( jstor )
Images of transformations ( jstor )
Ponds ( jstor )
Reflectance ( jstor )
Remote sensing ( jstor )
Signals ( jstor )
Stripes ( jstor )
Supernova remnants ( jstor )
Vegetation ( jstor )
Forest Resources and Conservation -- Dissertations, Academic -- UF
aquatic -- filtering -- hyperspectral -- imagery -- remotesensing
St. Johns River, FL ( local )
bibliography ( marcgt )
theses ( marcgt )
government publication (state, provincial, terriorial, dependent) ( marcgt )
born-digital ( sobekcm )
Electronic Thesis or Dissertation
Forest Resources and Conservation thesis, Ph.D.


High resolution hyperspectral imagery (airborne or ground-based) is gaining momentum as a useful analytical tool in various fields including agriculture and aquatic systems. These images are often contaminated with stripes and noise resulting in lower signal-to-noise ratio, especially in aquatic regions where signal is naturally low. This research investigates effective methods for filtering high spatial resolution hyperspectral imagery and use of the imagery in water quality parameter estimation and aquatic vegetation classification. The striping pattern of the hyperspectral imagery is non-parametric and difficult to filter. In this research, a de-striping algorithm based on wavelet analysis and adaptive Fourier domain normalization was examined. The result of this algorithm was found superior to other available algorithms and yielded highest Peak Signal to Noise Ratio improvement. The algorithm was implemented on individual image bands and on selected bands of the Maximum Noise Fraction (MNF) transformed images. The results showed that image filtering in the MNF domain was efficient and produced best results. The study investigated methods of analyzing hyperspectral imagery to estimate water quality parameters and to map aquatic vegetation in case-2 waters. Ground-based hyperspectral imagery was analyzed to determine chlorophyll-a (Chl-a) concentrations in aquaculture ponds.  Two-band and three-band indices were implemented and the effect of using submerged reflectance targets was evaluated. Laboratory measured values were found to be in strong correlation with two-band and three-band spectral indices computed from the hyperspectral image. Coefficients of determination (R2) values were found to be 0.833 and 0.862 without submerged targets and stronger values of 0.975 and 0.982 were obtained using submerged targets. Airborne hyperspectral images were used to detect and classify aquatic vegetation in a black river estuarine system. Image normalization for water surface reflectance and water depths was conducted and non-parametric classifiers such as ANN, SVM and SAM were tested and compared. Quality assessment indicated better classification and detection when non-parametric classifiers were applied to normalized or depth invariant transform images. Best classification accuracy of 73% was achieved when ANN is applied on normalized image and best detection accuracy of around 92% was obtained when SVM or SAM was applied on depth invariant images. ( en )
General Note:
In the series University of Florida Digital Collections.
General Note:
Includes vita.
Includes bibliographical references.
Source of Description:
Description based on online resource; title from PDF title page.
Source of Description:
This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Thesis (Ph.D.)--University of Florida, 2012.
Adviser: Abd-Elrahman, Amr H.
Statement of Responsibility:
by Roshan B Pande-Chhetri.

Record Information

Source Institution:
Rights Management:
Copyright Pande-Chhetri, Roshan B. Permission granted to the University of Florida to digitize, archive and distribute this item for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.
LD1780 2012 ( lcc )


This item has the following downloads:

Full Text




2 2012 Roshan Pande Chhetri


3 To my p arents and my wife


4 ACKNOWLEDGEMENTS I would like to express my deep est gratitude to my advisor, Dr. Amr Abd Elrahman for his excellent guidance, mentorship and continued support he provided to me, all the way from when I was first considering applying to the doctoral program through to completion of this degree. Dr. Abd Elrahman only by his genuinely good nature down to earth humility and generosity I am privileged to have had the opportunity to work with him. I would also like to thank my committee members, Drs. Scot Smith, Bon Dewitt, Gurpal Toor and Charles Jacoby for their invaluable guidance and support I am grateful to have an opportunity to take interesting and inspiring classes from Dr. Dewitt and Dr. Smith that enhanced and widened my knowledge base in Geomati cs Special thank goes to Dr. Jacoby for his tremendous support to my research by providing hyperspectral aerial imagery, related information and invaluable suggestions I would like to thank all my colleagues who worked with me in the field or lab oratory including Justin Harris, Matt h ew Croxton, Mary Thornhill, Joe Latvis, Naveen Anne and especially Nicole Hewitt who also helped me in reviewing my manuscript And my t hanks to very supportive administrative team of Plant City Center, especially Carolyn Hi nson who is always willing to help us like a guardian in official or personal matters. I appreciate my parents for their guidance, support wisdom and faith in me that always inspired me to better myself not just for the accomplishments but as a person. I thank my brother Deepak and my sister Shova for their encour agement and regards for all my endeavors Finally and importantly I would like to thank my wife, Roshani, for her support, encouragement, patience and unwavering love that ha ve energized past


5 de cade of my life and provided enough boost to accomplish this degree I acknowledge the innumerable sacrifices she made in shouldering far more than her fair share of the parenting and household burdens while I pursued my graduate degree s


6 TABLE OF CONTENTS p age ACKNOWLEDGEMENTS ................................ ................................ ............................... 4 LIST OF TABLES ................................ ................................ ................................ ............ 9 LIST OF FIGURE S ................................ ................................ ................................ ........ 10 LIST OF ABBREVIATIONS ................................ ................................ ........................... 12 ABSTRACT ................................ ................................ ................................ ................... 14 CHAPTER 1 BACKGROUND : HYPERSPECTRAL IMAGERY ................................ ................... 16 1.1 Introduction ................................ ................................ ................................ ....... 16 1.2 Statement of Problems ................................ ................................ ..................... 18 1. 2.1 Filtering Hyperspectral Imagery ................................ ............................. 18 1.2.2 Analyzing Hyperspectral Imagery for Estimating Water Quality Parameters and Detectin g Aquatic Vegetation ................................ ...... 20 1.3 Goals and Objectives ................................ ................................ ........................ 23 1.4 Organization of the Dissertation ................................ ................................ ........ 24 2 DE STRIPING HYPERSPECTRAL IMAGERY USING WAVELET TRANSFORM AND ADAPTIVE FREQUENCY DOMAIN FILT ERING ................... 26 2.1 Introduction ................................ ................................ ................................ ....... 26 2.2 Background ................................ ................................ ................................ ....... 30 2.2.1 Fourier Transform ................................ ................................ .................. 31 2.2.2 Wavelet Transform ................................ ................................ ................ 33 2.3 Methodology ................................ ................................ ................................ ..... 36 2.3.1 Propos ed Adaptive De striping Method ................................ ................. 37 2.3.2 Combined De striping and De noising Method ................................ ...... 42 2.3.3 Other Used De striping Methods ................................ ........................... 43 Spatial filter: simple linear matching (SLM) ............................... 43 Wavelet based filter ................................ ................................ .. 44 2.3.4 Quality Assessment ................................ ................................ ............... 46 Visual assessment ................................ ................................ .... 46 Quantitative assessments ................................ ......................... 46 2.4 Imagery and Experiments ................................ ................................ ................. 48 2.4.1 Used Images ................................ ................................ ......................... 49 2.4.2 Selection of Filter Parameters and Algorithm Implementation ............... 50 2.5 Results and Analysis ................................ ................................ ......................... 53 2.5.1 Filtering of Noise simulated Image ................................ ........................ 53


7 Visual inspection ................................ ................................ ....... 54 Quantitative assessment using RMSE and PSNR values ......... 55 SNR of s ample profiles ................................ ............................. 55 2.5.2 Filtering Hyperspectral Images ................................ .............................. 56 2.6 Discussion ................................ ................................ ................................ ........ 61 2.7 Conclusion ................................ ................................ ................................ ........ 63 3 FILTERING HIGH RESOLUTION HYPERSPECTRAL IMAGERY IN MAXIMUM NOISE FRACTION DOMAIN USING WAVELET BASED DE STRIPING .............. 74 3.1 Introduction ................................ ................................ ................................ ....... 74 3.2 Methodology ................................ ................................ ................................ ..... 78 3.2.1 Maximu m Noise Fraction (MNF) ................................ ............................ 78 3.2.2 Filtering ................................ ................................ ................................ .. 80 3.2.3 Quality Assessment ................................ ................................ ............... 83 3.3 Experiments and Results ................................ ................................ .................. 86 3.3.1 Hyperspectral Imagery ................................ ................................ ........... 86 3.3.2 Image Filtering ................................ ................................ ....................... 87 3.3.3 Parameter Selection ................................ ................................ .............. 90 3.4 Assessment and Discussion ................................ ................................ ............. 92 3.4.1 Visual Assessment ................................ ................................ ................ 92 3.4.2 Quantitative Assessment ................................ ................................ ....... 94 Analyzing autocorrelation indices and ratios ............................. 94 Analyzing signal to noise ratios. ................................ ............... 95 3.4.3 Discussion ................................ ................................ ............................. 96 3.5 Conclusi on ................................ ................................ ................................ ........ 97 4 ANALYSIS OF HYPERSPECTRAL IMAGERY TO ESTIMATE WATER QUALITY PARAMETERS IN CASE 2 WATERS OF AQUACULTURAL PONDS 106 4.1 Introduction ................................ ................................ ................................ ..... 106 4.2 Background ................................ ................................ ................................ ..... 110 4.2.1 Bio optical Models ................................ ................................ ............... 110 4.2.2 Modeling Chlorophyll a Concentration in Case 2 Waters .................... 111 4.3 Materials and methods ................................ ................................ .................... 113 4.3.1 Site Description ................................ ................................ .................... 113 4.3.2 Water Sampling and Laboratory Analysis ................................ ............ 116 4.3.3 Hyperspectral Image Acquisition and Calibration Target Platform ....... 117 4.3.4 Data Processing ................................ ................................ .................. 119 4.4 Results ................................ ................................ ................................ ............ 123 4.5 Discussion ................................ ................................ ................................ ...... 125 4.6 Conclusi on ................................ ................................ ................................ ...... 128 5 ANALYSIS OF HYPERSPECTRAL IMAGERY TO DISCRIMINATE AQUATIC VEGETATION IN A RIVER ESTUARINE ECOSYSTEM ................................ ...... 133 5.1 Introduction ................................ ................................ ................................ ..... 133


8 5.2 Site and Background ................................ ................................ ....................... 135 5.2. 1 Site Description ................................ ................................ .................... 135 5.2.2 Data ................................ ................................ ................................ ..... 136 5.2.3 Preprocessing Methods ................................ ................................ ....... 137 5.2.4 Radiation Transfer Models ................................ ................................ ... 138 5.3 Methodology ................................ ................................ ................................ ... 139 5.3.1 Image Normalization ................................ ................................ ............ 139 5.3.2 Depth Invariant Transformed Image ................................ .................... 140 5.3.3 Classifiers ................................ ................................ ............................ 141 5.3.4 Indices for Detection and Categorization ................................ ............. 143 5.3.5 Assessment Methods ................................ ................................ .......... 144 5.4 Experiments and Results ................................ ................................ ................ 145 5.4.1 Depth Invariant Transform ................................ ................................ ... 145 5.4.2 Supervised Classification ................................ ................................ ..... 146 5.4.3 Indices ................................ ................................ ................................ 147 5.4.4 Quality Assessment ................................ ................................ ............. 147 5. 4.5 Discussion ................................ ................................ ........................... 149 5.5 Conclusion ................................ ................................ ................................ ...... 150 6 CONCLUSIONS ................................ ................................ ................................ ... 155 6.1 Image Filt ering ................................ ................................ ................................ 155 6.2 Hyperspectral Image Analysis for Aquatic Applications ................................ .. 156 LIST OF REFERENCES ................................ ................................ ............................. 160 BIOGRAPHICAL SKETCH ................................ ................................ .......................... 173


9 LIST OF TABLES Table p age 2 1 RMSE and PSNR values computed from various filtering methods. .................. 65 2 2 SNR estimation of high and low frequency noise for row 960 of the noise induced image. ................................ ................................ ................................ ... 65 2 3 SNR estimation for original and de striped data computed at rows 30 and 830 of the Pond image (band 30). ................................ ................................ ...... 65 2 4 SNR estimation in frequency domain for original and de striped data computed at columns 40 and 438 of the Vegetation image (band 32 ). ............... 66 2 5 SNR estimation for original and de striped images computed at row 365 of the Hyperion image (band 195). ................................ ................................ ........ 66 4 1 Total nitrogen, total phosphorus, and chlorophyll a in aquaculture ponds used in the study. ................................ ................................ ............................. 115 4 2 Chl a model calibration and validation results using two band and three band indices. ................................ ................................ ................................ ............. 130 4 3 Chl a model calibration and validation results using two band and three band indices (reflectance computed by subtracting dark current values obtained b efore pond image acquisition). ................................ ................................ ........ 130 4 4 Calibration and validation results for total P and total N models. ...................... 130 5 1 Assignment of field polygons for classification training and for assessment. .... 137 5 2 Attenuation coefficient ratios for 12 band pairs ................................ ................. 145 5 3 Accuracy assessment results using different classification methods ................ 148


10 LIST OF FIGURE S Figure p age 2 1 Schematic diagram illustrating proposed adaptive de striping and combined de striping and de noising methods. ................................ ................................ ... 38 2 2 Wavelet decomposition of an image. ................................ ................................ .. 39 2 3 Schematic diagram of pro posed wavelet Fourier adaptive filter (step 2) applied on image of water body with floating targets. ................................ ......... 42 2 4 Frequency domain signal and noise. ................................ ................................ .. 48 2 5 Test images. ................................ ................................ ................................ ....... 67 2 6 Filtering with the proposed WFAF filter. ................................ .............................. 68 2 7 PSNR versus k threshold values using different decomposition levels and wavelet types ................................ ................................ ................................ ...... 69 2 8 Noise induced tomato image filtered using various methods. ............................ 70 2 9 Filtering results of the pond image ................................ ................................ ...... 71 2 10 Filtering results of the vegetation image ................................ ............................. 72 2 11 Filtering results of the Hyperion image ................................ ............................... 73 3 1 Schematic diagram illustrating Wavelet based Torres and WFAF de striping methods. ................................ ................................ ................................ ............. 82 3 2 Schematic diagram of MNF transform combined with wavelet based de striping methods. ................................ ................................ ................................ 83 3 3 Image profile and fourier values. ................................ ................................ ....... 85 3 4 Test images. ................................ ................................ ................................ ....... 87 3 5 Zoom in portion of test images bands 20, 150 & 204. ................................ ........ 88 3 6 MNF transform percent eigen value graphs ................................ ....................... 91 3 7 First 10 MNF bands of vegetation scene images. ................................ ............... 98 3 8 Filtering results of vegetation scene image cube. ................................ ............... 99 3 9 Filtered Vegetation images (bands 20, 150, 204) ................................ ............. 100


11 3 10 Filtered pond images (bands 20, 150, 204) ................................ ...................... 101 3 11 Geostatistical indices (autocorrelation with horizontal one pixel lag and horizontal to vertical autocorrelation ratio) for original and de striped images. 103 3 12 Signal to noise ratios SNR low (for low noise peak) and SNR high (for high noise peak ) computed along homogenous profiles (see Figure 3 4) for original and filtered images. ................................ ................................ .............. 105 4 1 Tropical aquaculture facility an d ponds utilized in this study ............................ 116 4 2 Hyperspectral image acquisition. ................................ ................................ ...... 118 4 3 Linearly stretched RGB representation of the Hyperspectral image of pond K demonstrating used ROIs. ................................ ................................ ................ 119 4 4 Reflectance spectra of four representative ponds with varying Chl a content collected from pixels in two ROIs ................................ ................................ ...... 123 4 5 Relationship between Chl a and reflection ratios ................................ .............. 131 4 6 Relationship between Chl a and three band index ................................ ........... 131 4 7 Total P and N relationship with reflectance ratios and three band indices. ...... 132 4 8 Total P and N relationship with reflectance ratios and three band indices after removing pond D (Oct 13, 2009) observations. ................................ ................ 132 5 1 Center wavelength (nm) positions of the image bands ................................ ..... 136 5 2 Typical spectra of aquatic vegetation and bare bottom. ................................ ... 143 5 3 Depth invariant transform example. ................................ ................................ .. 151 5 4 Classification of Image 17 normalized for water surface reflectance ................ 152 5 5 Classification of depth invariant transform bands of Image 17 ......................... 153 5 6 Indices based classification of image 17 ................................ .......................... 154


12 L IST O F ABBREVIATIONS ANN Artificial neural network CASI Compact Airborne Spectrographic Imager CCD Charge coupled device C DOM Colored Dissolved Organic Matter Chl a Chlorophyll a DC Direct current (of signal) DFT Discrete Fourier Transform F FT Fast Fourier Transform FWT Fast Wavelet Transform HYDICE Hyperspectral Digital Image Collection Experiment HS Hyperspectral IOP Inhere nt Optical Properties LOO Leave one out (validation algorithm) MERIS Medium Resolution Imaging spectrometer MNF Maximum Noise Fraction MODIS MODerate Resolution Imaging Spectroradiometer N Nitrogen NASA National Aeronautics and Space Administration P Phosp horous PCA Principal Component Analysis PSNR Peak Signal to Noise Ratio R 2 Coefficient of determination RMSE Root Mean Square Error ROI Region of Interest


13 SAM Spectral angular mapper SLM Simple Linear Matching SNR Signal to Noise Ratio SVM Support vector machine TM Landsat Thematic Mapper TN Total Nitrogen TP Total Phosphorous WFAF Wavelet frequency adaptive filter WO Water only (spectra)


14 Abstract of Dissertation presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy FILTERING HIGH RESOL UTION HYPERSPECTRAL IMAGERY AND ANALYZIN G IT FOR QUANTIFICATION O F WATER QUALITY PARA METERS AND AQUATIC VEGETATION By Roshan Pande Chhetri December 2012 Chair: Amr Abd Elrahman Major: Forest Resources and Conservation High resolution h yperspectral imagery ( airborne or ground based ) is gaining momentum as a useful analytical tool in various fields including agriculture and aquatic systems. These images are often contaminated with stripes and noise resulting in lower signal to noise ratio, especially in aquatic regions where signal is naturally low This research investigates effective methods for filtering high spatial resolution hyperspectral imagery and u se of the imagery in water quality parameter estimation and aquatic vegetation classification The striping pattern of the hyperspectral imagery is non parametric and difficult to filter. In this research, a de striping algorithm based on wavelet analysis and adaptive Fourier domain normalization was examined The result of this algorithm was found superior to other available algorithms and yielded highest Peak Signal to Noise Ratio improvement. The algorithm was implemented on individual image bands and on selected bands of the Maximum Noise Fraction (MNF) transform ed images. The results showed that image filtering in the MNF domain was efficient and produced best results


15 Th e study investigated methods of analyzing hyperspectral imagery to estimate water q uality parameters and to map aquatic vegetation in case 2 waters. Ground based hyperspectral imagery was analyzed to determine chlorophyll a (Chl a) concentrations in aquaculture ponds. Two band and three band indices were implemented and the effect of us ing submerged reflectance targets was evaluated Laboratory m easured values were found to be in strong correlation with two band and three band spectral indices computed from the hyperspectral image. Coefficients of determination ( R 2 ) values were found to be 0.833 and 0.862 without submerged targets and stronger values of 0.975 and 0.982 were obtained using submerged targets. A irborne hyperspectral image s were used to detect and classify aquatic vegetation in a black river estuarine system. Image normalizat ion for water surface reflectance and water depths was conducted and non parametric classifiers such as ANN, SVM and SAM we re tested and compared Quality assessment indicated better classification and detection when non parametric classifiers were applied to normalized or depth invariant transform images. Best classification accuracy of 73% was achieved when ANN is applied on normalized image and best detection accuracy of around 92% was obtained when SVM or SAM wa s app lied on depth invariant images.


16 CHAPTER 1 BACKGROU N D: HYPERSPECTRAL IMAGERY 1.1 Introduction Hyperspectral imagery is getting momentum in research and applications in multidisciplinary fields. Hyperspectral imagery is imagery with high spectral resolution which contains tens or, more often, hundreds o f narrow and continuous wavelength bands. This spectral richness facilitates many types of analysis while increases noise extent. Due to advancement of computer processing and decreasing cost of commercial grade hyperspectral sensors, use of hyperspectral imageries with higher spatial resolution is rapidly increasing. Airborne hyperspectral systems have been widely used in various applications. Most recently, ground based hyperspectral systems are being developed as laboratory or mobile systems. Ground base d systems can be mounted in mobile systems for faster and efficient capture and offer greater scope due to higher resolution in both spectral and spatial dimensions (Abd Elrahman et al. 2011). High resolution hyperspectral imagery has been used in applica tions including detection and classification of plant species, evaluation of plant health and moisture, characterization of ecosystems and more recently, in the study of aquatic environment. Images captured by hyperspectral systems with a 2D pushbroom arra y, such as Hyperion, CASI and HYDICE, are susceptible to in track striping and smile effects in addition to atmospheric effects (Schowengerdt, 200 7 ). These effects contaminate the imagery and can affect further analyses. For regions of low brightness such as aquatic environment, contamination can lead to further lower signal to noise ratio and difficult ies in analysis. Proper preprocessing is, therefore, an important component in the analysis of hyperspectral imagery. Striping and smile effects can be reduc ed by careful one time


17 calibration if these effects are unchanging from image to image. But in many cases, striping effects change or valid calibration coefficients for the image are not available. Stripes are of non periodic nature which complicates the f iltering process. For higher spatial resolution imagery, reduction of striping effect is more difficult and requires careful filtering. One of the objectives of this is to develop an effective and efficient method of reducing stripes and noise from high sp atial resolution hyperspectral imagery. In the past decades, applications of hyperspectral imagery have increased in many fields including agricultural, natural resources and aquatic ecosystems. Hyperspectral imagery has been used in complex applications i ncluding assessing plant stress ( Champagne et al., 2003), identifying invasive species (Okamoto, 2007), agricultural crop classification and yield estimation (Ye et al., 2006 ) study of rocks and minerals and characterization of ecosystems (Schmidt and Ski dmore, 2003 ) Remote sensing analysis of aquatic environment has always been challenging due to the complex optical interaction in water. Multispectral satellite imagery was utilized in earlier studies of aquatic environment, supplemented sometimes with sp ectrometer observations. Hyperspectral imagery, due to high spectral and spatial resolution, offers great prospect in such complex analyses and therefore, its use has expanded in the analyses of ocean and inland waters in recent years. Hyperspectral images have been used in bathymetry (McIntyre et al., 2006), estimating chlorophyll and water constituents (Albert and Gege, 2006), mapping coral reefs and benthic h abitats (Bertels et al., 2008 ). This research focuses on filtration of high resolution hyperspect ral imagery and its applications in characterization of the natural aquatic environment. It explores effective


18 and efficient way of de striping and filtering high spatial resolution hyperspectral imagery. It also explores analy sis of hyperspectral imagery for estimation of water quality parameters including chlorophyll and detection and discrimination of submerged aquatic vegetation. 1.2 Statement of Problems 1.2.1 Filtering H yperspectral I magery Hyperspectral and pushbroom imagery systems are subject to in track str iping as each line in the image is acquired when a narrow slit of incoming light is detected by hundreds and thousands of detectors in a cross track array. Striping is of non periodic nature caused by differences in detector responses to incident light. I mage striping degrades the image quality and diminishes signal to noise ratio. De striping is an important pre processing step which is performed through either sensor calibration or image enhancement. Absolute radiometric calibration is conducted and sens or calibration coefficients are determined which when applied to captured images results in stripe reduction. But detector responses in aerial and ground based hyperspectral systems may change over time and in frequent intervals. This requires sensor calib ration to be conducted frequently which might not be feasible practically and economically. Alternatively, methods based on detector response matching or filtering are used. Response matching algorithms, including s imple linear matching (Horn and Woodham, 1979), histogram modification and matching algorithms (Weinreb et al., 1989; Wegener, 1990), are designed to match each sensor response to a typical reference response in order to de stripe the image which is based on the assumption that each detector view s statistically similar sub scenes. This assumption is valid and the algorithms are effective only when the image area is large and have periodic stripes


19 such as Landsat imagery. For higher resolution hyperspectral imagery with non periodic stripes, these algorithms are less effective. Filtering in spatial domain has been used since earlier studies to reduce stripes and noise. Directional average filter and low pass filters used in reducing stripes are associated with loss of image information (Chen et al., 2003). Filtering in transformed domain has been more popular in modern filters. Frequency (Fourier) domain filtering is frequently used in removing striping effects especially of periodic nature. Striping pattern is captured in frequency domain as an elon gated pattern perpendicular to stripe direction and periodic stripes and appears at specific frequency position allowing filtering periodic striping patterns more effectively. However, in the case of non periodic stripes as in hyperspectral imagery, stripe s are mixed with some image information in frequency domain leading to some information loss. Most recently, filtering in wavelet transform is getting increased attention as an efficient preprocessing tool for remote sensing. In this approach, the image i s decomposed into wavelet components in different scale space levels and filtering is done on selected components in selected levels. Since Donoho and Johnstone (1994) introduced wavelet based de noising technique, it has been experimented, expanded and es tablished as reliable filtering tool in remote sensing (Fukuda and Hirosawa, 1999, Lu et al., 2007; Chen et al., 2003). Wavelet transform decomposes the image into different scale spacce levels and filtering is performed on these directional components. D espite great filtering potential, there are only few studies on wavelet based de striping. Torres and Infante (2001) introduced wavelet based de striping method by eliminating directional wavelet detailed components in the striping direction while other re searchers (Chen et al., 2006, Munch et al., 2009) refined the method by filtering rather


20 than eliminating the wavelet components. Still these methods produce better results for images with low resolution or periodic striping patterns. But for high resoluti on imagery with non periodic patterns, these methods still have limitations as they will either smoothen the image details or introduce the image artifacts. De striping in transformed domains has better prospects but requires further study developing or re fining the filter for non periodic stripes in high resolution hyperspectral or pushbroom sensor systems. Filtering hyperspectral imagery is further complicated due to presence of hundreds of bands. Filtering and parameter selection for each image band cons umes much time, effort and computing resources. Applying de striping or filtering algorithms on the dimension reduced transform such as Principal Component Analysis (PCA) or Maximum Noise Fraction (MNF) might be efficient. But this requires further research to eval uate the efficiency and effectiveness of such approach. This research explores a better and efficient filter for hyperspectral imagery. 1.2.2 Analyzing H yperspectral I magery for E stimating W ater Q uality P arameters and D etecting A quatic V egetation Water quality condition and aquatic vegetation are important ingredients of an aquatic environment and play an important role in an aquatic ecosystem and production. Water quality parameters such as Chlorophyll, Nitrogen and Phosphorous are important indicators of water quality These parameters are also interrelated with the a nthropologic impact on natural waters. Chlorophyll a concentration is of particular interest in natural water management. It is an indirect measure of algal biomass in water and also depends upon t he amount of Nitrogen and Phosphorous in water. Aquatic vegetation is found in natural waters and contributes to the ecosystem by providing oxygen, food and habitat to other organisms. Proper quantification of these aquatic


21 components is therefore integral to monitoring and management of aquatic ecosystems. The water column constituents and submerged aquatic vegetation alter the light propagating through and back from water. This alteration can be assessed and interpreted remotely. Remote sensing can thus p rovide a time and cost effective tool for frequent or large scale monitoring and characterization of aquatic environment. Many remote sensing studies have been conducted focusing on case 1 or open O cean waters. Case 2 waters, represented by turbid, near sh ore and estuarine and inland waters, further complicates the interpretation due to higher and varying concentration of organic matters and suspended particles. Hyperspectral imagery has great potential in the study of such aquatic environment. This resear ch is focused in the applications of high spatial resolution hyperspectral imagery in the quantitative analysis of chlorophyll and submerged aquatic vegetation in case 2 waters. Chlorophyll is an optically active ingredient that alters the light spectra an d can be analyzed from imagery. Estimation of chlorophyll concentration using remote sensing technique is challenging as various water constituents including sediments and organic matter along with chlorophyll influence the optical properties of water. In case 2 shallow waters, higher concentration of water constituents and reflection from the bottom must be considered. Various methods of estimating Chl a concentration in ocean waters have been proposed utilizing satellite imagery in early studies (Morel and Prieur, 1977) and using airborne imagery in later studies (Gitelson et al., 2008). Two bands or three bands indices are proposed. Most available methods were not tested in the case of very high concentration of Chl a, such as those in aquacultural pon ds. This study experimented the use of ground based hyperspectral imaging system in the estimation


22 of Chl a and nutrient (Nitrogen and Phosphorous) concentrations in highly eutrophic environment. Remote sensing can be used as a tool for detecting and disc riminating submerged aquatic vegetation by analyzing alteration in radiation spectra due to vegetation. Hyperspectral imagery can offer better interpretation due to abundant spectral information. Analysis of radiation spectra is, however, complicated by in herent optical properties (IOPs), optical geometry, varying water depths and optical properties and structure of plant species. If sufficient field samples are available, supervised classifier may be applied to original image data or normalized data. Lin ea r classifier such as Maximum L ikelihood, despite its popularity in upland classification, is not useful in applying to raw image data in aquatic region as light interaction and attenuation in water are non linear. Spectral Angular Mapper (SAM) is a popular hyperspectral classifier which is known to be insensitive to illumination as it compares only vector direction but not the magnitude between two spectra. But it needs to be investigated if and how well it can match spectral difference caused by variation in water depth and IOP variation. In recent years, non parametric classifiers such Artificial Neural Network (ANN) and Support Vector Machine (SVM) are gaining attention in classification and pattern recognition. Due to non parametric and non linear interp retation of training spectra of all sample pixels, ANN and SVM might be suitable options to be applied in classifying image data in aquatic region. There is a lack of studies that utilize ANN and SVM non parametric classifiers in the classification of aquatic region and vegetation. Further research is needed to evaluate the performance of classifiers on imagery normalized for glint and depth


23 effects. Image norma lization might be performed based on radiative transfer model. Semi analytical models of radiative transfer, which relates how light undergoes changes while propagating trough water column, have been utilized in many studies of aquatic systems (Lyzenga 197 8; Lee et. al., 1998). These models may be utilized for image normalization of aquatic region. Sky glint is the portion of reflectance which is directly reflected back from water surface and information about water column or bottom layer. One of the simplest methods is to deduct image data to deep water data band by band assuming that signal received from deep water is mostly glint. Depth normalized image can be created using simple radiative transfer models such as one proposed by Lyzenga (1978). Lyzenga used simplified model in which Landsat image bands are transformed into logarithmic bands, a pair of which from field sample regions are used to determine attenuation coefficient ratios in order to retrieve bathymetr y and generate depth invariant image. Schweizer et al. ( 2005 ) expanded the method using three pairs of bands from visible wavelength range and used parametric classifier to classify bottom layers. Further study is required to explore its application to sha llow fresh waters with reliable accuracy which might be achieved using more band pairs of hyperspectral imagery combined with various classifiers including parametric and non parametric classifiers which is proposed in this research. 1.3 Goals and Objectives T his research is focused on two aspects of hyperspectral imagery filtering and analyzing. Two main goals of the study and subsequent objectives to fulfill these goals are listed below:


24 GOAL 1 : To develop effective methods of filtering high resolution hype rspectral imagery. Objectives To develop filtering algorithm based on wavelet transform to filter (de stripe and de noise) high resolution hyperspectral or pushbroom imagery. To optimize filtering high resolution hyperspectral imagery in MNF domain using w avelet based de striping. GOAL 2 : To analyze hyperspectral imagery for the estimation of Chlorophyll and for the discrimination of submerged aquatic vegetation in case 2 waters. Objectives To analyze ground based hyperspectral imagery for the estimation o f Chlorophyll a concentration in case 2 waters of aquacultural ponds. To analyze airborne hyperspectral imagery for the detection/classification of submerged aquatic vegetation in river estuarine ecosystem. 1.4 Organization of the D issertation This dissertatio n is organized in several chapters as research article, each of which is related to one of the research objectives described above. Each chapter is divided into section s including in troduction, methodology, experi ments, results and conclusion. The l ast chapter (chapter 6) concludes the dissertation with overall discussion and scope for future studies. Chapters 2 and 3 focus on developing better and efficient algorithm/method of 2 describes an


25 algorithm for de striping non periodic stripes from hyperspectral and other pushbroom images based on wavelet decomposition and adaptive fourier zero frequency normalization and compares the results with other available methods. Chapter 3 i nvestigates suitability of applying the wavelet based de striping algorithm to the first several bands of minimum noise fraction transform (MNF) of the hyperspectral imagery. This allows filtering hyperspectral imagery efficiently as the algorithm can be a pplied on fewer MNF bands rather than on hundreds of image bands. The s the use of hyperspectral imagery in quantitative analysis of components of aquatic ecosystems. Chapter 4 details the analysis of hyperspectral imagery of aquacultural ponds in which methodology and indices are proposed for accurate estimation of Chlorophyll a in highly eutrophic waters. Chapter 5 investigates analysis of hyperspectral imagery in discriminati on of submerged aquatic vegetation in black water margining riverine and estuarine systems. Various classifiers including non parametric ones and various methods of image normalization based on radiative transfer model are utilized and tested.


26 CHAPTER 2 DE STRIPI NG HYPERSPECTRAL IMAGERY USING WAVELET TRANSFORM AND ADAPTIVE FREQUENCY DOMAIN FILTERING 2.1 Introduction Remote sensing imagery is frequently contaminated with stripes, mainly due to differential variations in detector sensitivity to perceived energy. Whiskbr oom scanners, such as Landsat Thematic Mapper (TM) and MODerate Resolution Imaging Spectroradiometer (MODIS) are subject to cross track striping due to detector failure or differences in response to incident light (Chander et al., 2002; Wang et al., 2006). Such patterns are periodic and correspond to the number of detectors per scan (Schowengerdt, 2007). Pushbroom sensors are subject to in track striping without scan periodicity as each line in the image is acquired simultaneously by hundreds or thousands o f detectors in a cross track array. Most hyperspectral imagery utilizes optical dispersion techniques to disperse the light entering through a narrow slit into a two dimensional array, where one dimension represents the spatial component and the other dime nsion represents the spectral component (e.g. Anger et al., 1996; Cocks et al., 1998). In these systems, each captured image represents a line in the formed hyperspectral image. Adjacent lines are captured by moving the sensor to form a hyperspectral image similar to images acquired by traditional pushbroom scanners and subject to the same striping effect. Image striping degrades image quality and risks its suitability for analysis. Image de stri ping is one of the standard image preprocessing steps that is performed either through sensor calibration and/or image enhancement. Absolute radiometric calibration is conducted before launch for space borne sensors. For airborne and ground based


27 sensors, absolute radiometric calibration is conducted periodically. However, cost and logistic issues may reduce the frequency of such calibration. Although, absolute radiometric calibration provides data that can be used to estimate sensor response, and hence lea ds to de striped images, sensor detectors frequently deviate from their calibration values, which results in striped images. Over the years, many algorithms were developed to de stripe remote sensing images captured by multiple sensors. Simple Linear Match one of the earliest methods used for image de striping (Horn and Woodham, 1979; Weinreb et al., 1989; Wegener, 1990). Gadallah et al. (2000) utilized a moment matching algorithm to de stripe Lan dsat satellite imagery by matching each sensor response to a typical response. Such algorithms are built on the assumption that each detector views statistically similar sub scenes and requires large (extended) images (Gadallah et al., 2000). Additionally, some of the histogram matching algorithms tend to clip the high and low ends of the histograms, which results in a loss of image details (Singh, 1985). Filtering has been widely used to reduce image striping. Filtering methods can be broadly divided into wavelet decomposition, and Component Analysis (CA). Image domain low pass disadvantage of removing some image information (Chen et al., 2003). Oimoen (2000) pass and high transformed domain, Mini


28 non directional noise reduction (Green et al., 1988). Filtering in Fourier transform periodic image striping artifac proposed Fourier transformation to characterize and reduce striping effects of Landsat images. Liu and Morg achieved satisfactory results in the case of periodic stripe patterns, such as the ones resulting from s ingle detector malfunctioning in Landsat images (Schowengerdt, 2007). Striping pattern is captured in the power spectrum as an elongated pattern in the direction perpendicular to the stripes. However, it is mixed with some image information, especially in the case of non periodic stripes. Filtering such frequency domain patterns and Morgan, periodic stripes, as in the case of pushbroom an d hyperspectral images, possess a range of low to high frequencies. Filtering these frequencies leads to distorted image content and overall image blurriness. noising and de based on decomposing the image into components in different scale space levels and


29 e decomposed components in these levels. Unlike Fourier transform analysis, where there is no direct link between the image representation in the spatial and frequency domains, wavelet decomposition maintains this link Donoho and Johnstone (1994) proposed wavelet based de noising techniques that reduces noise through hard and soft thresholding of wavelet detail components. Fukuda and Hirosawa (1999) authors highlighted h 2003). Torres an d Infante (2001) used wavelet transformation to analyze and reduce striping by eliminating directional wavelet detail components in the striping direction. This process is accompanied by a loss in other image information captured in the eliminated wavelet components. Other researchers (Chen et al., 2006; Munch et al., components. Wang and Fu (2007) developed a method inspired by the striping power spectrum properties and based on F based de convolution. Despite being a promising approach, wavelet based methods can produce unsatisfactory results as they have the potential to alter the content of the wavelet directional component and intro duce image artifacts. In this research, we propose an algorithm for de striping hyper spectral images based on wavelet decomposition and adaptive Fourier zero frequency normalization. In our method, the stripe dominated directional wavelet components are


30 stripe features in the image wavelet components. We improved on existing wavelet based de striping methods (e.g. Torres and Infante, 2001; Munch et al., 2009) to overcome many of the artifacts introduced when de striping high resolution edge rich remote sensing images. We tested our method on images captured by a hyperspectral imaging sensor for a water body and vegetation scene, a Hyperion image and an image contaminated with s imulated noise. We also tested a technique integrating our developed de striping method with a standard wavelet based de noising algorithm. We compared the results with three other de striping methods including the traditional simple linear matching method and more recent wavelet based methods using visual and quantitative assessments. This paper is organized as follows. Section 2.2 provides the theoretical background for the Fourier and wavelet transforms used in this study. In section 2. 3, our developed de striping methods are described and other methods used for comparison section we also discuss the quality assessment metrics used in the research. In section 2. 4, the test images used in this study are introduced, and sele Section 2. 5 presents and compares the results of all tested methods using visual and quantitative assessments. Finally, section s 2. 6 and 2. 7 discuss achieved results and provide the research conclusion, re spectively. 2.2 Background resolution hyperspectral images. This type of noise occurs in the image acquisition direction (on tr ack direction) and is non


31 periodic in nature. Wavelet transform, as a multi resolution and scale space representation of the image, has the ability to separate directional noise in certain directional wavelet components at different scale levels. On the ot her hand, Fourier transform can analyze image noise in the frequency domain. As these transforms were incorporated in our proposed method, basic concepts of these transforms and their 2.2.1 Fourier T ran sform different frequencies, amplitudes and phases. It transforms the image from spatial freq uency value that is the mean amplitude of the signal, also known as the Direct Current (DC) (Schowengerdt, 2007). For analyzing a discrete signal, Discrete Fourier dimensio nal DFT is mathematically expressed as: 1 (2 1) where the complex exponential term represents the sum of sine and cosine terms and i is imaginary number. F k Fourier transform generates real and imaginary components and can also be represented by amplitude and phase components. DFT is invertible and the image can be perfectly reconstructed using the Inverse Discrete Fourier Transform (Inverse DFT) expressed as: 1 (2 2)


32 DFT can be easily expanded to two dimensions (2 D) or multi dimensions by composing sequences of one dimensional DFTs along each dimension is in any order. 2 D DFT is expressed as: (2 3) where F kl Image stripes are condensed in the frequency domain to a narrow central band of high amplitude values in a direction orthogonal to the stripes. For instanc e, vertical stripes are presented as a horizontal central narrow band in Fourier domain. Filtering in Fourier transform involves examining and locating noise frequency components in the on the image spectrum and inverting back to image domain to obtain the de noised on. hard mask these methods r non periodic stripes, loss and distortion of image information can occur. In our proposed


33 individu al rows (or columns) of the wavelet components of the image dominated by the stripe noise as described in sub section 2. 3.1. 2.2.2 Wavelet T ransform Wavelet transform represents any arbitrary function as a superposition of wavelets (Antonini et al., 1992). Unlik e Fourier transform, wavelet transform retains both spatial and frequency information. Wavelet transform utilizes narrow groups of wavelets of different shapes (e.g. Debauchies, Morlet, Haar, Maxican Hat, etc.) that represent local and non periodic pattern s of a signal better than Fourier transform (Gonzalez and Woods, 2008). Unlike the continuous sinusoidal wave function that Fourier uses, a wavelet is a brief oscillation function that is localized in space. Wavelet transform is used in multi resolution an alysis to obtain different approximations of a signal function f(x) at different levels of resolution. (2 4) where a and b are scale and translation parameters along x axis. And function of transform or admissible wavelet is obtained by scaling and translation of a mother wavelet : (2 5) In discrete signal and according to the d yadic sampling, a and b are considered 2 j and k2 j (2 6) where j and k determine the position and width of wavelet along the x axis.


34 In multi resolution analysis, a scaling function x ) is used to create a series of approximations of the function; each differs by a factor of 2 in resolution. Wavelet functions are then used to encode the difference in information between approximation s (Gonzalez and Woods, 2008). Wavelet function can be expressed as a weighted sum of shifted, double resolution scaling function: = n h n) (2 7) where the h can be related to by the equation h (n) = ( 1) n h (1 n). Mallat (1989) developed Fast Wavelet Transform (FWT) algorithm for fast and e Wavelet Transform (DWT). It is based on the (W ) at scale j + 1 can be derived as: (2 8) (2 9) can be obtained by convolving W with the reversed scaling and wavelet vectors, h ( n) and h ( n) followed by the subsequent subsampling. In one dimensional (1 D) multi resolution analysis, signal f is decomposed into an approximation (low pass) and detail (high pass) components in one scale level. After decomposition, the size of thes e components is halved by down sampling. The approximation component can undergo iterative decomposition in the next scale level.


35 The theory of multi resolution analysis and wavelength can be extended to 2 D or higher dimensions. Two dimensional wavelet t ransform is a multi resolution, scale space representation of 2 D data, such as digital images. It is represented by one scaling function and three directionally sensitive wavelet functions at each scale level. scaling funct ion H wavelet horizontal function V wavelet vertical function D wavelet diagonal function H V D (x, y) represents directional wavelet functions in horizontal, vertical and diagonal directions, respectively. Two dimensional wavelet transform (decomposition) can be obtained by taking the 1 D DWT of the rows of the signal data f(x, y) and the subsequent 1 D DWT of the resulting columns. In a signal scale process, four quarter size sub images, one H W V W D ) are is further decomposed at the nex t scale level. In multi resolution scale levels, the signal is decomposed into one approximation H W V W D ) captures the horizontal variations (horizontal stripes, edges), vertical variations (vertical stripes, edges) and the variations along the diagonal directions, respectively. Wavelet capability of decomposing an image into directional detail compon ents in multiple scale levels is advantageous in detecting and eliminating noise and stripes.


36 Noises are detected in detail components of low scale (high frequency) levels while per iodic directional stripes have different frequencies and appear at many scale levels. Filtering ltering is performed in three main steps: (i) decomposition of image into wavelet components in the fact that directional components also capture non stripe signals, including contrasting feature edges in addition to the stripes. Non stripe signals become more er frequency) levels and blend with stripe information in reducing non periodic stripes. Results of available de striping algorithms show blurred or smear like artifa cts in the image area with contrasting features as will be discussed further in Section 2. 5. In this study, wavelet decomposition was used to separate directional stripes at multiple scale level components in preparation for a subsequent Fourier based adap 2.3 Methodology In this section, our proposed de striping method and combined de striping and de noising method are described, followed by a brief description of other de striping methods utilized in this study for comparison. Finally, quality assessment techniques adopted in this research to evaluate our proposed methods are described.


37 2.3.1 Proposed A daptive D e striping M ethod Our de striping method is based on a combination of wavelet decomposition and ering. In our method, the raw image is decomposed into wavelet components in a given number of scale levels. The wavelet directional domain. Discrete Fourier transfo rm (DFT) is applied on individual rows or columns (along stripe direction) of the stripe applied by detecting and excluding potential non stripe extreme pixels and normalizing the DC values adaptively in order to suppress the stripe effect only. Earlier efforts by stripe signal in the Fourier component, which led to image artifacts in remote sensing images. An inverse Fourier transform is the n applied to form the de striped wavelet component at each striped image. The proposed method, which is named WFAF (Wavelet Fourier Adaptive Filter) in the remaining sections of the ma nuscript, was implemented through custom Matlab ( ) and ENVI IDL ( tServices/ENVI.aspx ) code, which can be obtained through direct contact with the authors. A schematic diagram of the proposed WFAF method is illustrated in Figure 2 1. Please note that the de noising box in the bottom of Figure 2 1 is related to the combin ed de noising/de striping method described in sub section 2. 3.2. In the following, the two major steps in our adaptive de striping method are described. Wavelet decomposition First, the image is subjected to Discrete Wavelet Transform (DWT) and decomposed into a number of frequency scale levels. At each level, three directional components (horizontal, vertical and diagonal) are formed in


38 addition to the image approximation com ponent. This process separates the stripes (in addition to other image contents) in the detail wavelet components in the stripe direction (e.g. vertical stripes in the vertical detail components). Non periodic stripes, as in hyperspectral imagery captured by pushbroom sensors, appear in a number of scale levels. For most wavelet implementation, determining the wavelet type and number of scale levels is an experimental step that varies according to the stripe and image characteristics as discussed in sub s ec tion 2. 4.2. Figure 2 2 demonstrates three level wavelet decomposition of an image captured by a ground based hyperspectral sensor vertical components of the decompo sed image at all three scale levels. These e striped image. Figure 2 1. Schematic diagram illustrating proposed adaptive de striping and combined de striping and de noising methods.


39 A B Figure 2 2. Wavelet decomposition of an image. A ) Band 30 (wavelength = 436 nm) of a close range B) T hree level wavelet decomposition (produced using Matlab Wavelet Toolbox function). Fourier domain adaptive f iltering In this step, the directional wavelet compon ents (vertical or horizontal), corresponding to stripe direction, are transformed to the frequency domain using 1 D Fourier transform. The components are transformed as individual vectors, where each vector contains the digital numbers of a single column o r row in the striping direction. For the purpose of this introduction, we will assume that vertical component of the image wavelet transform. This process emphasize s the stripes in the frequency domain as variations in the DC (Direct Current or, amplitude value of the zero frequency) values of the Fourier transform of individual column. stripe the striped directional wavelet components by equalizing the DC value of each column Fourier


40 transform. One of the approaches that can be used to perform such normalization is to set the DC values to zero (or any constant value). The DC value represents the mean ampl itude of the signal column and is proportional to the column mean in the spatial domain. Equalizing the DC values in this sense can introduce artifacts, especially for small images and for images with highly contrasting features. Edges of such features (no n algorithm. Results may have less striping, however, smearing artifacts will be introduced due to the effect of the DC equalization process that alters values in the original i mage stripe signals in the directional wavelet component when normalizing the DC values in contrast feature edges) in the directional wavelet components are detected in each the column neighborhood me an. A new vector Y j free from these pixels is created by (2 10) where x i j j is the statistical mean of the digit al elet type, number of scale levels L, and threshold value k) need to be determined based on the


41 image content and stripe characteristics. The selection criteria and sensitivity analysis of these parameters are discussed in sub section 2. 4.2. lized DC value for each column can be computed as proportional computed as follows: F j norm = F j orig j j j (2 11) where F orig and F norm j is the mean of the column after excluding contrast features in the wave norm is zero; otherwise, the Figure 2 the second level wavelet component ( Figure 2 3 A ) of the water body image shown in Figure 2 2 A Figure 2 3 B shows the 1 D Fourier transform of the wavelet component columns. The striping effect is highlighted by the variations in the zero frequency (DC) component (central value of the 1 stripe signals (using a threshold k = 1.5) are shown in Figure 2 3 C Figure 2 3 D shows the de the Inverse FFT algorithm to reverse to the wavelet domain. Figure 2 3 D represents the striped image of only one vertical wavelet component of the original striped im age. This process should be applied to the vertical wavelet components of a number striped image.


42 Figure 2 3. Schematic diagram of proposed wavelet Fourier adaptive filter (step 2) applied on image of wate r body with floating targets. A ) Level 2 vertical wavelet component. B ) 1 D fa st Fourier transform. C ) Signal from hi gh contrast features (black). D ) De stri ped vertical wavelet component. 2.3.2 Combined D e striping and D e noising M ethod The proposed de striping method described in sub section 2. 3.1 is designed to reduce striping patterns only. Although this process noise content based on wavelet transform. Unlike wavelet based de striping methods that are implemented on wavelet components in the striping direction, wavelet based de noising algori thms are normally implemented on all wavelet detail components at certain wavelet decomposition level(s). Implementing wavelet based de striping algorithms encourages integrating other de noising algorithms in this process to achieve de striped and de nois ed image. We tested the use of a de noising algorithm suggested by Donoho and Johnstone


43 (1994), which utilizes soft thresholding with universal global threshold applied on the wavelet details component combined with the de striping algorithm (as shown sche matically in Figure 2 1). The de noising algorithm is applied on all three wavelet wavelet detail components are de c omponents are subjected to a soft thresholding process, where the detail value is set to zero if it is less than a certain threshold determined from the image noise level. Otherwise the detail value is set to the difference between the detail and the thres hold hereafter, was applied to evaluate the effectiveness of combining the proposed de striping method with a de noising algorithm on reducing both stripe and non systematic/non directional noise from the images. Using a wavelet based de noising algorithm that can be easily integrated within the proposed de striping algorithm makes 2.3.3 Other U sed D e stripin g M ethods We compared the performance of our proposed method with three other methods. One of these methods utilizes a spatial domain linear matching algorithm proposed by Horn and Woodham (1979). The other two methods are wavelet based and represent a bas Spatial f ilter: s imple l inear m atching (SLM) In the spatial domain, low raditionally used in stripe reduction. Directional low


44 including SLM, histogram matching, and moment matching assum e each detector to view statistically similar sub scenes which is a valid assumption for large low resolution images. For example, histogram and moment matching methods assume that each sensor detects all varieties of brightness values, which is valid in t he case of large number of pixels are available per each sensor. Despite its limitations, simple linear matching method (Horn and Woodham, 1979) is frequently applied and is used in this manuscript as a bench mark to assess the performance of our method co mpared to other de that each detector column in the clean image has similar mean and standard deviation and adjusts the stripe image pixels accordingly. Wavelet based f i lter Wavelet based de striping algorithms are promising due to the wavelet inherent ability to decompose the image into directional multi frequency components that separate the stripes in certain direction. Torres and Infante (2001) proposed an algorithm to red uce stripe noise by eliminating (zero padding) the wavelet detail components in the stripe direction under the assumption that by dropping any component of an orthogonal transformation, the results will still approximate the original image in the least squ are sense. In other words, it preserves the radiometric level of the image. The algorithm was tested on a Landsat image contaminated with periodic stripes and yielded satisfactory results. In the case of hyperspectral images, the stripes are non periodic w hich means stripes might appear in several scale levels and dropping stripes) leading to an over smoothed and distorted image signal. The algorithm requires two para meters (wavelet type and number of decomposition levels).


45 rather than eliminating the directional wavelet components that capture image stripes. Chen et al. (2006) targeted obliq components. The authors suggested that wavelet components induced by oblique stripes may be located by summing the frequency power spectra of each column or row and then averaging the sum. This method wa s designed mainly to handle periodic stripes in any direction. Many image stripes are non periodic and can appear at multiple leads to appropriate elimination of image stripes. Munch et al. (2009) extended this The algorithm requires three parameters (wavelet type, number of levels L, and the lgorithm was tested on tomographic imaging, such as X ray microtomography with good results. Unlike tomography images, high resolution remote sensing images may contain more edge features (directional details). In such case, wavelet directional components might contain details from highly contrasted directional features in addition to the stripes. Without detecting such feature In this research, we used the Torres and Infant e (2001) and the Munch et al. (2009), which are denoted the Torres and Munch methods, in addition to the Simple section 2. 3.1) and the comb ined de striping and de noising method (sub section 2. 3.2). Although, we experimented with these methods using different parameters (e.g. wavelet types and levels), we presented only


46 the results that provided best de striping while preserving image content The quality assessment techniques utilized in this evaluation are described in sub section 2. 3.4. 2.3.4 Quality A ssessment Visual assessment Visual assessment constitutes a primary step in assessing the quality of hyperspectral images and selecting the bands suitable for analysis. Extending visual assessment to evaluate the quality of image de striping and to help determine de striping parameters is not new to the hyperspectral image processing community. In this research, the de striped images resulting from applying different methods (and method parameters) were visually assessed and compared. For the test image with introduced (simulated) noise, de striping output was compared to the original clean image (see sub section 2. 4.1). For all tested images, differ the direction perpendicular to stripes were used to show changes in local brightness levels, reduction of noise and induced distortions. Quantitative assessments Root Mean Square Error (RMSE) and Peak Signal to Noise Ratio (PSNR). When a clean image is available (simulated noise case, see su b section 2. 4.1), Root Mean Square Error (RMSE) and Peak Signal to Noise Ratio (PSNR) are commonly Huynh Thu and Ghanbari, 2008). RMSE is used as an image noise measure and can be conveniently computed as follows: (2 12)


47 (2 13) where M and N refer to the image number of rows and columns, respectively; p(i,j) and q(i,j) are the pixel value for the clean and processed image, respectively, at row i and column j. PSNR, which is commonly used as a measure of signal (or image) quality, is computed as the ratio of maximum possible power of a signal to the power of corrupting noise and expressed in logarithmic decibel scale as expressed in the equation below: PSNR = 10 log 10 (MAX 2 / MSE) (2 14) where MAX is the maximum possible pixel v alue (e.g. 255 for 8 bit image) and MSE is the Mean Square Error described in E quation 2 12. Frequency domain Signal to Noise Ratio (SNR) estimation. Signal to Noise ratio is estimated indirectly in Fourier domain using selected array of DN values (Torres and Infante, 2001). For remote sensing images with no clean image available, SNR cannot be directly measured to provide quantitative measure of image quality. An array homogenous image section across the stripe direction can be used to estimate SNR values assuming that most high frequency changes in brightness come from stri pe and D Fourier amplitude values. The SNR (signal to noise) ratio was then estimated as follows (Torres and Infante, 2001): SNR (in dB) = 10 log 10 (A signal /A noise ) = 10 log 10 A signal 10 log 10 A noise (2 15) SNR Improvement = SNR destriped SNR original (2 16)


48 where A signal is the signal amplitude considered as the central component of the signal frequency vector and A noise is the noise amplitude considered as noise peaks. Figure 2 4 shows an example of signal and noise estimated from an image row. High frequency stripes and noises are related to high frequency noise peaks, while related artifacts are relat ed to low frequency noise peaks. Therefore, high frequency noise peaks and low frequency noise values were considered to compute SNR low (signal to low noise ratio) and SNR high (signal to high noise ratio), respectively. SNR values (in dB) and SNR improveme nt were computed for rows in the original and de striped images to evaluate the effect of the de striping algorithm. Figure 2 4. Frequency domain signal and noise. A ) Horizontal profile of original pond image at row 840. B ) Amplitude values (dB) in f our ier domain. 2.4 Imagery and Experiments We applied the proposed WFAF and Combined de striping method in addition to other tested algorithms (SLM, Torres, and Munch methods) on several test images. In our testing, we used an image with simulated noise, two hype rspectral images captured


49 by a ground based hyperspectral sensor, and a Hyperion satellite image sub scene. In the following sub sections, we introduce the tested images and present our experiments in determining and analyzing the sensitivity of our WFAF m parameter values. 2.4.1 Used Images We used an image with simulated noise and three hyperspectral images for testing striping n Figure 2 5. A clean high resolution image quality consumer grade Nikon camera (Fig ure 2 5 A ) was contaminated by adding random and striping noise (Fig ure 2 5 B ). The original Nikon image is a single frame image that is free from striping noise. We introduced simulated random and stripping noise by adding zero mean Gaussian white noise for each column followed by another Gaussian noise applied to all pixels in the image. It should be mentioned here that we experimented with multiple noise levels directional noise, respectively, to present in the manuscript, as they visually resembled the noise levels experienced in our tested hypersp ectral images. Quality assessment noise) and ground based hyperspectral sensor were also used as test images in this study. The sensor is composed of an ImSpecter imaging spectrograph made by Spectral Imaging Ltd. (www.s 2M30 Lynx Charge Coupled Device (CCD) camera made by Imperx Inc. ( As a pushbroom


50 scanner, each image captured by the Imperx camera forms a line in the created hyper spectral image. Adjacent lin es are captured, either by moving a sweeping mirror mounted in front of the sensor, or moving the sensor itself. The sensor provides 205 bands covering the Visible and Near Infrared (VisNIR) parts of the spectrum (400 1000 nm). Image striping occurs due to light energy, especially in low light conditions. This striping noise (in addition to other types of noise) increases towards both ends of the sensor sensed spectrum range primarily due to lower signal to noise ratio caused by the optical dispersion characteristics of the spectrograph. One of the used images (the pond image) is taken water quality assessment experiment. Band 30 (wavelength = 436 nm) of this image ( Figure 2 5 C ) was analyzed. The second image was for a vegetation scene taken as part of a study to identify invasive plant species. Figure 2 5 D shows Band 32 (wavelength = 440 nm) of this image with some light s triping noise. Finally, we tested http://aviris.jpl.nasa. gov/data/index.html Only a subset of the image was used in this study as shown in Figure 2 5 E 2.4.2 Selection of Fi lter P arameters and A lgorithm I mplementation Our de striping algorithm requires three parameters: (i) wavelet type, (ii) number of decomposition Selecting the number of decomposition levels (L) depends mainly on the on the stripe frequency decomposition characteristics. Non periodic stripes can possess multiple


51 frequencies and appear in a number of decomposition levels. The threshold (k) value create i mage artifacts. Choosing low k values means more pixels are excluded as extreme non stripe edges, while higher k values allow for these pixels to bias the striping method on the image induced with simulated nois parameters and to assess the sensitivity of the WFAF method, we experimented with multiple types of wavelets and different orders of the Daubechies (dbN) (Daubechies, 1988) wavelet family (e.g. db4 and db8). We recognize the large number of wavelet functions that can be used in this respect and note that the selected types were used wavelet functions is cumbe rsome and outside the scope of this research. We tested different wavelet levels (L = 2, 3, 4, 5, and 6) and threshold values (k = 0.5, 0.7, 1, 1.2, 1.5, 1.8, and 2). Filtering quality is assessed visually and quantitatively (RMSE and PSNR). Figure 2 6 sho de Figure 2 7 shows graphs of PSNR versus k val ues using different decomposition levels (L) and two wavelet types (db4 and db8). Figure s 2 6 C to 2 6E 3, 4, and 5) and same wavelet db4 and k =1. Figure 2 7 shows that images corresponding to L = 4 yielded slightly higher PSNR values when compared with those


52 Figure 2 7 also shows that the PSNR value increases with the increase in the number of wavelet levels until it reaches a maximum at certain L value, beyond which the PSNR value starts to decrease. Figure s 2 6 E to 2 6 G Daubechies wavelet family types db4, db3, and db8, respectively, with same values of (L = 5) and (k = 1). Visual assessment shows that the stripes were greatly reduced and image information was preserved without visible distortions. Despite the use of dif ferent less sensitive to the wavelet types used in this experiment. Nonetheless, db4 w avelet yielded slightly higher PSNR values and is therefore chosen for further experimentations involving this image. Based on our visual assessments to all test results (only few are presented in the manuscript due to space limitation), we found the db4 w avelet to give the best results for all tested images except for the pond image, where the db8 wavelet gave the best results. Figure s 2 6 H, 2 6E, 2 6I, and 2 6J with four different k values 0.7, 1, 1.5, and 1.8, respectively. Visual and quantitative inspection show that k values of 1 highest PSNR is around 27.2 at k = 1 .1). Selecting k values less than 1 will identify high number of pixels as non stripe (edge) pixels, which may not be desirable. Also, higher values of k tend to introduce slight distortions in this image (see Figure 2 6J ). Figure 2 7 shows that the PSNR v alues changes rapidly with the change in L values. This


53 suggests higher sensitivity to the number of decomposition levels (L) compared to changes in k or wavelet types. Depending on the image and stripe characteristics, k values in the range of 1 1.5 were found to be adequate to avoid introducing smear like artifacts into the image. 2.5 Results and A nalysis We applied the proposed WFAF de striping and the Combined methods on all test ltering methods. Parameter selection for noise induced tomato image described in sub s ection 2. 4.2 showed that visual and quantitative assessment results were matching. For each ues yielding best results. Filter parameters were primarily selected using visual assessment. In all experiments, an L value that ranges between 4 and 6 and k value between 1 and 1.5 produced the best results in eliminating the stripes from the images. As mentioned earlier, the db4 wavelet performed best for all images except for the pond image, where db8 wavelet results were slightly better. This could be attributed to the characteristics of the pond image, which represents a mostly homogeneous surface (wa ter) with few highly contrasting objects (calibration targets). The same set of parameters was used to implement the Combined method with the de level only. We compared these results with the results of implement ing the SLM, Torres, and Munch methods. 2.5.1 Filtering of N oise s imulated I mage The image with added simulated noise ( Figure 2 5 B evaluated using three different me thods: (1) visual inspection, (2) RMSE and PSNR


54 this image is induced which enabled the use of the pre noised image as a reference to compute the RMSE and PSNR quantitati ve measures. This facilitated accurate Visual inspection Figure 2 8. It shows that the performances of the WFAF and Combined methods were superior to other introduced artifacts ( Figure 2 8 A ). The implementation yielded almost stripe free images and retained most non stripe information. The Combined method pro duced a similar de striping effect in addition to random noise reduction ( Figure 2 8 B ). Torres method blurry images shown in Figure s 2 8 C and 2 8 D Our experiments showe d that increasing the number of decomposition levels (L) used in the Torres algorithm reduced the stripes but caused increased loss in image details and consequently more image blurriness. Munch method performed better than the Torres method in de striping and retaining image details However, smear like distortions were introduced along the positions of the feature edges as shown in Figure 2 8 E Increasing the sigma value of striping features extended in the striping direction were f aded ( Figure 2 8 F ) due to the bl urring effect of the Gaussian Figure 2 8 G shows the image resulting from applying the SLM method. It shows that the output of this method is of the least quality among all tested methods.


55 Quantitative assessment u sing RMSE and PSNR v alues RMSE and P Table 2 1. Lower RMSE and higher PSNR values indicate better performance of the Table 2 1 shows that the proposed WFAF and Combined methods produced the highest PSNR values com techniques. The WFAF method yielded a PSNR value of 27.19, which is higher than those obtained from other tested methods and approached the value obtained by the Combined method (27.97). The Combined method produc ed the highest PSNR (least RMSE) as it effectively de striped as well as de noised the image, followed by the results obtained by our WFAF method. The PSNR value obtained using the SLM method was worst (least PSNR and highest RMSE). Torres method performed better; however, it was short of the results obtained by our methods and the Munch method. The later provided PSNR value that is higher than Torres and SLM method but less than the values obtained by our method(s) due to introduced smear like artifacts. T hese results supported the visual inspection results discussed in sub section 2. 5.1 .1 SNR of s ample profiles homogeneous part of image for the SNR analysis. SNR for high and low frequency noise were computed and presented in Table 2 2 for the original and noise induced images in addition to t SNR computed for the original clean image showed some low and high frequency noise content due to the fact that the selected row has details and there is virtually no completely homogeneous induced image showed


56 decreased low and high frequency noise or increased SNR values. Our proposed WFAF method yielded SNR values higher or compar able to the Munch method but lower than the Torres method. Visual assessment showed that the image resulting from SNR values obtained for the Torres method can be attribu image details (over SNR. The Munch method yielded lower value of low frequency SNR due to added low frequency distortions. SLM method yielded the least SNR value s. The Combined method yielded best SNR values due to its reduction of both stripe and random noise (and may be some image details). Although not perfect, SNR estimation matched the visual and RMSE/PSNR quantitative assessment results. SNR assessment can p rovide available to accompany visual assessment results. 2.5.2 F iltering H yperspectral I mages We implemented the filtering algorithms on bands of the hyperspectral images discusse d in sub section 2. 4.1. Two of the images were high resolution (0.5 1 cm pixel size) images captured by our ground based hyperspectral sensor (Fig ures 2 5 C and 2 5 D ) while the third image was a lower resolution (30 meters) Hyperion image subset ( Figure 2 5 E ). Filtering results were evaluated using: (1) Visual inspection and (2) SNR of sample profiles. Visual i nspection Figures 2 9 to 2 11 show the implementation results of applying the proposed WFAF de striping method, Combined method and the three other (Torres, Munch, and SLM) on Bands 30 and 26 of the pond and vegetation images,


57 respectively, and Band 195 of the Hyperion satellite image. Parts of the images with reduced quality and/or noticeable introduced artifacts are marked by oval shapes on the performed better than the latter three methods by reducing stripes, minimizing introduced artifacts, and preserving image information. In addition to achieving sup erior de striping while retaining image content, the proposed WFAF method maintained the local average brightness level in the resulting image ( Figures 2 9 B and 2 10 B ). The se de striped image s however, still show random noise content. This content was effectively reduced using the Combined method, which can be seen in the results presented in Figures 2 9 C and 2 10 C Close inspection of Figures 2 9 D and 2 10 D shows that the Torres method blurr ed the images and low frequency stripes are still visible. Implementing the Munch method introduced smear like distortions due to the existence of contrasting feature edges, as shown in Figures 2 9 E and 2 10 E Applying the SLM algorithm resulted in losses in image content, distorted signal and non as shown in Figures 2 9 F and 2 10 F The relatively homogeneous characteristic of the pond image allowed visual inspection of individual rows to study the effect of the shown below Figure 2 (along row 500) and spanning homogeneous areas with different contrast (water body r proposed WFAF and Combined


58 pattern. Figure 2 11 B ) was better than the results of the other methods in terms of redu cing stripes and preserving details. The Combined method yielded good de striping result. However some small details were lower part of Figure 2 11 C As experienced before, the Torres method ( Figure 2 11 D ) yielded blurry imagery with low frequency striping patterns. The Munch method ( Figure 2 11 E ) yielded adequate de striped image except for some introduced small artifacts. In contrast to our previous observations, the SLM m ethod ( Figure 2 11 F ) performed well on the Hyperion image. However, close inspection of the output reveals different levels of contrast when moving from the left side of the image to the right. This could be attributed to the difference in the column histogram characteristics of the left side of the image (dominated by a lake and relatively homogeneous natural areas) and the right side that is dominated by highly contrasting urban areas. Although our basic motivation behind this research was to develop the WFAF method to handle the stripes in ground based high resolution hyperspectral imagery, the superior performance of the WFAF method on the lower resolution Hyperion image demonstrated the potential of this method for generic application on hyperspect ral imagery with different resolution. SNR of sample profiles Two rows (30 and 830) of the pond image were selected in relatively homogenous areas of the water body to estimate Fourier domain Signal to Noise ratio. SNR values (related to low and high frequ ency noises) and SNR improvement were computed for the original and de striped rows and tabulated in Table 2 3. Comparing the estimated


59 SNR shows that the proposed WFAF method yielded SNR values higher than original low (si and performed better than other methods. This can be attributed to the stripe reduction effect without introducing artifacts. The WFAF method improved SNR high (signal to high frequency noise ratio) just sligh tly as random noise still exists after de striping. The Combined method yielded the highest SNR improvement in both high and low frequency ranges, which indicates effective de striping and de noising without introducing low frequency artifacts. The SNR high value produced by Torres method was smoothing effect. The SNR low values yielded by Torres method were lower than our proposed produces SNR low proposed WFAF method due to the artifacts introduced by Munch method. The SNR high values show slightly better values for our method. The SLM method y ielded the least SNR values in both high and low frequency ranges due to introduced distortions all over the image. For the vegetation image, columns 40 and 430 (see Figure 2 5 D ) were analyzed (Table 2 4). The proposed WFAF method yielded improvement in S NR values compared to the original image values. higher than the WFAF method due to image over smoothing. SNR low produced by the Munch method was higher than our proposed loc show in edge areas and may not affect areas where directional contrast variations are


60 not prominent. SNR high which is a measure of high frequency content, was slightly hi gher in the Munch method results, which may indicate a small smoothing effect of this method. The SLM method downgraded the image quality and provided the lowest SNR for both low and high frequencies. water body ( Row 365, Figure 2 5 E ) was analyzed to estimate the SNR values (Table 2 5). Similar to previous results, the proposed WFAF method yielded improved SNR low values. The Combined method produced SNR values that are higher than all methods except th e SNR low of the SLM method. Torres method typically bl urred the image and yielded SNR high higher than the WFAF and Munch methods. The SNR low value for Torres method was lower than the one obtained by our WFAF method, which is attributed to the low frequenc y image distortions introduced by Torres method. As noticed visually, the SLM method performed better in the case of the Hyperion image compared to other high resolution images and yielded high SNR values in both low and high frequency ranges. As discussed in the visual assessment section, SLM provided good de striping results locally. However, a noticeable brightness change appeared when moving from the western part of the image to the eastern part.


61 2.6 Discussion The visual and quantitative assessment of superior performance of our proposed WFAF de striping method in reducing image stripes and preserving image details with the least introduction of distortions or artifacts. This method outperformed other tested methods, es pecially in the case of high resolution images, yielding higher PSNR (lower RMSE) and SNR low values. The Combined method performed even better in such images and yielded the highest PSNR and SNR values, which is attributed to the reduction of random noise in addition to image striping. For the low resolution Hyperion image, the Combined method tended to handle small image details as noise, which led to slight image blurriness and a loss of details. The results of the SLM method were consistently inferior to other methods. The method often produced images that were quantitatively worse than the original striped images. The SLM method produced images with radiometric distortions that can be ition to loss of image details. This method, nonetheless, yielded better results in the case of the low resolution Hyperion image. We can assume that this statistical method could be effective if applied on large low resolution images. The wavelet based To rres and Munch methods performed bet ter than the SLM method in the case of high resolution images. However, the Torres method, which is based on dropping whole wavelet components dominated by the stripes, resulted in a significant loss in image content wh ile still retaining low frequency stripes. This method yielded PSNR values similar to that of the striped image ( Table 2 1) and caused a strong increase in the estimated SNR of the high frequency range due to over smoothing (Tables 2 2 to 2 5). A major dis


62 introduction of smear like artifacts, which were more obvious in high resolution images, especially with the existence of highly contrasting features (edges) in the stripe direction. The filtered images yielded lower PSN R value and mostly lower SNR low (signal to low frequency noise ratio) when compared with the results obtained using our proposed method (Tables 2 1 to 2 5). Similar to other filtering techniques, our method, although automated and adaptive to image characteristics, needs three filtering parameters: wavelet type, number of decomposition levels, and a filtering threshold value. Our investigation with different p arameter values showed that the method was most sensitive to the number of wavelet levels parameter with least sensitivity to the wavelet type parameter. Although, we experimented with several wavelet types, the results of the three wavelet types from the Daubechies family (db3, db4 and db8) presented in Figures 2 6 and 2 7 were promising. A number of wavelet levels between 3 and 5 and a k value between 1 and 1.5 gave adequate results for the tested images. Due to the differences in image noise levels and i nherent image characteristics, it is advised that preliminarily experiments are conducted on the images before filtering using different parameters as a preprocessing step to determine the most appropriate parameters for the filter. Our experiments showed that visual and quantitative assessments of the filtering results were mostly in line, which indicates that careful visual image inspection may be sufficient for the purpose of selecting the filtering parameters. We think that the range of parameter values suggested or used in this study can be used as a guide to initiate the filter parameter determination.


63 Implementing the de striping and de noising algorithms on individual bands of a high spectral resolution image could be computationally expensive. Fut ure research is recommended to utilize existing strong correlation between hyperspectral image bands to optimize the use of the de striping algorithms and to test their overall effect on the image spectral characteristics. Future research is also needed to test the use of 3D wavelet transform, which is mainly used in image compression and de noising applications, to de noise hyperspectral imagery. 2.7 Conclusion In this study, we introduced the WFAF de striping method based on a wavelet multi resolution image analysis and adaptive Fourier transform filtering. The method separates image stripes using wavelet decomposition and adaptively normalizes the zero frequency components of individual vectors (rows or columns) in the direction of the stripes. The adaptive nature of the method subdues the artifacts that could happen in small sized images or in the case of the existence of contrasting features. We also tested a Combined method that implements our introduced de striping algorithm with a standard wavelet based de noising algorithm to filter random noise content in the images. The Combined method is implemented using a soft threshold de noising algorithm on the first level wavelet component. We compared the results of our algorithm and the Combined method with t hree other different de striping methods visually and quantitatively. Two of these methods were based on wavelet decomposition and represented a basic wavelet de striping and a non adaptive power spectrum filtering methods. The third method was the fundame ntal simple linear matching technique. All methods were tested on an image with simulated noise, two high resolution ground based hyperspetral images, and a


64 subset of a Hyperion satellite image. We applied the quantitative assessment using Root Mean Square Error (RMSE) and Peak Signal to Noise Ratio (PSNR) on the noise induced image and the frequency domain Signal to Noise Ratio (SNR) on all other tested images. Our comparison showed that the proposed WFAF de striping algorithm performed excellently when ap plied to all tested images, which was shown visually and evidenced by the quantitative assessment results. Our testing showed the applicability of the Combined method to reduce both types of noise (striping and random), while preserving image details, espe cially in high resolution imagery. In the case of low resolution images, the Combined method should be carefully applied as small details can be lost.


65 Table 2 1 RMSE and PSNR values computed from various filtering methods. Method Parameters RMSE PSNR Noise induced t omato image Stripes and random noises added 16.83 23.61 Proposed WFAF method db4; L=5; k = 1 11.14 27.19 Combined method db4; L=5; k=1 10.19 27.97 Torres method db4; L=3 16.30 23.89 Munch method db4; L=5; s = 1 13.56 25.49 SLM method 78.76 10.20 Table 2 2. SNR estimation of high and low frequency noise for row 960 of the noise induced image. Image Signal SNR (Low freq.) SNR (High freq) Clean tomato i mage 84.91 30.77 38.81 Noise induced i mage 86.28 24.29 24.17 W F AF method 86.22 25.38 28.60 Combined method 86.20 29.12 35.37 Munch method 86.15 23.34 28.43 Torres method 86.35 29.56 33.75 SLM method 95.27 23.31 27.54 Table 2 3 SNR estimation for original and de striped data computed at rows 30 and 830 of the Pond image (band 30). Image Signal SNR (Low freq.) Improved by SNR (High freq.) Improved by Row 30 results : Original 50.99 21.16 21.29 WFAF method 50.99 22.90 1.73 21.85 0.55 Combined 50.99 23.55 2.38 23.70 2.41 Torres method 50.99 22.68 1.51 23.04 1.75 Munch method 50.99 22.79 1.62 21.77 0.47 SLM method 51.14 18.38 2.78 17.87 3.42 Row 830 results : Original 50.98 21.22 19.65 WFAF method 50.98 21.55 0.32 20.88 1.22 Combined 51.02 23.13 1.91 23.69 4.03 Torres method 50.98 21.09 0.13 22.93 3.27 Munch method 50.98 18.79 2.43 20.81 1.16 SLM method 51.14 18.40 2.82 18.12 1.53


66 Table 2 4. SNR estimation in frequency domain for original and de striped data computed at columns 40 an d 438 of the Vegetation image (b and 32 ). Image Signal SNR (Low freq.) Improved by SNR (High freq.) Improved by Row 40 results : Original 49.35 21.81 21.25 WFAF method 49.35 22.42 0.61 21.98 0.72 Combined 49.34 23.01 1.19 23.01 1.75 Torres method 49.35 22.29 0.48 22.95 1.69 Munch method 49.35 20.28 1.52 22.10 0.84 SLM method 49.88 18.13 3.68 18.20 3.05 Row 438 results : Original 48.46 21.28 21.03 WFAF method 48.46 21.32 0.03 21.82 0.79 Combined 48.44 21.96 0.68 24.28 3.25 Torres method 48.46 21.38 0.10 22.68 1.65 Munch method 48.46 21.45 0.17 21.91 0.87 SLM method 48.96 19.46 1.81 18.06 2.96 Table 2 5 SNR estimation for original and de striped images computed at r ow 365 of the Hyperion image (b and 195). Image Signal SNR (Low freq.) Improved by SNR (High freq.) Improved by Original 47.34 24.23 25.42 WFAF method 46.88 25.25 1.02 28.08 2.66 Combined method 46.87 25.59 1.35 33.07 7.65 Torres method 46.87 24.54 0.30 29.74 4.32 Munch method 46.88 24.20 0.03 27.87 2.45 SLM method 46.97 26.50 2.26 28.40 2.98


67 Figure 2 5. Test images. A ) origi nal clean tomato field image, B ) noise induced tomato field imag e (stripes and noises added), C ) pond i mage (hyperspectral Band 30), D ) vegetation i mage (hyperspectral Band 32), E ) Hyperion image (Band 195). Highlighted boxes are zoom in areas used later for visual comparisons. Dashed blue lines mark the location of the profiles drawn across stripes along relatively homogeneous areas for SNR estimation.


68 Figure 2 6. Filtering w ith the proposed WFAF filter. A ) Original clean tomato image. B) Noise induced toma to image before filtering. C J ) F iltered images using different values of filter parameters as labeled.


69 Figure 2 7. PSNR versus k threshold values using diff erent decomposition levels and wavelet types A) db4 wavelet and B ) db8 wavelet, applied to the tomato noise induced image.


70 Figure 2 8. Noise induced tomato image filtered using various methods. A) P roposed WFAF method B ) Combined method C ) Torr es method (L = 3) D ) Torres method (L = 4) E ) Munch method (s = 1) F) Munch method (s = 2). G ) SLM method.


71 Figure 2 9. Filter ing results of the pond image band 30. A) Original image, B) WFAF de striping method, C) Combined method, D) Torres metho d, E ) Munch meth od, and F ) SLM method. Charts underneath the images represent profiles at row 500.


72 Figure 2 10. Filtering re sults of the vegetation image band 32 A) Original image, B) WFAF method, C) Combined method, D) Torres method, E) Munch method, and F ) SLM method


73 Figure 2 11. Filtering results of the Hyperion image band 195. A ) Original image, B) WFAF method, C) Combined method, D) Torres method, E) Munch method, and F ) SLM method.


74 CHAPTER 3 FILTERING HIGH RESOLUTION HYPERSPECTRAL IMAGERY IN MAXIMUM NOISE FRACTION DOMAIN USING WAVELET BASED DE STRIPING 3.1 Introduction Most hyperspectral sensing systems have 2D pushbroom system design. Such systems utilize optical dispersion techniques to disperse the light entering through a narrow sli t into a two dimensional array, where one dimension represents the spatial component and the other dimension represents the spectral component (e.g. Abd Elrahman et al., 2011; Anger et al., 1996; Cocks et al., 1998). The image cube is formed by consecutive ly capturing digital camera frames, where one dimension represents the spatial lines in the formed hyperspectral cube and the other dimension represents different spectral response for each spatial line. The final hyperspectral image is formed by stitching these adjacent lines, captured while the platform (or the sensor) is moving. This image formation technique causes in track striping artifact due to sensitivity differences between individual pixels in the spatial line forming each spectral band. The stri ping patterns formed by such system design are non periodic and are unique for each band. Besides striping, hyperspectral image cubes show non systematic noise in the spatial and spectral domains caused mainly by the physical characteristics of the sensing material (e.g. Charge Coupled Display and Complementary Metal Oxide Semiconductor) and inherent electronic noise. Stripes and non systematic noise degrade image quality and risk its analysis potential unless properly reduced. One of the approaches for no ise reduction is by conducting radiometric calibration. However, calibration coefficients can change over time and frequent calibration may not be possible due to cost and logistics issues. Filtering is another approach for noise reduction that can be cat egorized into:


75 (i) De striping (reduction of stripes) and, (ii) De noising (reduction of non systematic noise). De striping algorithms are usually challenged by the non periodic nature of the stripes and the large number of bands that need to be filtered. Over the years, various methods have been proposed for image filtering in the spatial or transformed domains. In most cases, filtering leads to a loss in signal information and/or introduces artifacts. Linear and non linear low pass filtering (e.g. Gaussian or median filters) were traditionally used for noise reduction in the image domain. Other spatial filters have been introduced to reduce noise and preserve sharpness such as the Sigma filter (Lee, 1983) and the Nagao Matsuyama filter (Nagao and Matsuyama, 1979). All these spatial convolution filters cause artifacts in the image, especially in the case of high spectral resolution images. Spatial methods used to reduce striping noise include directional average filter, simple linear matching, histogram modification and matching algorithms (Horn & Woodham, 1979; Weinreb et al., 1989; Wegener, 1990). Gadallah et al. (2000) utilized a moment matching algorithm to de stripe Landsat satellite imagery by matching each sensor response to a typical re sponse. Histogram matching, moment matching and simple linear matching algorithms were built on the assumption that each detector will view statistically similar sub scenes and require large images (Gadallah et al., 2000). This assumption is not valid for small images or images with non periodic stripes. Filtering was also performed on image transforms such as the Principal Component (PC) transform, Fourier transform, and wavelet decomposition. Frequency domain filtering was successfully applied to de stri pe images by applying the algorithms on the transform of each individual band (Srinivas et al., 1988, Liu and Morgan, 2006). Despite achieving satisfactory results in the case of images with periodic stripes (e.g.


76 Landsat), these methods did not yield effe ctive results in the case of images with non periodic stripes. Recently, wavelet transform has received increasing attention in image compression and noise reduction. Donoho and Johnstone (1994) proposed wavelet based de noising techniques to reduce noise from signal and image through hard and soft thresholding. Wavelet transform has potential to analyze and capture non periodic stripes in its directional and scaled components. Few researchers have applied wavelet transformation in de striping. Torres and Infante (2001) used wavelet transform to analyze and reduce striping effects in images by eliminating directional wavelet detail components in the direction of the stripes. This process is accompanied by a loss in other image information captured in the el iminated wavelet components. Other researchers (Chen et al., 2006; Munch et al., 2009) filtered the wavelet components rather than eliminating them. Munch et al. (2009) proposed filtering wavelet directional components in Fourier domain using a gaussian fi lter. This method minimizes the signal loss but induces smear like distortions along contrasted features in high resolution images. Pande Chhetri and Abd Elrahman (2011) modified this wavelet based method by introducing adaptive filtering to accommodate ar tifacts that could appear in high resolution imagery. The method was successfully tested on individual bands of several hyperspectral images. Maximum Noise Fraction (MNF), which can be defined as Noise adjusted Principle Components (NAPC), has been popula r in dimensionality and noise reduction for high spectral resolution images since its introduction by Green et al. (1988) as transformed components in the order of image quality (Green et. al, 1988, Lee et al., 1990). Various researchers studied refinement s of the MNF process focusing on improving noise


77 estimation (e.g. Chang and Du, 1999 and Liu et al., 2009). Maximum noise fraction filtering can reduce non systematic noise simply by discarding noise dominated components provided that noise covariance stat istics is accurately estimated a priori. However, MNF was rarely tested by itself or in combination with other de striping methods to reduce striping noise. Although filtering based on MNF transformation is promising in reducing non systematic noises from multispectral and hyperspectral images (Lee et al., 1990, Liu et al., 2009), MNF alone is not effective in stripe noise reduction and de striping methods were often applied separately on the original image bands (Gersman et al., 2006 ). This approach may su it multispectral images; however it is inefficient in the case of hyperspectral images where tens or hundreds of bands need filtering. This paper investigates the suitability of applying de striping algorithms on the image MNF transform for high resolutio n hyperspectral imagery and evaluates the results of applying combined de striping and de noising algorithm on the image spectral characteristics. Spatial domain filtering of highly dimensional hyperspectral imagery is time consuming and impractical, espec ially when filtering parameters need to be fine tuned for each band. Other noise content could also confuse the de striping algorithm leading to inferior results. We hypothesize that applying the filter on fewer MNF bands could improve the filtering proces s quality and efficiency. It could also lead to superior results through integration with MNF based de knowledge, applying de striping algorithms on the MNF transformed bands has not been experimented before to de s tripe hyperspectral imagery. As mentioned earlier in this section, spatial filters and frequency domain filters are not appropriate to filter non periodic stripes in high resolution images. Due to the high potential of wavelet based


78 filtering to handle dir ectional non perioidic stripes better, we propose using wavelet based de striping algorithms to be applied on the first few MNF transform bands of the image before transforming them back to image domain. We experimented with two wavelet based algorithms, o ne suggested by Pande Chhetri and Abd Elrahman (2011), and the other proposed by Torres et al. (2001). For comparison, these two de striping algorithms were also applied directly on the original image bands followed by applying MNF based de noising. The f ollowing section (Section 3. 2) describes the methods and algorithms applied in this study including MNF transform and filtering, wavelet based de striping algorithms and the methods of quality assessment. Section 3. 3 introduces the test hyperspectral image s, discusses the filtering procedure, including parameter selection and illustr ates the filtering results. In s ection 3. 4, quality of the applied filtering methods is assessed and co mparatively analyzed. Finally, s ection 3. 5 provides the study conclusion. 3.2 Methodology 3.2.1 Maximum Noise Fraction (MNF) Green et al. (1988) introduced Maximum Noise Fraction (MNF) transformation designed to yield components ordered by image quality or signal to noise ratio, which is commonly employed for dimension and noise reduction MNF uses linear transformation which can be mathematically expressed as: Z i = a i T X i (3 1) where Z i refers to MNF component values, X i refers to original image data, and a i T refers to transform matrix. Unlike Principle Component (PC) transformation (Richards and Jia, 1999), which maximizes variation, MNF maximizes signal to noise ratio. Transform matrix a i T N N


79 are t he covariance matrices of the image and noise respectively. Maximum noise fraction components are then created in the increasing order of noise fraction {Var(N i ) / Var(X i )} or decreasing order of signal to noise ratio. The MNF transform requires noise N ) estimation in addition to image data. Estimating noise covariance accurately is critical to the effectiveness of MNF based filtering methods. Various methods have been proposed for this purpose. One of these methods utilizes dark image to estimate the noise covariance matrix if it has been captured by the same sensor at the time of image capture. Other methods use the original image data or its subset for the estimation. Green et al. (1988) recommended estimating noise covariance from the same image using the minimum/maximum autocorrelation factor procedure proposed by Switzer & Green (1984). Noise covariance can also be estimated by shift differencing a homogeneous subset of the image. This algorithm was utilized in this research to transf orm hyperspectral imagery from the original image domain to MNF domain. The algorithm is based on the spatial autocorrelation principal as differences between adjacent pixels, especially in a homogeneous region of the image, is most likely due to image noi se. This principal was used to compute the covariance matrix for the error using pixel values in a homogeneous region and their counter parts obtained by shifting one pixel in the horizontal and vertical direction. When the MNF transform is performed, the first several MNF bands are signal dominant while the amount of noise increases gradually and eventually the last bands become dominated by noise only. Filtering using MNF transform involves transforming image data into MNF components, spatially filtering and/or simply setting each of the


80 noisy components to a constant value, and finally transforming these bands back to the original image spectral domain (Lee et al., 1990). The nature of the applied filter and the way noise dominated bands are treated gove rn the filtering performance. This technique can filter out non systematic noise successfully; however, it fails to suppress stripe noise. 3.2.2 Filtering The ability to de stripe the hyperspectral image by de striping few MNF bands of a hyperspectral cube prov ides significant computational advantage and increases the effectiveness of the applied filter. Widely used destriping methods including histogram matching methods and frequency domain filtering are useful in filtering low resolution images or images with periodic stripes such as satellite images. Wavelet transform decomposes image into directional and scaled components and has potential to handle non periodic vertical or horizontal stripes. Two wavelet based de striping methods are used in this research an d applied to original image bands as well as MNF bands. The first method proposed by Torres et al., 2001 (named Torres method in this research) is based on eliminating the directional components in the striping direction for few wavelet levels. Directional components, for instance vertical components for vertical stripes, are usually replaced by corresponding component means (see Figure 3 1). This method normally removes the striping artifacts in addition to other image information contained in the eliminat ed components. Implementation of Torres algorithm requires deciding on the wavelet type and the number of wavelet components to filter. The second filtering method used in this study is Wavelet Frequency Adaptive Filter (WFAF) method proposed by Pande Chhetri and Abd Elrahman (2011) as a modification of the method proposed by Munch et al. (2009). The WFAF method filters


81 stripes from direct ional wavelet component adaptively in the Fourier transform domain by identifying non stripe features in the wavelet components and minimizing their effect. The method reduces the stripes from high resolution images and retains image signal with minimal ar tifacts. The WFAF method is applied to individual image or MNF bands tainted with stripes. Each band is decomposed into wavelet components in a number of scales, which separates the stripes into the wavelet component at the corresponding direction at each level ( Figure 3 1). The wavelet directional components containing stripes are filtered with an adaptive filter in the Fourier domain by normalizing zero frequency values or Direct Current (DC) values adaptively considering non stripe image content (Pande C hhetri and Abd Elrahman, 2011). This content (non stripe features) in the directional wavelet components is detected in each array (column or row) along the striping direction as those beyond a statistical threshold from the array neighborhood mean. A new vector Y j free from these extreme pixels is created by excluding influential values from the original wavelet vector so that (3 2) where x i j is the value for pixel i in array j, j is the mean of the digital numbers of Chhetri and Abd Elrahman, 2011). Normalize d DC value Fjnorm for each array is then calculated as F j norm = F j orig j j j (3 3) where F j orig j j are array means before and after excluding extreme non stripe pixels. Inverse Fourier transform is applied after replacing


82 the DC values with the normalized ones. Application of inverse wavelet transform using the filtered wavelet components results in th e de striped band. Implementation of the WFAF algorithm required determining the wavelet types, number of wavelet levels and the k factor (adaptive threshold). Figure 3 1. Schematic diagram illustrating Wavelet based Torres and WFAF de striping meth od s In this study, the filtering methods were named Direct Torres and Direct WFAF in case of applying the Torres and the WFAF de striping algorithms on all image bands before implementing the MNF transform. When applying these de striping algorithms on th MNF Torres and MNF WFAF, respectively ( Figure 3 2). The number of filtered bands was determined through visual inspection of the transformed bands and their corresponding MNF eigen values. The WFAF method was implemented through custom Ma tlab ( ) code, which can be obtained through direct contact with the authors. The Torres meth od is implemented based on the method described by Torres et al. (2001). The MNF transform was


83 implemented through the ENVI software ( US /ProductsServices/ENVI.aspx ). Figure 3 2 Schematic diagram of MNF transform combined with wavelet based de striping methods. 3.2.3 Quality Assessment Quality assessment is vital to validate the quality of the applied filtering methods. Ideally, successful fi ltering should be able to reduce all stripes and noises and should retain the local signal levels without smoothing them or inducing distortions. In this study, both visual and quantitative methods were used to compare different filtering methods. Visual c omparison between pre and post filtering images can indicate the efficiency of the filtering method. For hyperspectral images, not only individual bands can be assessed, but also spectral profiles of individual pixels can be evaluated. Quantitative assess ment using geostatistical indices and frequency domain signal to applies to signal information in the image but not to noise content. Geostatistical indices


84 were used to indica te the amount of systematic (striping) and non systematic noise in (3 4) where C(h) is the covariance between the data sets separated by a lag distance h and can be expressed as: (3 5) h ij is the lag distance, which is the distance b etween the pairs of data points located at i and j. h +h ar e means of the data points separated by lag h. h +h are standard deviations of these data points, and N(h) is the number of data point pairs separated by h. Autocorrelation values for small lag value, as a measure of similarity between adjacent pix els, are indirectly proportional to noise and stripes. For an image with vertical stripes, the autocorrelation values with small horizontal lag inversely correlate to striping and non systematic noise. After noise and stripes are suppressed, these values a re expected to increase. On the other hand, autocorrelation values with small vertical lag are related to non systematic noise only. Ratios of autocorrelation along two directions (horizontal to vertical) computed for original image bands with vertical str ipes will be less than one; but after de striping those images, values are expected to increase approaching a value of one. Signal to noise ratio (SNR) is a commonly used index for quantitative assessment of image quality. In the case of non synthetic (actual) imagery, the amount of noise is


85 unknown and the signal to noise ratio needs to be computed indirectly. Cross section pr ofiles across relatively homogeneous area are converted into frequency amplitudes (using Fourier transform) and signal to noise ratios are computed in Decibel (db) (e.g. Torres et al., 2001). In our study, we used the SNR computed as shown in E quation 3 6 for quality assessment. SNR (in db) = 10 log 10 (A signal /A noise ) = 10 log 10 (A signal ) 10 log 10 (A noise ) ( 3 6) where A signal is signal amplitude considered to be central component of frequency vector, and A noise is noise amplitude considered as noise pea ks. In the Fourier domain, high frequency noise peaks are related to high frequency stripes and noise, while low frequency noise peaks are related to induced distortions and low frequency stripes. Figure 3 3 illustrates an example of an image profile and corresponding Fourier transform showing signal and noise peaks. Maximum noise Figure 3 3 Image profile and fourier values A ) Horizontal profile of original pond image Band 20 at row 25 B ) Amplitude values (dB) in f ourier domain.


86 peak in both high frequ ency and low frequency ranges are considered to compute SNR high (Signal to high noise ratio) and SNR low (Si gnal to low noise ratio) using E quation 3 6. These values are computed for homogeneous row profile in original and filtered images. 3.3 Experiments and Results 3.3.1 Hyperspectral Imagery Images used in this study were captured by a hyperspectral image acquisition system developed at UF/Geomatics lab (Abd Elrahman et al., 2011) which is mainly comprised of an ImSpector imaging spectrograph made by Spectral Imag ing Ltd. ( and an Imperx IPX 2M30 Lynx Charge Coupled Device (CCD) camera made by Imperx Inc. ( Like most current commercial hyperspectral sensors, light passing through a narrow (30 microns) spectrograph sl it of this system is dispersed optically to different wavelengths and received by the CCD camera attached to the back of the spectrograph. After dispersion, a two dimensional image is formed in the camera, with one dimension representing the spatial compon ent and the other dimension representing the spectral component. Each captured image creates a spatial line in the formed hyperspectral image. Adjacent lines are captured, either by rotating a mirror mounted in front of the sensor, or moving the sensor sys tem with a moving platform (e.g. airplane, or ground vehicle for terrestrial close range applications). Each line in the formed hyperspectral cube consists of 800 spatial pixels. With its binning property, the Imperx camera can produce images with differen t numbers of spectral bands. In our case, we utilized a binning factor of four, which produces 270 spectral bands (only 204 bands are usable), covering the Visible and Near Infrared (ViNIR) part of the spectrum (400nm 1000nm).


87 Testing the filtering method s was implemented on two image sets ( Figure s 3 4 and 3 5). The first image set is a vegetation scene with calibration targets captured for a tomato experimental field to analyze growth and monitor plant health. The second image is for an aquaculture pond w ith floating reflectance targets captured for the purpose of analyzing bio optical properties of water constituents. A B Figure 3 4 Test images. A ) Original vegetation image b and 15 0. B ) Original pond image b and 150. Red boxes represent zoom in portions used later for visual comparison. Yellow lines represents homogenous profile along row 820 in vegetation image and along row 25 in pond image and are used for SNR estimation. 3.3.2 Image Filtering We used the images described above to examine two d ifferent techniques to de stripe and de noise high resolution hyperspectral imagery using two different de striping algorithms (WFAF and Torres filtering). In t he first technique (t wo methods for this technique corresponding to the Torres and WFAF algorith ms and namely: Direct Torres and Direct WFAF methods), each band in the images was de striped. This step was


88 A B Figure 3 5 Zoom in portion of test image s (red box areas of Figure 3 4) b ands 20, 150 & 204 A ) Original vegetation scene images B ) Original pond images. followed by applying an MNF transform on the de striped image and examining the quality of the MNF bands. As indicated earlier, band quality degrades increasingly from low order MNF bands to higher order on es. Our images showed that depending on the image and algorithm used, after ten to fifteen bands of the MNF transformed image, almost no signal information persists and bands become dominated by noise. In order to capture most of the signal content, we dec ided to accept the first ten to fifteen MNF bands and ignore the remaining bands as noise only. Eigen values associated with the MNF bands suggested a similar pattern. Figure 3 6 A shows the percent eigen values obtained for the vegetation image with Direct WFAF method. The values rapidly decreased for the first few bands then leveled off afterwards. This pattern indicates that


89 each band explains fewer variations in the image data (less signal to noise ratio) than previous bands. Finally, these accepted MNF bands are transformed back to the original image domain using inverse MNF transform. In the second technique (MNF Torres and MNF WFAF methods), the test image sets were first transformed into the MNF domain. Unlike the Direct Torres and Direct WFAF method s, de striping algorithm was applied on the selected MNF bands rather than on the image bands. In this experiment, the first ten MNF transformed bands of only vegetation scene images are shown in Figure 3 7 A due to space limitation. The images show that th e first MNF band contains minimal stripe or non systematic noise. Striping effect and noise gradually increase with the increase of the band number of the MNF transformed imagery. Beyond band ten, images were heavily dominated with stripes and noise. Eigen values of the MNF bands for the original vegetation image are shown in Figure 3 6 B which also suggest that most of the image variation is contained in the first few MNF bands then levels off after around Band 10. Only the first ten bands were de striped using the WFAF and Torres algorithms (for the MNF WFAF and MNF Torres methods). The results of de striping the MNF bands de striped using the WFAF and Torres algorithms are shown in Figure s 3 7 B and 3 7 C The de striped MNF bands for each of the tested im ages were then subjected to MNF backward transform to original spectral domain to yield the de striped and de noised image sets. Examples of one band of the vegetation image before and after filtering with the MNF WFAF technique and corresponding horizonta l and spectral profiles (of a vegetation pixel) are shown in Figure 3 8.


90 Final de striped and de noised images resulting from above mentioned two techniques (four methods) are presented in Figure s 3 9 and 3 10. Figure s 3 9 A to 3 9 D are resulting vegetati on images filtered using Direct WFAF, Direct Torres, MNF WFAF and MNF Torres methods, respectively. Figure s 3 10 A to 3 10 D are final pond images filtered with the same four methods. 3.3.3 Parameter Selection Because the stripe intensity varies from one band to another, different parameter values were used for various bands in the image and MNF transform domain. Torres de striping algorithm requires two parameters wavelet type and number of decomposition levels. The WFAF de striping algorithm requires the adapt ive threshold value (k) in addition to wavelet type and decomposition levels. Our experiments showed that a threshold value (k) between 1 and 1.5 would give adequate results for most image bands and cause minimal artifacts (Pande Chhetri and Abd Elrahman, 2011). Combinations of different types of wavelets from the Daubechies wavelet family (Daubechies, 1988) and a number of decomposition levels were tested. Visual differences in the filtering quality but the number of decomposition levels can significantly affect the filtering results. In our experiments, the db4 wavelet (one of the Daubechies wavelet family) was chosen for all tests. Different numbers of decomposition levels (L) were examined for each band to select the number of decomposition levels producing best filtering results. The number of decomposition levels was chosen high enough to eliminate the stripes but not too high that it would introduce artifacts or cause over smoothing. When implementing the algorithms directly on the original image bands (first technique), we found that two to eight decomposition


91 levels were sufficient to achieve adequate de striping results for WFAF filtering while maximum decomposition level s were limited to six for Torres de striping because of resulting blurry images at higher decomposition levels. When the de striping algorithm was applied on each band in the image domain directly, a higher number of decomposition levels was needed for the first and last group of bands. In other words, the number of decomposition levels needed to de stripe the image was less when approaching the middle group of bands. This can be attributed to the optical characteristics of the hyperspectral sensors gener ally, including the one used to capture the images tested in this study (Abd Elrahman et al., 2011). When de striping the MNF transformed images (second technique), the de striping algorithm was applied on the first ten bands only. Two or three wavelet decompositions were sufficient to de stripe the first three MNF bands and a maximum of eight (or six fo r Torres method) wavelet decomposition levels were needed to de stripe the last three bands (MNF bands 8, 9 and 10). A B Figure 3 6 MNF transform percent eigen value graph s (in log scale) derived from A ) d e striped vegetation im age (Direct W FAF method), and B ) the original vegetation image.


92 3.4 Assessment and Discussion 3.4.1 Visual Assessment Visual assessment of filtering quality was conducted to compare the results of the two filtering techniques, applying the de striping algorithms on the image ban ds directly (Direct WFAF and Direct Torres) versus applying them on some bands of the image MNF transform (MNF WFAF and MNF Torres). The assessment shows that applying the de striping algorithms on the MNF bands (MNF WFAF and MNF Torres) yielded better res ults than applying the algorithms directly on the original bands followed by MNF de noising (Direct WFAF and Direct Torres) ( Figure s 3 9 and 3 10). Visual inspection indicated a better filtering quality associated with de striping the MNF bands. Among de striping algorithms applied on original image bands, the Direct WFAF method was much better than Direct Torres. Results from Direct Torres show blurry images due to over smoothing and yet insufficient stripe reduction in many bands ( Figure s 3 9 B and 3 10 B ) Images filtered with Direct WFAF show good results but some smear like vertical distortion lines were visible in some bands of the processed images ( Figure s 3 9 A and 3 10 A ). Application of the de striping algorithms on the MNF bands yielded higher qualit y images with minimum distortions. When the images were transformed to MNF bands before de striping, the first few bands with higher eigen values had less striping ( Figure 3 7), which suggests effective filtering and less potential for artifact introductio n. Filtering the MNF bands by the WFAF method greatly reduced stripes and retained most signal information with minimal distortion induced ( Figure 3 7 B ). For the Torres algorithm, increasing the number of filtered wavelet levels resulted in a more blurry image. The MNF Torres method over smoothed the MNF bands,


93 especially the higher ones that require higher wavelet levels, making them blurry ( Figure 3 7 C ). Final images filtered with MNF Torres, however, were found to be of better quality because the first few MNF bands with higher Eigen values, which account for more than 99% of variance in the image data ( Figure 3 6 B ), had minimum amount of stripes and could be de striped using minimum (i.e. two) wavelet decomposition levels. Images filtered with MNF WFAF had the best quality with reduced stripes/noise and preserved details and crispness ( Figure s 3 9 C and 3 10 C ). No induced distortions were visible in most of the image bands except for faint smearing effects in a few end bands. In the case of the vegetatio n image set, results of MNF Torres filtering ( Figure 3 9 D ) were slightly blurry but still comparable to results of the MNF WFAF filtering ( Figure 3 9 C ). For the pond image set, MNF WFAF method yielded results ( Figure 3 10 C ) that are clearly better than MNF Torres method ( Figure 3 10 D ) in visual assessment. Overall, the methods of de striping MNF bands (MNF WFAF and MNF Torres) improved image quality by reducing stripes and noise across bands of various noise intensities in the tested images. The MNF WFA F method, especially, maintained the image details and crispness with minimal distortions. Even the heavily contaminated bands (e.g. last band 204) were outstandingly clear and sharp with strong signal information in the filtered images. Comparison of hor izontal profiles before and after filtering ( Figure 3 8) showed image profiles (both horizontal and spectral) that are smoother than the original by eliminating most noise patterns and yet followed local mean brightness and sharp signal changes.


94 3.4.2 Quantitat ive Assessment Analyzing a utocorrelation i ndices and r atios For assessing the quality of fil tered images, autocorrelation (E quation 3 4 ) and frequency domain signal to noise ratio ( E quation 3 6 ) were computed for original and filtered image sets and compar atively discussed. Assessment analysis included images filtered by applying the de striping algorithms on original image directly (Direct WFAF and Direct Torres), on MNF transform (MNF WFAF and MNF Torres) as well as image de noising only applied to the im age MNF transform (eliminating high order MNF bands). Autocorrelation values with a horizontal lag of one pixel and autocorrelation horizontal to vertical ratios were computed for original and filtered image sets for all bands. Figure s 3 11 A and 3 11 B show these indices for the vegetation scene image bands and Figure s 3 11 C and 3 11 D for the pond image bands. The first set of bands (e.g. first 20 bands) and last bands (e.g. higher than band 125) of the original image were less auto correlated indicatin g heavy stripes and noise content. In those bands, autocorrelation in horizontal direction is less than in vertical direction yielding lower autocorrelation horizontal to vertical (H/V) ratios. After filtering, autocorrelation horizontal values in these ba nds increased significantly and autocorrelation ratios approached one. It should be noted here that de noising only by eliminating all bands beyond Band 10 in the image MNF transform (followed by inverse MNF transform to original image domain) improved the se values only partially. Autocorrelation horizontal values and H/V ratios obtained from other filtering methods were higher and comparable for middle bands (Band 15 to Band 180) for both images but the values differed for bands at both ends, which are co ntaminated heavily with stripes and noise. In the vegetation test image, the autocorrelation values and


95 autocorrelation ratios obtained from the filtered images were comparable for all bands. However, MNF Torres yielded autocorrelation ratios more than one for the bands at both ends. In the case of the pond image, autocorrelation values of these bands obtained from MNF WFAF and Direct WFAF filtering were higher than original image but lower than those obtained from MNF Torres and Direct Torres filtering. G enerally, the WFAF algorithms improved autocorrelation ratios close to but less than one while Torres algorithms yielded autocorrelation ratios higher than one. Autocorrelation ratios that are more than one indicate that horizontal direction is made smooth er than vertical direction. This supports the directional over smoothing effect of the Torres algorithms, as suggested also by the visual inspection results, which is attributed to the elimination of whole wavelet directional components. Analyzing s ignal t o n oise ratios. Signal to noise ratios of original and filtered images corresponding to low and high frequency noise peaks, as computed using E quation 3 6 were plotted against corresponding band numbers ( Figure 3 12). Signal to noise ratio graphs show si gnificant improvement in image quality using the MNF WFAF and MNF Torres methods across all bands in respect to both low frequency and high frequency noise. Results from direct de striping methods (Direct WFAF and Direct Torres) also showed improved signal to noise ratios but inconsistent values across adjacent bands. As shown before, MNF based de noising only (no de striping) improved the SNR values only partially when compared to other methods because the filtered bands still contain stripes. In the veget ation test image, Signal to low frequency noise ratios (SNR low ) resulting from the MNF Torres method were comparable to those obtained using the MNF WFAF method ( Figure 3 12 A ). In the case of the pond image, the SNR low resulting


96 from the MNF WFAF method we re higher than those from the MNF Torres and other filtering methods for the bands heavily contaminated with stripes and noise (some initial bands and bands beyond Band125) but were lower for middle bands where noise and stripes are minimal ( Figure 3 12 C ). MNF Torres method yielded signal to high frequency noise ratio (SNR high ) higher than MNF WFAF and other methods for both test images ( Figure s 3 12 B to 3 12 D ). Visual assessment showed that MNF Torres filter over smoothed the images ( Figure s 3 9 D and 3 10 D ), that is probably the reason high frequency noise values are lower for this method. 3.4.3 Discussion Overall, both visual and quantitative analysis showed that de striping MNF bands yields much better results than de striping the original image bands directly followed by MNF de noising for the WFAF and Torres methods. The two MNF based methods (MNF WFAF and MNF Torres) performed better and faster. These methods have significant advantage and convenience in determining and fine tuning filtering parameters for an d de stripe only a few (ten in our experiment) bands in the MNF domain rather than determining these parameters for each of the over 200 image bands. This means that only a fraction of the experimentation and filtering time/computational resources are requ ired to filter hyperspectral image. MNF Torres filtering method improved the image quality significantly and yielded much better results than direct de striping methods but tended to eliminate image details in addition to noise content, which resulted in b lurriness along feature edges in some bands. Even though quantitative assessment (autocorrelation and SNR high ) were better in some cases for the MNF Torres method, visual assessment and autocorrelation ratios showed that the


97 MNF WFAF method retained signal information better than MNF Torres method and yielded higher quality filtered images. This study investigated integration of MNF transform with wavelet based de striping methods. For future research, 3D wavelet analysis might be investigated in developin g method of filtering hyperspectral image to filter both spatial and spectral dimensions. Besides wavelet based analysis, there are fractural function methods, most notably Detrended Fluctuation Analysis (DFA) that is applicable in scaling analysis of mult ifractals in higher dimensions and has been successfully applied to DNA sequencing, heart rate dynamics, air pollution and image pattern recognition (Peng et al., 1995, Varotsos et al., 2005). Such methods have potential to be used in image filtering and m ight be considered in future study. 3.5 Conclusion This study compares de noising/de striping algorithms/methods applied to high resolution hyperspectral imagery. Two wavelet based image de striping algorithms were compared when applied directly to individual bands in the image domain (Direct Torres and Direct WFAF) or to the selected bands in the image MNF transform domain (MNF Torres and MNF WFAF). When applying the de striping algorithms on MNF bands (MNF Torres and MNF WFAF), the results were superior and the process was more efficient in parameter selection and computation due to the reduced number of bands to be de striped. However, the Torres algorithm, which involves eliminating wavelet components in the striping direction, resulted in noticeable over s moothing effect. The MNF WFAF filtering method effectively de striped image bands with different amount of stripes and retained signal information without introducing significant distortions.


98 A B C Figure 3 7. First 10 MNF b ands of vegetation scene images A) Original MNF bands. B ) After de striping them with WFAF alg orithm (MNF WFAF method). C ) After de striping them with Torres algorithm (MNF Torres method).


99 Figure 3 8 Filtering results of vegetation scene image cube. A) Ori ginal image band 150. B ) Filte red image band 150 using MNF WFAF method. T heir corresponding horizontal profiles (original in bl ue and filtered in red) along C) Row 450 and D) Row 725. S pectral profiles of vegetation pixel (as indicate d by small red square box) of E) Original image, and F ) Filtered image.

PAGE 100

100 A B C D Figure 3 9 Filtered Vegetation image s (b ands 20, 150, 204) using A) Direct WFAF method, B) Direct Torres method, C) MNF WFAF, and D ) MNF Torres filter. (Refer to Figure 3 5 A for original vegetation images.)

PAGE 101

101 A B C D Figure 3 10 Filtered p ond images (b ands 20, 150, 204) using A) Direct WFAF method, B) Direct Torres method, C) MNF WFAF, and D ) MNF Torres filter. (Refer to Figure 3 5 B for original pond images.)

PAGE 102

102 A B

PAGE 103

103 C D Figure 3 11 Geostatistical indices (a utocorrelation wit h horizontal one pixel lag and horizontal to vertical a utocorrelation ratio) for origi nal and de striped images. A) a utocorrelat ion for vegetation images, B) a utocorrelation ra tio for vegetation images, C) a utocorrelation for pond im ages, and D) a utocorrelation ratio for Pond images.

PAGE 104

104 A B

PAGE 105

105 C D Figure 3 12. Signal to n oise ratios SNR low (for low noise peak) and SNR high (for high noise peak) computed along homogenous profiles (see Figure 3 4) f or original and filtered images. A ) SNR low for veg etation images along row 820, B ) SNR high for veg etation images along row 820, C ) SNR low for pond images along row 25, and D ) SNR high for pond images along row 25.

PAGE 106

106 CHAPTER 4 ANALYSIS OF HYPERSPECTRAL IMAGERY TO ESTIMATE WATER QUALITY PARAMETERS IN CASE 2 WATERS OF AQUACULTURAL PONDS 4.1 Introduction Many studies have been developed to model the optical properties of natural water in terms of water constituents. Water is broadly classified as (1) Case 1 water when optical properties of water are dominated by Chlorophyll a (Chl a) concentration such as in open ocean waters and (2) Case 2 water when other water constituents (e.g. dissolved organic matter and inorganic mineral particles) have strong influence on the optical properties of water such as in turbid inland lakes (Lee and Hu, 2006; Morel and Prieur, 1977). Case 2 waters represented by turbid, nea r shore, and m ost hyperspectral sensing systems have 2D pushbroom system design. Such systems utilize optical dispersion techniques to disperse the light entering through an estuarine waters are of great ecological and economic importance. Water quality pa rameters such as Chl a, total Nitrogen (N), and total Phosphorus (P) are important parameters widely used to determine human impacts on water quality. Chl a is an indicator of algal biomass that depends upon N and P availability in the water bodies (Carlso n and Simpson, 1996). The interrelationships between Chl a and N and P have been explored during the past few decades (e.g. Brown et al., 2000; Kamarainen et al., 2009). These relationships suggest the potential use of Chl a to estimate N and P concentrati ons. Aquaculture ponds are a unique and economically important subset of Case 2 waters. Farm gate values of aquaculture in the United States exceed $1 billion, with nearly $700 million originating from the southeastern United States (SARC, 2008). In aquacu lture systems, there is an expectation that nutrient and chlorophyll levels will fluctuate more greatly than in large lakes and ponds where human manipulation occurs on a more limited

PAGE 107

107 scale. Optimal conditions for breeding and raising particular fish speci es are often closely guarded secrets in aquaculture operations. There may also be strong production incentives to maintain consistent water quality parameters over time. Estimating Chl a of Case 2 waters using remote sensing techniques is challenging beca use optical properties of water are significantly influenced by mineral and sediment particles, in addition to phytoplankton and associated organisms (Meaden and Kapetsky, 1991; IOCCG Report #8, 2009). Aquaculture ponds can be more challenging due to their shallow depth, high turnover (harvesting rate), and regular manipulation. Three methods commonly used to estimate Chl a are in situ sampling, mathematical modeling, and remote sensing (Dekker et al., 1996). In situ sampling is currently one of the most co mmon methods for gathering data to assess the health of small water bodies (e.g. aquaculture ponds) (Rodgers, 2008; Tabinda and Ayub, 2010). Collecting and analyzing water samples is a laborious and relatively expensive process. On the other hand, mathemat ical modeling of water quality parameters requires significant input and model assumptions that make estimation reliable only at regional or smaller scales (Dekker et al., 1996). Case 2 waters contain phytoplankton intermixed with Colored Dissolved Organic Matter (CDOM) and suspended inorganic material, as well as yellow substances (gelbstoffe) that affect the optical properties of water and pose a measurement challenge for remote sensing beyond that presented by open ocean waters (i.e., Case 1 conditions) (Morel and Prieur, 1977; Morel, 1988). Isolating spectral contributions of multiple water components is a significant technical challenge, but extensive development of algorithms for Chl a estimation have been improvised (Schalles, 2006).

PAGE 108

108 Chl a estimation using spectral analysis of satellite imagery is becoming standard, especially for Case 1 waters in large water bodies and open ocean, where dominance of a phytoplankton signal is the defining water spectral characteristic (Morel and Prieur, 1977; IOCCG Re port #3, 2000). Estimating Chl a from satellite and airborne imagery has gained momentum in the past decade. For example, Gitelson et al. (2008) used a combination of two or three bands from the Medium Resolution Imaging Spectrometer (MERIS) and the Modera te Resolution Imaging Spectrometer (MODIS) instruments to assess Chl a concentration consistently in turbid waters. Thiemann and Kaufmann (2000) used a two band (705 and 678 nm) approach with an airborne hyperspectral (HS) sensor to estimate chlorophyll le vels in German lakes. Generally, satellite systems lack the spatial resolution to examine small scale Case 2 waters such as aquaculture ponds, while airborne systems are generally too costly to use in day to day monitoring of a small scale operation. The use of remote sensing techniques for analyzing Chl a is subject to effects of upwelling reflectance coming from the water surface, bottom substrata impact in shallow waters, and variations in water turbidity (Bhattia et al., 2009; Mobley, 1999; Rundquist e t al., 1995). Several techniques have been used to overcome these problems. Gitelson et al. (2008) used in situ spectral data collected using spectrometer observations taken just beneath the water surface to avoid surface water reflectance. In satellite re mote sensing, reflectance values in the infrared bands were used to estimate water surface reflectance under the assumption that deep water would absorb all downwelling energy (Gitelson et al., 2007; Ruddick et al., 2001). This assumption is violated for C ase 2 waters, where suspended material can contribute to reflectance signal in these band

PAGE 109

109 values. Bhattia et al. (2009) suggested that there is currently no method available for removing the complex spectral perturbations due to surface water reflections. Some research utilized statistical modeling of water surface movement (waves) to estimate its effect on the upwelling radiance received by the sensor (Doxaran et al., 2004). The same concept also applies to reflectance from the bottom strata, where composi tion and water depth affect the upwelling reflectance, adding uncertainty to estimated Chl a concentration. In this study, we developed a technique to utilize a ground based hyperspectral ( HS ) imaging sensor to e stimate Chl a for shallow eutro phic / hyper eutrophic Case 2 waters in aquaculture ponds. The sensor can be mounted on a boat and synchronized with navigation sensors to provide georeferenced HS imagery or can be mounted on a crane for near shore imaging. Our HS remote sensing system enabled observa tions at a scale appropriate for water quality analysis of aquaculture ponds. Due to their small surface area and close proximity, neither satellite nor aerial capture options seemed cost effective or capable platforms for this particular application. Unli ke in situ spectral based Chl a measurement systems that utilize spectrometer measurements, our imaging system provides numerous pixels (samples) for the water body and can be utilized onboard a moving platform. We tested the use of floating calibration ta rgets to calibrate for atmospheric conditions and submerged targets to increase signal strength. In our experiments, we set out to answer the following basic questions: (1) Can we model Chl a (and consequently N and P) concentration(s) in shallow aquacultu re ponds with high water quality variability using ground based HS imaging? and (2) To what degree does the use of reflectance targets improve the ability to produce accurate Chl a

PAGE 110

110 estimates in shallow Case 2 waters, where multiple factors may exert undesi rable influence on the upwelling signal? 4.2 Background 4.2.1 Bio optical M odels Remote sensing reflectance R is widely used in the bio optical analysis of a water body. Adopted from Mobley (1999) and ignoring the sun and the sensor relative geometry, R can be derived mathematically as follows: w d ( 4 1) w is water leaving spectral radiance from light beneath the surface (excluding radiance from water surface reflection), and E d is downwelling irradiance incident to the water surface. For a given wavelength, the remote sensing reflectance R can be expressed as: R = (L t r )/L g g s )/L g g ) ( 4 2) where L t is the total radiance measured by the sensor poin ting downward, L r is the radiance reflected from water surface, L s is the downwelling sky irradiance, L g is the calibration target upwelling radiance, and R g is the (constant) reflectance coefficient of the sun sensor orientation, wavelength, wind speed, sensor field of view, and sky radiance distribution. Assuming L over the detector field of view (Mobley, 1999). Many studies L for specific direction. For a viewing angle near nadir, L 0.02 is adopted. Melack and Gastil (1994) indicated that the difference in R t value is less than 2%.

PAGE 111

111 In shallow waters, determination of R depends on knowledge of the water properties, water depth, and may be bottom reflectance properties. In this study, we used ratios between reflectance (two band and three band) values as descr ibed in the next sub section. This technique enabled further elimination of the constants R g We postulated that approximate radiometric calibration of the sensor may generate acceptable accuracy, and that L t and L g values in E quation 4 2 can be repl aced by the sensor digital numbers of the water and calibration target. Mobley (1999) suggested the use of band values in deep water regions of the image to estimate water surface s ); this approach was implemented in a SeaWiFS satellite imag e analysis protocol (Mueller et al. 1995). A similar method was used in this study, as will be discussed in sub section 4.3.4 4.2.2 Modeling C hlorophyll a C oncentration in Case 2 W aters Modeling Chl a concentration using optical spectrometry relies on analyzi ng water upwelling radiance to build relationships between sensed values and water constituents. Case 2 waters represent a challenge due to the presence of CDOM and tripton, factors that cause great problems for algorithms used to estimate Chl a in Case 1 ocean waters ( IOCCG 2000; Xu et al., 2009; Zimba and Gitelson, 2006). Summarizing a previous study, Schalles (2006) asserted that using a simple ratio of the reflectance signals from a near infrared (NIR) band near 700 nm and the red absorption band near 670 nm isolated the Chl a signal from the influence of other pigment and CDOM signals, thus, making it a candidate for two band analysis. Gons (1999) provided theoretical explanations to this argument using a bio optical model. Multi band methods and their associated algorit hms have continued to proliferate. A three band index has been proposed and used for determining Chl a concentration. The index was proposed

PAGE 112

112 mainly to overcome violations of the two band model assumptions of constant Chl a absorption coefficient and flores cence quantum yield, which depend on the dynamics of al., 2009). Gitelson et al. (2008) used MERIS wavelength ranges of 660 670, 703 713, and 750 758 nm in a three band model to estimate Chl a in lakes and reservoirs in Iowa and Nebraska. We used two band (simple reflectance ratio) R band indices [R 1 2 )] 3 ) to model chlorophyll a (and hence N and P) concentrations. It should be m entioned here that although we used reflection ratio and three and Gitelson, 2005; Zimba and Gitelson, 2006) in this field, we conducted a preliminarily analysis using multiple band s and band combinations. For example, we tested the use of Partial Least Squares (PLS) using all observed spectral bands. PLS is designed for highly correlated/dimensional covariates (e.g. hyperspectral bands). It reduces the number of covariates to a set of uncorrelated components using least squares principal and then applies the regression. We also experimented with band ratios of different bands representing peaks and troughs in eutrophic water spectrum, where several bands in each peak and trough regio n were tested. Among all tested models, the coefficients of determination and validation results obtained using the reflection ratio of the chlorophyll florescence band (700 nm) and the chlorophyll absorption band in the red region (680 nm) were superior. A three band model involving these two bands and a further infrared band provided similar results. In our study, the use of simple reflectance ratio and three band models was sufficient to demonstrate the efficiency of our

PAGE 113

113 suggested technique (including th e use of submerged targets) in modeling Chl a, N; and P concentrations in eutrophic or hypereutrophic waters. Zimba and Gitelson (2006) tested the three band and two band (simple ratio) models on 0.9 m depth aquaculture ponds with Chl a concentrations that exceeded 3000 g/L in at least one sample. They utilized a dual fiber spectrometer where one of the fibers collected upwelling water radiance just underneath the water surface and the other fiber collected downwelling sky irradiance. The two band and thre e band models were able to explain R 2 value 0.78 and 0.71 of the data, respectively, which highlighted the difficulties of estimating Chl a content in hypereutrophic aquaculture ponds, especially with spatially varying algal blooms. Rundquist et al. (1995) investigated the changes in spectral reflectance of algal concentrations using bright and dark bottoms at various depths in a 9500 L pool and manually controlled the algal concentration in the pool. In this research, we utilized reflectance targets at two depths to investigate the effect of this technique on Chl a estimation, which to our knowledge has neither been investigated using HS systems nor applied to eutrophic and hypereutrophic waters such as those of aquaculture ponds. Our use of HS imaging sens or, instead of commonly used spectrometers in in situ reflectance studies, enabled capturing images of the water with different floating and submerged calibration targets. We hypothesized that using highly reflective targets could improve the signal to noi se ratio and minimize the effect of turbidity on the observed spectra. 4.3 Materials and methods 4.3.1 Site D escription We collected HS images and water samples from 14 aquaculture ponds at the University of Florida, Tropical Aquaculture Laboratory in Ruskin, Florid a, USA on seven

PAGE 114

114 dates between August 20, 2009 and January 12, 2010. Prevailing temperature conditions varied considerably among data collection dates, with average daily temperature ranging from 2.8 to 28 C. Some of the ponds were surveyed more than one t ime, and on different dates, for a total of 30 observations. Ponds were approximately 0.2 ha in size and 2.5 m in depth and were selected because of their consistent size, access, and variable water quality conditions. Most ponds are maintained on medium ( 6 months to 1 year) to long (>1 year) production cycles. Some are maintained for short cycles of less than 6 months depending on the ornamental fish production cycle. All ponds are created by water seepage through ground water table with little input of we ll water except during the late spring near the end of the dry season, a time which was outside our observation period. Ponds are de watered and washed to begin a production cycle. Water is pumped from the pond into a drainage system (drainage ditches and retention/detention ponds). Soft, organic sediments are removed by washing and pumping. Hydrated lime is applied as a disinfectant before seepage refills the pond. Commercial fish feed is fed to most species, which contributes to algal growth. An aeration system consisting of blower supplied airstones is used for aeration and water circulation. Ponds are aerated in warmer months, in order to prevent oxygen depletion and are allowed to draw down during seasons of low precipitation. Figure 4 1 shows an aeria l image of the facility and the ponds used in this research. Ponds are labeled with letters for identification. Aeration hoses and bubble plumes can be seen at the surface of most ponds, where they resemble a beads on string formation ( Figure 4 1). Table 4 1 lists the observed ponds and their dates of observation.

PAGE 115

115 Table 4 1 Total nitrogen, total phosphorus, and chlorophyll a in aquaculture ponds used in the study. Date Pond TP TN Chl 8/20/2009 1B 59 1390 13.7 8/20/2009 2C 67 1550 49.6 8/20/2009 3K 106 2260 67.6 10/13/2009 D 483 6510 120.8 10/13/2009 J 324 4840 494.4 10/13/2009 L 205 3790 209.8 10/13/2009 N 50 930 17.3 11/17/2009 C 93 1730 32.7 11/17/2009 G 102 1770 70.6 11/17/2009 I 100 1550 74.6 11/19/2009 D 149 2300 125.0 11/19/2009 H 131 2760 65.6 11/19/2009 J 20 1770 3.0 11/19/2009 L 128 3630 215.6 11/19/2009 N 40 930 13.0 12/21/2009 A 97 1000 81.3 12/21/2009 E 43 580 36.6 1/11/2010 C 28 330 14.4 1/11/2010 G 50 470 29.3 1/11/2010 H 113 830 105.6 1/11/2010 I 36 430 31.8 1/11/2010 J 17 320 0.8 1/11/2010 K 57 770 92.8 1/11/2010 L 57 1140 65.1 1/11/2010 M 61 660 2.7 1/11/2010 N 18 400 6.1 1/12/2010 A 52 430 39.2 1/12/2010 B 26 290 12.1 1/12/2010 E 28 440 14.2 1/12/2010 F 21 370 7.9 Mean 92.03 1539 70.4 Std Dev 98.04 1482.8 97.5 Max 483 6510 494.4 Min 17 290 0.8

PAGE 116

116 Figure 4 1. Tropical a quaculture facility and ponds utilized in this study 4.3.2 Water S a mpling and L aboratory A nalysis Water samples were collected for laboratory analysis immediately following spectral data collection for each pond. Samples were taken from the water adjacent to th e in water reflectance targets in 1 L high density polyethylene Nalgene flasks and placed on ice in a darkened cooler. Chl a concentrations (g/L) were deter mined spectrophotometrically following pigment extraction with 90% ethanol (Sartory and Grobbelaar 1984). Total N concentrations were determined by oxidizing water samples with persulfate and determining nitrate nitrogen with second derivative spectroscopy 1985; Wo llin, 1987). Total P concentrations were determined on persulfate digested samples (Menzel and Corwin, 1965) following Murphy and Riley (1962).

PAGE 117

117 4.3.3 H yperspectral I mage A cquisition and C alibration T arget P latform Images were collected using a hyperspectral ( HS ) imaging sensor acquired from AutoVision, Inc. The sensor is composed of a V10E Specim Im Spector holographic grating (Aikio, 2001; http :// ) spectrograph, manufactured by Spectral Imaging Ltd (Finland). T he grating spectrograph disperses the energy passing through a Schneider mount lens and by an Imperx ( ) IPX 2M30H L digital camera attached to the back of the spectrograph. The energy entering the spectrograph slit is distributed by the spectrograph optics into different wavelengths and captured by th twodimensional Charge Coupled Device (CCD). One dimension on the CCD represents the spatial component and the other dimension represents the spectral component. Each image captured by the Imperx camera represents a line in the formed HS i mage. Images are built up line by line either through platform movement or through a rotating mirror mounted in front of the lens. The latter configuration was used in this research to produce an 800 825 pixel image. The sensor is capable of producing 27 0 band images in the range from 400 to 1000 nm with around 2 3 nm band width. Bands close dominated by noise. This was especially true for the water parts of the image where low signal was expected due to water absorption. Visual inspection of the bands revealed appropriate signal at the range from 440 to 958 nm. Each series of image captures for a pond was preceded by acquiring a dark current frame. A hand operated crane (Kessler Crane, Inc. Plymouth, IN) was used to hold the HS sensor above the water surface and away from plants at the margin of the pond so

PAGE 118

118 Figur e 4 2 A shows the HS sensor mounted on the Kesler crane during one of the acquisition sessions. A floating target platform was designed and manufactured using PVC pipes to mount multiple calibration targets. Low cost ($30/target), reusable white targets (20 cm30 cm) (Kontrollkart:Novoflex Memmingen, Germany) were deployed onto the floating target platform. A three level design was made to facilitate data reflective target s suspended just above water level. A second level of targets was held beneath the water surface at approximately 10 cm in depth. Finally, the third level was con f igure d at a distance approximately 20 cm below the second ( 30 cm from the water surface). Th e submerged targets (ST) at 10 cm and 30 cm depth will be denoted ST 10 and ST 30, respectively. Figure 4 2 B shows the target platform as captured from the pond bank using a consumer digital camera. A B Figure 4 2. H yperspectral i mage acquisition A ) image of our HS sensor mounted on a crane; B ) p hoto of our floating platform carrying different targets.

PAGE 119

119 Figure 4 3. Linearly stretched RGB representation of the H yperspectral image of pond K demonstrating used ROIs. 4.3.4 Data P rocessing Data from the HS imaging spectrometer was processed using ENVI 4.5 (ITT Corporation, White Plains, NY). Regions of Interest (ROI) were manually digitized in each image to select pixels from (1) above water calibration target; (2) 10 cm below water target ( ST 10) ROI; (3) 30cmbelow water target (ST 30) ROI, and (4) pond water regions, which will be denoted water only (WO) ROI in the manuscript, as shown in Figure 4 3. Basically, a polygon was digitized to mark all pixels belonging to a region of interest (ca libration Water samples were collected for laboratory analysis immediately following target or just water). The spectra of the pixels within each region of interest were extracted. For example, all spectra of the pixels captured for the submerged target at 10 cm depth (ST 10) for a certain pond were extracted and averaged. This means that for each pond, we ended up with 4 different spectra to analyze representing the 4 different conditions (ROI) listed above. The above water calibration target data was used as a white reference to calibrate for atmospheric conditions. The submerged

PAGE 120

120 targets were used to evaluate our hypothesis that white targets will enhance the upwelling signal leading to a better Chl a estimation models. The number of pixels per ROI was usu ally more than 3000 pixels. In a few cases of highly turbid ponds or the existence of platform shadows, smaller numbers of pixels were used for some of the ST 30 ROI due to hardship in recognizing the boundary of the submerged calibration card (at 30 cm de pth). Unlike in situ spectrometer measurements where multiple spectral readings (observations) can be sequentially acquired and averaged to reduce noise, each pixel collected by our imaging system represents a single spectral reading (observation). Pixels belonging to the same ROI are considered multiple observations for the same condition (e.g. submerged target at 10 cm depth ST 10). Averaging the digital numbers of all pixels belonging to the same ROI provides single spectra representing this ROI. Averag ing these values suppresses the noise common to this technology. In order to filter out outliers, we sorted the digital numbers of all pixels belonging to each ROI and averaged only a certain percentage of the central digital number values. In other words, we excluded (cut off) a certain percentage at each end of the sorted data set. We tested our model by averaging all pixels (0% cutoff), the central 90% of the pixels (10% cutoff), the central 80% of the pixels (20% cut off), and the central 60% of the dat a (40% cutoff). Our results indicated slight differences in model results using a cutoff percentage of 10% or more. This set of data (average of the central 90% of the sorted digital numbers) was used in the analysis and reported in the remaining sections of the manuscript. At each pond, a dark current signal was acquired before image acquisition. Although the time gap between dark current acquisition and image capture was short in

PAGE 121

121 duration, our testing showed that this time gap was often sufficient for no ise level changes in the captured images. Mobley (1999) suggested that subtracting the reflectance value at further infrared bands (e.g. 750 nm) in deep water regions, where water absorbs almost all downwelling irradiance, could correct for the water surfa ce reflectance in Sea WiFS satellite imagery. Several end bands (926 958 nm) at the edge of the sensor spectral sensitivity range in the near infrared region receive low signal due to high water absorption in this region, spectrograph technology (grating a nd CCD), and/or relatively lower downwelling irradiance in this region of the spectrum. Nonzero signals received at these bands may serve as an estimate for the base signal level (signal offset). In our analysis, the average of the digital number values fo r bands 926 958 nm of the WO ROI spectrum was subtracted from the pixel digital number of each ROI. Next, individual band values for the two submerged (ST 10 and ST 30) and the water only (WO) ROIs were divided by the corresponding band values of the above water spectra for each image data set (pond/date) as depicted in E quation 4 3 to compute the reflectance values for each of these ROIs. R j DN j w (926 958)] / [ DN r w (926 958)] ( 4 3) type (10 cm submerged target ST 10, 30 cm submerged target ST 30, or water only (WO)), and R j j ROI j. Similarly, DN r or the above water reference w (926 958) term represents the average digital number of the water only ROI pixels for bands 926 958 nm, respectively.

PAGE 122

122 Figure 4 4 shows the reflectance for the data extracted from the wat er only (WO) and 10 cm submerged target (ST 10) ROIs for ponds B (Chl a = 12.1 g/L), H (Chl a = 65.6 g/L), K (Chl a = 92.8 g/L), and L (Chl a = 215.6 g/L). The f igure shows that the a concentration curves. It also shows that the open water data has higher noise content due to the low signal coming out of the water compared to the strong signal resulting from the submer ged calibration target. As mentioned in sub section 4.2.2 we used two band (reflection ratio) index utilizing the 680 and 700 nm reflectance values and a three band index utilizing the 680, 700 nm, and the 769 nm reflectance values in our analysis. The fi rst two bands are the most prominent trough and peak, respectively, indicating increased Chl a concentration as shown in Figure 4 Gitelson, 2005), while band 769 in the three band model is used to normalize for surface water reflection (Mobley, 1999). Equations 4 4 and 4 5 depict the reflectance ratio and three band indices used in this study. Visual inspection of the relationship between the reflectance ratio (or the three band index) and Chl a con centration indicates a clear linear relationship. This observation is supported by the great majority of research in this area (e.g. Mittenzwey et al., 1992; Gons, 1999; Schalles et al., 1998; ar regression model was used to model Chl a concentration resulting from the water sample analysis and the spectral indices (reflectance ratio and three band index). Reflectance ratio j = R rs j (700) / R rs j (680) ( 4 4) Three band index j = (1/R j 1/R j (700))R j (769) ( 4 5)

PAGE 123

123 A B Figure 4 4. Reflectance spectra of four representative ponds with varying Chl a content collected from pixels in two ROIs: A) the water only (WO) and B) the Submerged Target at 10 cm depth (ST 10) ROI. 4.4 Results Tab le 4 1 shows that, across 30 pond samples, Chl a, total P, and total N concentrations were 0.8 494 g/L, 17 483 g/L, and 290 6510 g/L, respectively. Simple linear regression models were created to estimate Chl a concentration from reflectance ratio and t hree band index values. Laboratory results of Chl a concentrations were the response variables and the reflectance ratios and three band indices for each ROI were the predictors. Due to the small sample size (n = 30), a Leave one out (LOO) algorithm (Marte ns and Dardenne, 1998) was used to validate the models. The Leave one out validation algorithm repeatedly builds the model using all data except one data sample (one pond observation). The model is used to predict the value for the excluded sample and comp ute the residual (difference between predicted and observed values). The technique is repeated for all data points. R 2 is computed using the sum of the square of the residuals and used to validate the model. The coefficient of determination R 2 and the RMSE for model calibration results and the RMSE for model validation are presented in Table 4 2.

PAGE 124

124 Six models were created using the reflectance ratio and the three band index values computed for each of the three ROI data sets (WO, ST 10, and ST 30 ROIs). Figu res 4 5 and 4 6 and Table 4 2 show the results of using the reflectance ratio and the three band models, respectively, for the WO, ST 10, and ST 30 data sets. The largest R 2 (0.982) (smallest LOO RMSE = 13.4 g/L) was observed for the three band model when the ST 10 data was used ( Figure 4 6 B ). The corresponding reflectance ratio model achieved an R 2 value that is slightly less (0.975) ( Figure 4 5 B ) with a 20.7 g/L LOO RMSE. The WO data set gave R 2 values of 0.862 ( Figure 4 6 A ) for the three band model and 0.833 ( Figure 4 5 A ) for the reflectance ratio model. When using data extracted from the ST 30 ROI, the obtained R 2 values were the lowest among the three ROI results. This could be attributed to the hardship in determining the ST 30 ROI boundaries due to w ater turbidity and shadow cast by the target platform, and/or the uneven spatial distribution of Chl a concentration at 30 cm depth in the ponds. Other models were created using the same ROI data sets with dark current values subtracted from the observed digital number of each band instead of subtracting the average of 926 958 nm band values DN j (926 958 nm) as depicted in E quation 4 3 These dark current values were recorded immediately before capturing each HS image at every pond. The objective of creati ng these models is to study the advantage of using data extracted from the image to calibrate for signal level offset compared to using dark current data recorded before image acquisition. Table 4 3 summarizes the R 2 of the model calibration results and th e RMSE of the model calibration and LOO validation results when dark current values are subtracted. The table shows that R 2 values were less than corresponding ones shown in Table 4 2 (subtracting average 926 958 nm

PAGE 125

125 band values) with a maximum of 0.912 obt ained for the ST 10 data using the reflectance ratio model. Total P and N concentrations were analyzed using the same reflectance ratios ( Figure 4 7) and three band values ( Figure 4 8). The R 2 values for these relationships were relatively low (e.g. 0.478 and 0.512 for the P and N concentrations, respectively, using the three band model). Table 4 4 summarizes the model calibration and LOO validation results. These values are comparable to the R 2 values of the models created between the Chl a and the N and P concentrations. It should be noted that N and P concentrations for pond D (observed on Oct 13, 2009) were consistently identified by the model as outliers. These concentrations were found to be the maximum among all observed ponds (TN = 6510 g/L and TP = 483 g/L). Removing this sample increased the R 2 values for all models as shown in Figure 4 8. For example, the R 2 values for P and N models increased to 0.823 and 0.678, respectively, using the three band model values. 4.5 Discussion Hoyer et al. (2005) rep orted that the mean Chl a concentration across 84 Florida lakes was 13 g/L. Mazumder and Havens (1998) found that subtropical lakes where small herbivore zooplankton were present had a mean Chl a concentration of 9 g/L. Aquaculture ponds have much greate r nutrient concentrations and are eutrophic or hypereutrophic due to their spatial characteristics and diverse/dynamic nutrient cycle. This was evidenced in our study where Chl a concentrations ranged between 0.8 and 494.4 g/L (Table 4 1). The use of a HS imaging sensor in Chl a estimation increased the sample size (number of pixels), which enabled noise minimization and accounted for local spatial variability in water quality. In our analysis, we averaged the spectra of

PAGE 126

126 thousands of pixels in each ROI and used averaged spectra in the Chl a models. Unlike most in situ studies, where more than one spectrometer was used simultaneously to observe water (some experiments used subsurface readings) and collect calibration spectra, our method captured a whole scen e, including upwelling water radiance, above water calibration target spectra and upwelling radiance from submerged targets. Although it was not implemented in this study, this setup enables the use of HS imaging system mounted on a mobile platform (such a s a boat) to continuously acquire imagery in larger or extended water bodies such as stormwater ponds, streams and wetlands. The use of a submerged target at 10 cm depth (ST 10) provided the strongest correlation (R 2 of 0.975 for two band reflectance ratio and R 2 of 0.982 for three band model) between Chl a concentration and the spectrally derived indices. Studying the difference between the spectra of pixels in open water (WO) and in the ST 10 areas ( Figure 4 4), we notice much smoother spectra in the ST 10 indicating a higher signal to noise level. This can also be extrapolated to the effect of signal enhancement on minimizing other error sources expressed in the water quality bio optical (Gege and Albert, 200 6) model such as surface water reflectance and variations in the water column optical properties. Overall, we conclude that the results obtained using reflectance ratios (two band) and the three band models were not significantly different in most cases. T able 4 2 shows that, in general, the three band model performed slightly better than the two band model, especially in the WO case. The LOO validation results were consistent in this case with the model calibration results with a RMSE of 13.2 g/L for the validation results and 13.4 g/L for the calibration results.

PAGE 127

127 The R 2 of Chl a estimation models using data in the WO region (R 2 = 0.833 for two band reflectance ratio model and R 2 = 0.862 for three band model) matched many of the results obtained in Case 2 waters (e.g. Gitelson et al., 2007; Jiao et al., 2006; Sudduth et al., 2005). In addition, they surpassed the results reported by Zimba and Gitelson (2006) for hypereutrophic ponds (R 2 = 0.71 and 0.78 for two and three band models, respectively). Chl a m odels of ST 30 data were consistently lower than the results of the ST 10 and WO models probably due to the small sample size (number of pixels) obtained for the deeper target. Difficulties in identifying and selecting the deepest ROI due to lower target v isibility contributed to the obtained lower correlations and provide no reason for using deeper targets. The results depicted in Table 4 3 were inferior compared to their counterparts presented in Table 4 2, respectively. For example, the R 2 value for the three band model applied on ST 10 data was 0.982, while the corresponding result in Table 4 3 was 0.901. These results indicate the effect of using data extracted from the image to account for signal level instead of using dark current data captured before the actual image acquisition. Our results suggest the possibility of signal level change between the time of dark and pond image acquisitions. Subtracting the average digital number (WO ROI) of bands close to the end of the sensor sensitivity limit in the infrared region accounted for dark current variation. Investigating the relationship between the reflectance indices and the total N and P concentrations ( Figure 4 7 and Table 4 4) is actually a study on the suitability of Chl a concentration to be used as a measure of N and P concentrations in our aquaculture pond systems. Several research studies have sought to correlate Chl a with P and other water quality parameters (e.g. Fagerbakke et al., 1996; Kamarainen et al., 2009) in

PAGE 128

128 different water systems. Al though we observed moderate correlation between the spectral indices used to model Chl a concentration and P and N concentrations, our finding that eliminating one of the pond samples (sample with highest P and N concentrations) increased the R 2 from 0.478 to 0.823 for P and from 0.512 to 0.678 for N ( Figure 4 8) may reflect the potential existence of other N and P forms such as P associated with particulates or N present in organic forms which are not available to algae. This also indicates the need for fu rther studies to incorporate other bio geochemical parameters such as different forms of N and P in water bodies and other parameters, such as silicon affecting the eutrophication process, before being able to use the spectral indices sensitive to Chl a to accurately estimate nutrient concentrations. We believe that the results of this manuscript open the doors for using hyperspectral imagery on close range mobile platforms (e.g. mounted on a vessel) for water quality studies. The results also indicate the possibility of manufacturing standalone sensors that employ absorption spectroscopy concepts through the utilization of a light source, photodiodes (to measure reflectance in 680, 700, and 769 nm bands) and a white target. The device would simulate the s ame configuration used in this study and uses a white target to amplify the energy passing through the water to accurately measure Chl a, P and N concentrations in eutrophic waters at different water depths. 4.6 Conclusion We evaluated the utility of a new gro und based HS imaging sensor, in conjunction with a set of floating and submerged calibration targets, for estimating Chl a in shallow nutrient rich aquaculture ponds. Unlike in situ spectrometer systems for Chl a

PAGE 129

129 estimation, our imaging system provided num erous pixels for the water body. The sensor captured spectra for water and calibration targets simultaneously, eliminating the need for two (or more) spectrometer observations. Using submerged targets at 10 cm depth significantly enhanced the received sign al and provided the highest correlation (R 2 = 0.982 and validation RMSE = 13.4 g/L) using a three band model. Spectral indices extracted from water were less correlated with Chl a concentration. We found that accounting for signal noise by subtracting the average of the last few bands in the near infrared region (926 958 nm) gave better results than subtracting dark current obtained before image acquisition. Total P and N models with three band index values resulted in a moderate level of correlation. Remo ving a single observation led to a 2 from 0.478 to 0.823 for the P model using three band index values, which highlighted the need for more research to directly model nutrient concentrations from spectral information. N onetheless, our results for Chl a estimation are promising and indicate that HS images can be successfully used to determine prevailing Chl a concentrations in eutrophic and hypereutrophic aquaculture ponds. Further research is needed to calibrate the proc edure for other types of water bodies such as storm water retention ponds and wetlands.

PAGE 130

130 Table 4 2 Chl a model calibration and validation results using two band and three band indices. Model Data extracted from Model calibration LOO validation R 2 RMSE LOO RMSE Reflectance ratio (two band) Water only (WO) 0.833 40.6 52.0 Submerged target at 10 cm depth (ST 10) 0.975 15.7 20.7 Submerged target at 30 cm depth (ST 30) 0.551 69.9 88.7 Three band index Water only (WO) 0.862 36.9 37.1 Submerged target at 10 cm depth (ST 10) 0.982 13.2 13.4 Submerged target at 30 cm depth (ST 30) 0.644 62.2 89.9 Table 4 3 Chl a model calibration and validation results using two band and three band indices (reflectance computed by subtracting dark current values obtained before pond image acquisition). Model Data extracted from Model calibration LOO validation R 2 RMSE LOO RMSE Reflectance ratio (two band) Water only (WO) 0. 301 83.0 87.2 Submerged target at 10 cm depth (ST 10) 0.9 12 29.4 43.2 Submerged target at 30 cm depth (ST 30) 0. 239 91.0 98.8 Three band index Water only (WO) 0.739 51.9 64.7 Submerged target at 10 cm depth (ST 10) 0.9 01 31.2 43.8 Submerged target at 30 cm depth (ST 30) 0.394 86.6 112.9 Table 4 4 Calibration and validation results for t otal P and t otal N models. Item Data extracted from Model calibration LOO validation R 2 RMSE LOO RMSE Total P Reflectance ratio (two band) 0. 438 74.8 77.5 Three band index 0.478 72.1 82.6 Total N Reflectance ratio (two band) 0.496 1071.0 1162.8 Three band index 0.512 1054.1 1322.3

PAGE 131

131 A B C Figure 4 5. Relationship between Chl a and reflection ratios for A) water only (WO), B) submerged targets at 10 (ST 10) and C) 30 cm depth (ST 30) ROIs. A B C Figure 4 6. Relationship between Chl a and three band index for A) water only (WO), B) submerged targets at 10 (ST 10) an d C) 30 cm depth (ST 30) ROIs.

PAGE 132

132 Figure 4 7. Total P and N relationship with reflectance ratios and three band indices. Figure 4 8. Total P and N relationship with reflectance ratios and three band indices after removing pond D (Oct 13, 2009) observations.

PAGE 133

133 CHAPTER 5 ANALYSIS OF HYPERSPECTRAL IMAGERY TO DISCRIMINATE AQUATIC VEGETATION IN A RIVER ESTUARINE ECOSYSTEM 5.1 Introduction Proper classification of aquatic vegetation in river estuaries is a vital component for managing fresh water ecosystems. Field surveying, which is the conventional way to map and estimate aquatic vegetation communities is labor intensive, expensive and time consuming. Remote sensing offers an efficient tool for discriminating vegetation in shallow waters, where the imprint of veg etation beds in the radiation spectra can be analyzed. There are very few studies to classify aquatic vegetation in black water riverine system. Black and fresh water is characterized by higher absorption of visible lights by Color Dissolved Organic Matter s (CDOM), which possesses unique optical properties and brings different analytical challenges than most turbid case 2 waters. Interpretation of radiation coming from aquatic vegetation is complicated by the Inherent Optical Properties (IOPs) of the wate r, sensing geometry, varying water depth and reflectance properties of different plant species as well as the complex structure of these plants. Unlike other bottom materials, aquatic plants are not flat and their reflectance properties are complex depend ing on many factors including plant species and density, canopy architecture, vertical structure, leaf orientation and whether the plants are emerging to water surface or submerged. Light is reflected directly from plant parts emerging to the surface, whil e for submerged region, radiation gets attenuated while propagating through the water before it hits the submerged plant and after it gets reflected.

PAGE 134

134 Various supervised classifiers have been used for land cover and wetlands class ification. Maximum L ikeli hood is widely used parametric classifier which assumes that the spectral information for each class in each band is normally distributed. The classifier calculates the probability of a pixel belonging to a specific class (Jensen, 2004). The Spectral Angul ar Mapper (SAM) algorithm is a popular classifier that is often between pixel spectrum and reference spectrum in band dimensions and assigns a pixel to the reference class that yields the smallest angle (Kruse et al, 1993). Non parametric methods such as Artificial Neural Network (ANN) and Support Vector Machine (SVM) received more attention in remote sensing in recent years, especially for land cover classification and pat tern recognition (Corsini et al., 2003; Gualtieri and Chettri, 2000; Heermann and Khazenie, 1992; Jensen et al., 2009; Melgani and Bruzzone, 2004, Pal and Foody, 2010). Semi analytical models of radiation transfer have been utilized in many studies to ret rieve water depths, properties of water column, and bottom types from natural waters. Earlier studies used simplified models for direct parameter retrieval from remote sensing data of oceanic or coastal waters (Lyzenga et al, 1985, Fey et. al, 1990, Bierwi rth et. al, 1993, Schweizer et al., 2005). These methods were examined mostly using multispectral satellite imagery and required field sampling data. Inverse modeling approaches were used in recent studies. These approaches utilize analytical models to ret rieve parameters such as water constituents, bathymetry and bottom properties simultaneously and indirectly either by inverse modeling through convergence and iteration to minimize error (Lee et. al., 1998, Albert and Mobley, 2003, Klonowski et al.,

PAGE 135

135 2007, Brando et al., 2009), or by spectrum matching with model derived spectra (Mobley et al., 2005, Hedley et al., 2009). These methods are highly sensitive to data quality. They require high signal to noise ratio imagery with narrow contiguous bands as well as additional observations such as illumination, sky condition, water constituents and bottom types. Radiance reflected from black water body has low signal to noise ratio which mostly produces unstable solutions. This study investigates the implementation of various classifiers on normalized images and evaluates their suitability to detect or discriminate aquatic vegetation in black river estuary ecosystem. We hypothesize that the performance of non parametric non linear classification methods is better tha n that of traditional linear classifiers for the complex case of aquatic vegetation. Airborne hyperspectral imageries of case 2 black water of the lower St. Johns River basin in Florida were analyzed while field data taken with conventional survey methods was used for training. Classifiers were implemented on the original images and images normalized for surface water reflectance and water depth. Simple indices were also developed to detect aquatic vegetation, to separate emerging and submerged vegetation a nd to mask broad categories of upland and deep water pixels. Quality assessment is performed using field data set aside for this purpose. 5.2 Site and Background 5.2.1 Site D escription Hyperspectral images of the southern segment of the lower St. Johns River basin w ere analyzed to classify aquatic vegetation. St. Johns River is a northward flowing black water river in the east coast of Florida. The southern segment of the river is broad with slow flow, which is inhabited by various aquatic vegetation that provide s fo od and

PAGE 136

136 shelter for many organisms including fish and invertebrates and grazing bed for manatees. St. Johns River water contains chlorophyll, CDOM and non algal particles in appreciable quantities (Gallegos, 2005). The CDOM concentration which is responsibl e for the black color of the water varies spatially and temporally. It is usually high during the rainy season and is less during the dry season. Water quality parameters measured in April 2006 at water quality stations in the river indicated Chl a concent total suspended solids (TSS) from 7 to 31 mg/L with 16 mg/L mean value; and Color value from 70 to 100 CPU with 90.5 mean value. It should be noted here that such water quality concentrations, espec ially Chl a and suspended particles, may vary significantly in the near shore areas, where aquatic vegetation grows. 5.2.2 Data Eight hyperspectral images of 1.2 meters spatial resolution and 17 narrow bands with 7.4 nm average band width were used in this study These bands are not contiguous and their positions in the wavelength spectrum are shown in Figure 5 1. The images have 0.6 km width and lengths varying from 6km to 15km. They were acquired in the short duration from April 13, 2006 till April 16, 2006. Da ta collection was done during clear weather in the morning and afternoon times such that sun zenith angle is maintained between 25 to 45 (Eason et al., 2006) in order to minimize sun glint. Figure 5 1. Center wavelength (nm) positions of the image bands 400 450 500 550 600 650 700 750 800

PAGE 137

137 Field polygons were surveyed along the shoreline at random locations in one month period from April 5, 2006 to May 4, 2006. Data included species type, aggregate percent coverage, plant foliar density, w ater depth, average plant height, and substrate type. Field polygons were then shrunk by 1 pixel (1.2 m) along the border in order to compensate for potential positional errors. Field survey data includes 58 polygons, among which 35 polygons were found to be pure polygons representing specific classes with 80% 100% coverage. Half of these pure class polygons were selected to train the classification and the remaining half was used for accuracy assessment. Assignment of class polygons are shown in the Table 5 1 below. Field polygons represented four aquatic vegetation species and bare bottom. In addition to field polygons, regions of interest for 3 additional classes, upland vegetation, upland ground/built up and deep water, were selected from the images. Ta ble 5 1 Assignment of field polygons for classification training and for assessment. Class Class D escription Classification polygons Assessment polygons Bare Bare bottom 6 6 Val Vallisneria Americana 6 5 Najas Najas Guadalupensis 3 5 Algae Plants heavily covered with Algae 1 1 Emer Emergent plants 1 1 Sum 17 18 5.2.3 Preprocessing M ethods Captured images were atmospherically corrected and converted to reflectance values using the ENVI FLAASH module which is based on MODTRAN4 atmospheric model (Adler Golden et al., 1999, Eason et al., 2006). Then, they were geo rectified using proprietary so ftware to the design accuracy of 2 pixels (~2.4m). Water has low signal to noise ratio, with negative values observed in some bands, especially in deeper

PAGE 138

138 regions. Maximum Noise Fraction (MNF) (Green et al., 1988) based de noising was applied on these ima ges by selecting only MNF bands with significant signal to invert back to image domain. 5.2.4 Radiation T ransfer M odels Total reflectance (R tot ) from water region recorded in imagery of an above surface sensor is composed of a portion directly reflected from th e water surface, R surf (water surface reflectance), and another portion upwelling from water body (R). R tot = R surf + R ( 5 1) Water surface reflectance (R surf or bottom properties and therefore eliminated or minimized in most studies. Sun glint is usually avoided by planning proper imaging time and direction. When light propagates through water, it gets attenuated due to absorption and scattering by water and water constituents. Remote sensing r eflectance from deep water (ratio of upwelling radiance to downwelling irradiance just under water), which corresponds to water column only, depends on absorption and backscattering coefficients of water column and modeled in many studies as: b / (a + b b ) ( 5 2) where, a and b b are absorption and backscattering coefficients of water column. In case of shallow water, reflectance from water (R sh ) has two components reflectance from water column and reflectance from bottom layer. In general radiative transfer model for shallow waters can be expressed as: R sh A 1 e (Kd+Kuw)H 2 e (Kd+kuB)H R b ( 5 3)

PAGE 139

139 where, K d K uW and K uB are attenuation coefficients of downwelling irradiance, upwelling radiance from water and upwelling radiance from bottom, respectively and H refers to water depth in meters; R b 1 and A 2 are closer to one (A lbert and Mobley, 2003). By approximating A 1 2 uW K uB u this equation can be rewritten as: R sh R = (R b R). e (Kd+Ku)H ( 5 4) Taking natural l og from both sides of E quation 5 4 LN (R sh R) = LN (R b R) (K d +K u )H ( 5 5) Deducting total reflectance of shallow water to deep water reflectance, R sh tot R tot = ( R sh + R surf ) ( R + R surf ) = R sh R ( 5 6) Equation 5 6 demonstrates that (R sh tot R tot ) cuts out water surface reflectance and can be substituted to (R sh R) in E quations 5 4 and 5 5. Above expressions show that reflectance value gets attenuated non linearly with negative power function of water depth and relates to bottom reflectance and de ep water reflectance. Equation 5 4 shows that shallow wate r reflectance value normalized by deducting deep water reflectance, relates to normalized bottom reflectance times negative power function of water depth and attenuation coefficients. Equation 5 5 linearizes this relationship. 5.3 Methodology 5.3.1 Image N ormalizatio n In this study, image reflectance values are normalized for water surface reflectance by deducting deep water spectra. Deep water region of interest ( ROI ) is selected for each image at open deep water area without sun glint or cloud shade where water dept h is more than 3 meters. Spectral data is extracted from the ROI.

PAGE 140

140 Rather than using minimum spectral values, low deep water spectra is determine d by calculating (Mean value 2 Standard Deviation) in order to avoid outliers of darker spectra. Image norma lization is then conducted by deducting low deep water spectral values band by band. This normalization improves the imagery of aquatic region by removing water surface reflectance, which then relates directly to bottom reflectance and negative power (non linear) function of water depths and attenuation coefficients as shown in E quations 5 4 and 5 6. 5.3.2 Depth I nvariant T ransformed I mage The image data is transformed into depth invariant bands based on a method proposed by Lyzenga et al (1981) and modified here to suit hyperspectral imagery. Natural log is computed using the normalized im age values (as in left side of E quation 5 5), which yields log value for each band ( log band X) and can be expressed as, X = LN(R sh R) ( 5 7) According to Lyzenga me thod, scatter plot between two bands for pixels with varying water depths and the same bottom type are linear and parallel to scatter plots from different bottom strata. The offset value derived from the band pair scatter plot is unique for each type of bottom. Lyzenga et al. (1981) utilized a visible spectrum band pair of Landsat corresponding to Band 1 (Green band) and Band 2 (Red band) to derive a depth invariant transform image which is then used to discriminate major bottom strata types. In our case of Hyperspectral imagery of shallow waters, multiple log b and pairs, as calculated using E quation 5 7, were evaluated for same bottom layer. For this purpose, field polygons of same bottom type (e.g. bare sand bottom) with varying water depths were used as regions of interest for analysis. Scatter plots of pairs of log bands from the

PAGE 141

141 selected regions of interest were analyzed and band pairs with high correlation were selected. Ratio of attenuation coefficients then was calculated for each of these pairs and a depth invariant transform band (Y) was created using following equations. Using selected regions computed as follows. a = (var i var j ) / 2cov ij ( 5 8) where X i and Xj are log values of i and j band pair; var i and var j are variances of X i and X j ; cov ij is covariance of X i and X j for selecte d regions of same bottom strata with varying depths. Then the ratio of attenuation coefficients (K i /K j ) is calculated as (5 9) Finally, depth invariant transform band is created for each band pair using: Y ij = X i (K i /K j )X j ( 5 10) where, Y ij is depth invariant band using band values X i and X j K i /K j is the attenuation coefficient ratios derived. 5.3.3 Classifiers Three non parametric classifiers, (i) Artificial Neural network (ANN), (ii) Support Vector Machine (SVM ) and (iii) Spectral Angular Mapper (SAM), were investigated and compared with Maximum Likelihood (ML) classifier. These four classifiers were applied to filtered original images, normalized images and depth invariant transform images using the selected t raining sets. Visual inspection of the images revealed that the first four bands are heavily contaminated with noise and stripes. Preliminary results showed exclusion of these four bands leads to better classification rather than using all bands.

PAGE 142

142 According ly, classifiers were tested on only 13 bands of 17 original bands. In a different experiment, to increase S/N ratio, we averaged each successive two bands of the first four bands and tested the classifiers on 15 band imagery. The ANN classifier is a mathe matical model consisting of interconnected group of nodes. It is used as non linear data modeling tool and is often used in classification and pattern recognition (Haykin, 1994, Keiner and Yan 1998). In our experiments, layered feed forward network ANN cla ssification technique was employed using back propagation for supervised training. Different activation functions and parameter values were tested and non linear classification was opted for by selecting one or more hidden layers for separation. The SVM classifier is another non parametric classifier that separates the classes with a decision surface, known as optimal hyperplane, such that it maximizes the margin between the classes (Cortes and Vapnik 1995). It provides better classification results for complex and noisy data and can be implemented as non linear classifier through the use of non linear kernels. We experimented with different kernel types in the SVM classifier including the non linear second degree polynomial kernel and radial basis functi on and with different parameter values to find the optimal ones. The SVM the reference classes. The classified image is created by assigning to each pixel the class with highest probabilities. SAM is a non parametric spectral classifier commonly used with hyperspectral images, which compares image spectra with reference spectra in terms of solid angle between these spectra as n dimensional vectors. It is known to be i nsensitive to

PAGE 143

143 illumination or albedo as it compares only vector direction but not the magnitude of the two spectra (Kruse et al., 1993; Van der Meer, 2006 ). 5.3.4 Indices for D etection and C ategorization Study of spectra from field survey polygons show ed that typically Band 9 (Red 682nm ) is deeper and Band 12 (short Infrared 710nm ) is higher in aquatic vegetation than in bare bottom. Figure 5 2 shows some typical spectra in aquatic region. Emerging vegetation yield s higher values for i nfrared bands. Uplan d pixels are separable due to brighter values. Deep water region has all values lower except for regions with higher glint Figure 5 2. Typical spectra of aquatic vegetation and bare bottom. S imple indices were developed empirically and tested for separ ating aquatic vegetation, bare bottom, upland pixels and deep water regions, which are as follows: Vegbare index: (Band12 Band9) / (Band8 -100 -50 0 50 100 150 200 250 300 350 400 450 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 Bare bottom Vegetation Vegetation Vegetation Deep water Bare bottom

PAGE 144

144 Aquaveg index: (Band14 Band11) / (Band11 Band9), se Deep index: (Band12 Optimum index thresholds were determined using the training data and assessing the results. A c ombination of the se indices was used to categorize whole area broadly into uplands, shallow water and deep water and to detect aquatic vegetation. 5.3.5 Assessment M ethods Quality assessment was performed using the 18 pure class field polygons previously set aside for assessment These polygons contain more than 700 pixels in total. Two types of accuracies were considered for results of implemented classifiers: 1. Classification accuracy Accuracy to classify data into the five aquatic classes listed in table 5 1. Upland and deep water classes were not included in the assessment because only aquatic region is our central focus of this analysis, moreover, the inclusion of upland pixels is expected to overestimate the classification accuracies. 2. Detection accurac y Accuracy to detect aquatic vegetation versus bare bottom. In addition to the accuracy obtained using the tested classifiers, the performance of a simple 3 band index was evaluated for submerged vegetation detection accuracy. Overall Accuracies (classifi cation and detection) were first computed in two different ways pixel wise and polygon wise. Overall pixel accuracy is computed as the ratio of sum of correctly assigned pixels to total number of pixels. Overall polygon accuracy is derived by computing t he percentage accuracy of each polygon. Then the average of individual polygon accuracies was computed to give the overall polygon wise

PAGE 145

145 accuracy. Accuracy values were presented as average of overall pixel and polygon percent accuracies. 5.4 Experiments and Res ults 5.4.1 Depth I nvariant T ransform Image 25 of the dataset was used to derive depth invariant parameters due to availability of field polygons of bare sand bottom type with varying water depths up to 0.7 m. Scatter plots of pairs of log bands for pixels in the se polygons were analyzed and 12 pairs were found to be highly correlated. These parameters were used to create depth invariant images for the other seven images used in the study. Attenuation coefficients were calculated for these 12 band pairs as shown in table 5 2. Then, twelve depth invariant bands were created using Equation 5 10 An e xample of depth invariant bands is shown in Figure 5 3 Table 5 2 Attenuation coefficient r atios for 12 band pairs Y Band Band pair Correlation Attenuation c oefficient ratio 1 K5:K6 0.990 9 0.9451 2 K7:K6 0.9990 0.998 1 3 K8:K6 0.9987 0.958 8 4 K9:K6 0.98 40 0.6184 5 K10:K6 0.9916 0.6948 6 K11:K6 0.9912 0.6964 7 K12:K6 0.990 2 0.763 6 8 K13:K6 0.986 7 0.960 2 9 K12:K15 0.9921 0.609 8 10 K13:K15 0.9915 0.7662 11 K16:K15 0.9902 0.8671 12 K17:K15 0.9924 0.9145

PAGE 146

146 5.4.2 Supervised Classification As mentioned earlier, classification of the image was performed using non parametric classifiers ANN, SVM and SAM in addition to the Maximum Likelihood. These four classifiers were applied on the following set of images: 1. Filtered original images (13 bands): Original image filtered with MNF de noising. 2. Normalized images (13 bands): Image data after deducting low spectra of open deep water region. 3. Normalized images (15 bands ): Image data after deducting low spectra of deep water and averaging successive two band pairs of the first four bands of the images. 4. Depth invariant transform images: Image data normalized for water depth. Multilayer perceptron Artificial Neural Netwo rk with back propagation learning was and one hidden layer provided better classification of these images. Figure 5 4 A and 5 5 A show results of ANN applied to the normalized and depth invariant images. Classified images were visually inspected and found appealing and less pixelated. Upland classes (Trees and Ground) and deeper water class were classified fine as well. Support Vect or Machine (SVM) was used with non linear kernels. Polynomial 2nd degree was selected due to better performance in preliminary results. Figure s 5 4 B and 5 5 B show results of classifying the normalized and depth invariant images using SVM. Visual inspection showed that classified images were less pixelated and matched more or less the quality of the ANN results. Spectral angular mapper (SAM) was used on the images with maximum angle set to zero. Results of SAM classified images are shown in Figure 5 4 C and 5 5 C Classification in shallow water looked visually good. Deeper

PAGE 147

147 water regions were more pixelated due to noise in the spectra and upland regions were misclassified in some areas due to matching spectral angle of upland ground pixels with those of aquatic classes. Maximum Likelihood (ML) was implemented on the same images with probability threshold set at zero in order to allow all pixels to be classified. Figure 5 4 D and 5 5 D show the ML classification results on the normalized and depth invariant transfo rm images. In the aquatic region, classification looked more pixelated compared to all other classifiers. Accuracy of classifying shallow aquatic region is sub section. 5.4.3 Indices Combination of indices as m entioned in sub section 5.3.4, was used to separate aquatic vegetation, bare bottom, upland and deep open water. This serves as simple method for detection of aquatic vegetation. It also helps to broadly categorize (mask) uplands and deep water as a supple ment to improve other classification method (e.g. SAM classification). Figure 5 6 shows indices based categorized image. Aquatic vegetation (green and dark green) matches well with results of non parametric classifiers. Upland and deep water regions were c ategorized remarkably well. Accuracy sub section. 5.4.4 Quality A ssessment Assessment of the classification results (refer to Table 5 3) indicates that non parametric classifiers (SAM, ANN and SVM) perform ed better with accuracy over 60% for classification and over 84% for detection when applied to filtered original images with 13 bands (excluding first four noisy bands). The Maximum L ikelihood was found to be least accurate for both classification (45% acc uracy) and detection (59% accuracy) of aquatic vegetation. We can attribute this partially to its inability to incorporate the non

PAGE 148

148 linear optical characteristics of aquatic environment. Image normalization for water surface reflectance eliminates the water surface reflectance. That possibly led to improvement in accuracies when non parametric non linear classifiers (e.g. ANN and SVM) were applied to the normalized images. We believe that the ANN and SVM classifiers, which incorporate all training pixels, co uld have been able to learn the non linear optical complexity of aquatic environment. Result of SAM classification on original filtered and normalized images, however, showed unchanged accuracies. Use of normalized image with additional first two bands, r esulting in total 15 bands, improved the accuracies slightly in most classification results. The ANN classifier yielded the best results, 73% for classification and 91% for detection. Result of classification of the normalized images using Maximum Likelih ood improved slightly but still significantly low. Index based detection yielded good accuracy of 84.8% which is better than accuracy of all ML detection results and comparable to the results of other classifiers applied to original images. Table 5 3 Accuracy assessment results using different classification methods Classifier Original clean image Normalized image Depth Invariant Image 13 bands 13 bands 15 bands 12 bands Classification Overall accuracy : ANN 62.16% 68.73% 73.12% 59.71% SVM 61.91% 64.42% 64.68% 62.35% SAM 64.53% 58.60% 61.18% 64.85% ML 45.37% 49.01% 49.85% 57.17% Detection Overall accuracy : ANN 89.03% 89.72% 90.90% 88.73% SVM 84.49% 88.99% 89.94% 91.57% SAM 86.43% 87.92% 87.98% 92.95% ML 59.12% 71.51% 68.78% 81.25% Indices 84.8%

PAGE 149

149 The use of non parametric classifiers (ANN and SVM) on the depth invariant images produced mixed results. Accuracies, especially detection accuracy, of SVM improved, however ANN accuracy declined. Applying SAM on depth invariant images resulted in high er detection accuracy. Maximum L ikelihood classifier remarkably improved for both classification and detection accuracies. It should be noted that depth invariant images were derived using same attenuation coefficient ratios for all images because of insufficient field data. If sufficient field data is available, coefficient ratios for individual image s can be computed which might improve the classification of the depth invariant images even further. 5.4.5 Discussion Overall, the non parametric classifiers, SVM, ANN and SAM performed better than traditional ML classifier in classifying and detecting aquatic vegetation. Normalization by deep spectra deduction was found to improve image classification when using ANN and SVM. These non parametric cl assifiers likely incorporated spectral variation induced by variations in water depth. The SAM classifier showed good results with origin al clean images. In all cases, Maximum L ikelihood classifier was found least accurate despite some improvement with dep th invariant images. Depth invariant transform seemed to have enhanced image interpretation by allowing ML, SVM and SAM to improve accuracy, especially that of detection. The ANN classifier, which performed best with furth er with depth invariant images. As mentioned earlier in this section a ttenuation coefficient ratios used for depth invariant image transform of all images were determined from a single image due to sample insufficiency, which might have limited the accura cy of classifying such images

PAGE 150

150 Adequate samples for each image would allow to compute individual coefficient ratios and to produce improved depth invariant images. I ndices developed empirically based on study of spectral patterns, was useful to separate ou t broad categories su ch as upland and deeper water pixels and as a supplement to other classifiers such as SAM. Indices were also used to detect aquatic vegetation with moderate overall accuracy of 85% besting detection accuracy of ML classifier. Refinemen t of indices is recommended for future studies with ample field samples and multiple sets of images This study is based on available limited field sample size 1 7 field site pure class polygons with over 1000 inner pixels were used for training classifiers and similar sample size for assessment Number of site polygons was lower than generally recommended for classification but were large enough in size to have sufficient number of pixels for analysis. Th is study and the results are worth while to make general comparison between various methods of processing and classification and to recommend prospects of further studies Further study with sufficient sample data is needed to conduct subtle comparison between performances of these non linear class ifiers. Depth invariant image transform can be improved if sufficient field samples are available and coefficient ratios for all images are computed. With adequate sample size and study of different types of hyperspectral images, indices for detection migh t be refined and improved. 5.5 Conclusion This research investigated non parametric classifiers ANN, SVM and SAM and traditional classifier ML applied to the filtered original images, images normalized for water surface reflectance and depth invariant transfo rm of CASI hyperspectral images.

PAGE 151

151 R esults showed that the Maximum L ikelihood is not as suitable in the case of aquatic environment. Non parametric classifiers (ANN, SVM and SAM) yielded better results. The ANN and SVM produced visually smooth classification map in all aquatic and upland regions. The SAM classifier was better in analyzing shallow water regions but would require other methods to broadly categorize upland and deep open water classes. In our experiments, best classification accuracy of 73% was y ielded when ANN was implemented on normalized images and best detection accuracy of around 92% resulted from applying SVM or SAM to depth invariant images. Indices were developed to categorize broad classes and to detect aquatic vegetation. Index based det ection of aquatic vegetation yielded good accuracy but short of those from non parametric classifiers. We recommend further study with sufficiently large sample size to improve depth invariant transform and to conduct subtle comparison between performances of non parametric classifiers. A B Figure 5 3. Depth invariant transform example. A) Original image ID17 and B ) Depth invariant bands (Y7:Y4:Y1) of the same image.

PAGE 152

152 A B C D Figure 5 4 Classification of Image 17 normalized for w ater surface reflectance using A) ANN, B) SVM C ) ML, and D ) SAM. (Green, dark green and orange color represents aquatic vegetation while cyan and blue colors represent bare bottom and deep water.)

PAGE 153

153 A B C D Figure 5 5 Classification of depth invariant tran sform bands of Image 17 using A ) ANN B ) SVM C) ML, and D ) SAM. (All green and orange color represents aquatic vegetation while cyan and blue colors represent bare bottom and dee p water.)

PAGE 154

154 Figure 5 6. I nd ices based classification of image 17 Emerging vegetation (Green), submerged vegetation (dark green), bare bottom (cyan), deep water or non data (dark gray), upland (maroon).

PAGE 155

155 CHAPTER 6 CONCLUSION S The study investigated and evaluated optimized methods of filtering hyperspectral imagery for signal to noise ratio improvement and analyzing such imagery for water quality parameter estimation and aquatic vegetation classification in case 2 waters. 6.1 Image Filtering High resolution hyperspectral imagery is usually contaminated with non periodic stripes and noise which are difficult to reduce. High contrast edges and relatively smaller image coverage further complicate filtering methods. This research teste d the WFAF (wavelet frequency adaptive filter) de striping algorithm, based on wavelet decomposition and adaptive Fourier transform filtering. The WFAF de striping algorithm is applicable to any pushbroom type image including high spatial resolution hypers pectral image bands tainted with vertical or horizontal stripes Experiments showed that the application of other de striping methods results in over smoothed image and/ or induced artifacts. T he WFAF method decomposes the image into wavelet components T he detail (high frequency) components in the direction of stripes are then filtered in the F ourier domain by adaptively normalizing zero frequency components of vectors (rows or columns). This adaptive nature of filtering subdues artifacts that would other wise be induced due to feature contrasts in the case of high spatial resolution hypersp ectral or pushbroom images. The algorithm requires user input of three parameters wavelet type, number of decomposition level (L) and threshold value (k). Based on res earch results the dB4 wavelet type and k value s from 1 to 1. 5 performed better for various images while L in the range of 2 to 8 is recommended depending

PAGE 156

156 upon the amount of stripes. Comparative results showed the algorithm performed superior to other de striping methods in visual inspection and quantitative assessment yielding greatest improvement in Peak Signal to Noise Ratio improvement ( 23.6 to 27 decibe l). Applying the de striping algorithm on all of the original hyperspectral image bands is time consuming and inefficient as hyperspectral imagery can have numerous bands. Wavelet based de striping methods were tested on selected MNF transformed bands of the image and on original image bands. Applying the de striping method on first several selected MNF bands offered higher efficiency in parameter selection and computation due to r educed number of bands This technique combines wavelet based de striping an d MNF based de noising and reduces both stripes and random noise. Assessment showed that the results of de striping MNF bands were superior to de striping original image bands. Among all tested de striping algorithms, use of the WFAF on MNF bands yielded b est results. In this research, different filtering a lgorithm s w ere tested using only the Daubechies (dB ) wavelet types and the number of decomposition levels was selected manually based on of the amount of stripes in the tested imagery to obtain best resu lts Further study is recommended to evaluate other wavelet types and develop automated methods to determine the number of decomposition levels. 6.2 Hyperspectral I mage A nalysis for A quatic A pplications The m ain focus of th is part of the study is to investigate methods to an a lyz e hyperspectral imagery for water quality parameter estimation and submerged aquatic vegetation classification in case 2 waters. We investigated optimized method for estimating Chlorophyll a and nutrients (N and P) in water from ground based

PAGE 157

157 hyperspectral imagery of eutrophic and hypereutrophic aquaculture ponds using floating and submerged targets Image data was normalized by subtracting average near InfraRed band (926 n m to 958 n m) values and reflectance values were computed using reflectance targets. Using ground based hyperspectral imagery, instead of traditional use of spectro radio meter measurements allowed extract ion of spectral values from different targets and open water and enabled collecting large number of samples ( individual pixels) from single scanned image. Despite high variability in Chl concentrations which ranged from 0.8 to 500 ug/L, model values provided high correlation with concentration s measured in the laboratory Models from water only spectra (without submerged targets) provided high correlation, yielding coefficient of determination (R 2 ) values of 83% for the 2 band model and 86% for the 3 band model Using targets submerged at 10cm depth enhanced the signal and significantly improved the correlation ( R 2 values of 97.5% for the 2 band model and 98.3% for the 3 band model were achieved ) The models were also evaluated for the N itrogen (N) and Phosphorous (P) water quality parameters, and produced moderate correlation. When a single outlier sample was removed from analysis, the R 2 improv ed from 0.478 to 0.823 for P and from 0.51 to 0.68 for N using 3 band model The experiments were conducted on various aquacultural ponds located in one site Further study is needed to expand this method and calibrate t he procedure for Chl a estimation on other types of water bodies and to relate spectral readings with nutrients (N and P) in water. The research investigated the use of high resolution aerial hyperspectral images to detect and classify submerged aquatic ve getation in black riverine water. It evaluated

PAGE 158

158 non parametric classifiers, ANN, SVM and SAM, as well as parametric Maximum L ikelihood classifier as applied to filtered images images normalized for surface water reflectance and depth invariant transformed images. The r esults showed that Maximum L ikelihood classifier generally underperformed all other methods in classification and detection. Non parametric classifiers performed better using normalized or transformed images. Visual inspection showed that ANN and SVM classifiers yielded smooth (less pixelated) classification maps in all regions. The SAM classifier yielded better classification in submerged vegetation region s but classification of upland and open deep water areas was not impeccable and require d other methods to broadly categorize those areas. The ANN method yielded best classification accuracy of 73% when applied to the normalized images. The SVM and SAM classifiers produced best detection accuracy of around 92% when applied to depth invariant transformed image while the ANN classifier yielded detection accuracy of 91% when applied to normalized image. Indices were developed as simple method s for detecting aquatic vegetation and categorizing broad classes. Index based detection results yielded accuracy of 85%, which was better than the Maximum L ikelihood results but short of the best accuracies obtained by other classifiers. This study supports the use of hyperspectral imagery for aquatic vegetation mapping in case 2 waters if images are preprocessed and appropriate classifiers were adopted Experiments were carried out using limited samples in which application of non parametric classifiers on normalized or transformed image s yiel ded higher accuracies. Further study is recommended using sufficiently large sample size to

PAGE 159

159 confirm the accuracies, to improve depth invariant image transform s and to determine subtle difference in performances between non parametric classifiers.

PAGE 160

160 LIST OF REFERENCES Abd Elrahman, A., R. Pande Chhetri, and G. Vallad, 2011. Design and Development of a Multi Purpose Low Cost Hyperspectral Imaging System, Remote Sensing 3(3): 570 586. Adler Golden, S., A. Berk, L. Bernstein, S. Richtsmeier, P. Acharya, M. Matthew, G. Anderson, C. Al lred, L. Jeong, and J. Chetwynd, 1998. FLAASH, a MODTRAN4 atmospheric correction package for hyperspectral data retrievals and simulations, Summaries of the Seventh JPL Airborne Earth Science Workshop JPL Publications 97 21, 1:9 14. Adler Golden, S.M., M.W. Matthew, L.S. Bernstein, R.Y. Levine, A. Berk, S.C. Richtsmeier, P.K. Acharya, G.P. Anderson, G. Felde, J. Gardner, M. Hoke, L.S. Jeong, B. Pukall, J.Mello, A. Ratkowski, and H. H. Burke, 1999. Atmospheric correction for short wave spectral imagery based on MODTRAN 4, SPIE Proceeding, Imaging Spectrometry V 3753: 61 69. Aikio, M., 2001, Hyperspectral prism grating prism imaging spectrograph Espoo 2001 Technical Research Center of Finland, VTT Publications 435 121p. Albert, A., and C. Mobley, 2003. An analytical model for subsurface irradiance and remote sensing reflectance in deep and shallow case 2 waters, Optics Express 11(22): 2873 2890. Albert, A., and P. Gege, 2006. Inversion of irradiance and remote sensing reflecta nce in shallow water between 400 and 800 nm for calculations of water and bottom properties, Applied Optics 45(10): 2331 2343. Anger, C., S. Mah, T. Ivanco, S. Mah R. Price, and J. Busler, 1996. Extended operational capabilities of the casi, Proceeding s of the 2 nd International Airborne Remote Sensing Conference and Exhibition, San Francisco, C alifornia 124 133. Antonini, M., M. Barlaud, P. Mathieu, and I. Daubechies, 1992. Image coding using wavelet transform, IEEE Transactions on Image Processing 1(2): 205 220. Bachmann, R.W., and D.E. Canfield, 1996. Use of an alternative method for monitoring total nitrogen concentrations in Florida lakes, Hydrobiologia 323(1): 1 8. Benediktsson, J.A., P.H. Swain, and O.K. Ersoy, 1990. Neural network approac hes versus statistical methods in classification of multisource remote sensing data, IEEE Transactions on Geoscience and Remote Sensing 28(4): 540 552.

PAGE 161

161 Bertels, L., T. Vanderstraete, S. Van Coillie, E. Knaeps, S. Sterckx, R. Goossens, and B. Deronde, 200 8. Mapping of coral reefs using hyperspectral CASI data; a case study: Fordata, Tanimbar, Indonesia, International Journal of Remote Sensing 29(8): 2359 2391. Bhatti, A.M., D. Rundquist, J. Schalles, L. Ramirez, and S. Nasu, 2009. A comparison between a bove water surface and subsurface spectral reflectances collected over inland waters, Geocarto International 24(2): 133 141. Bierw irth, P., T. Lee, and R. Burne, 1993. Shallow sea floor reflectance and water depth derived by unmixing multispectral imager y, Photogrammetric Engineering and Remote Sensing 59(3): 331 338. Brando, V.E., J.M. Anstee, M. Wettle, A.G. Dekker, S.R. Phinn, and C. Roelfsema, 2009. A physics based retrieval and quality assessment of bathymetry from suboptimal hyperspectral data, Remote Sensing of Environment 113(4): 755 770. Brown, C.D., M.V. Hoyer, R.W. Bachmann, and D.E. Canfield Jr, 2000. Nutrient chlorophyll relationships: an evaluation of empirical nutrient chlorophyll models using Florida and north temperate lake data, Ca nadian Journal of Fisheries and Aquatic Sciences 57(8): 1574 1583. Carlson, R., and J. Simpson, 1996. A Coordinator's Guide to Volunteer Lake Monitoring Methods. North American L ake Management Society, 96p Champagne, C.M., K. Staenz, A. Bannari, H. McNa irn, and J.C. Deguise, 2003. Validation of a hyperspectral curve fitting model for the estimation of plant water content of agricultural canopies, Remote Sensing of Environment 87(2): 148 160. Chander, G., D.L. Helder, and W.C. Boncyk, 2002. Landsat 4/5 band 6 relative radiometry, IEEE Transactions on Geoscience and Remote Sensing 40(1): 206 210. Chang, C.I., and Q. Du, 1999. Interference and noise adjusted principal components analysis, IEEE Transactions on Geoscience and Remote Sensing 37(5): 2387 23 96. Chen, C., 1997. Trends on information processing for remote sensing, Geoscience and Remote Sensing, 1997. IGARSS'97 IEEE International 3: 1190 1192.

PAGE 162

162 Chen, J., H. Lin, Y. Shao, and L. Yang, 2006. Oblique striping removal in remote sensing imagery ba sed on wavelet transform, International Journal of Remote Sensing 27(8): 1717 1723. Chen, J., Y. Shao, H. Guo, W. Wang, and B. Zhu, 2003. Destriping CMODIS data by power filtering, IEEE Transactions on Geoscience and Remote Sensing 41(9): 2119 2124. Ci pollini, P., G. Corsini, M. Diani, and R. Grasso, 2001. Retrieval of sea water optically active parameters from hyperspectral data by means of generalized radial basis function neural networks, IEEE Transactions on Geoscience and Remote Sensing 39(7): 150 8 1524. Cocks, T., R. Jenssen, A. Stewart, I. Wilson, and T. Shields, 1998. The HyMapTM airborne hyperspectral sensor: the system, calibration and performance, Proceedings of the 1st EARSeL workshop on Imaging Spectroscopy 37 42. Corsini, G., M. Diani, R. Grasso, M. De Martino, P. Mantero, and S. Serpico, 2003. Radial Basis Function and Multilayer Perceptron neural networks for sea water optically active parameter estimation in case II waters: a comparison, International Journal of Remote Sensing 24(20) : 3917 3932. Cortes, C., and V. Vapnik, 1995. Support vector networks, Machine Learning 20(3): 273 297. C ox C., and W. Munk, 1954. Measurement of the roughness of the sea surface from photographs of the sun's glitter, Journal of the Optical Society of America, 44(11): 834 850. Crumpton, W.G., T.M. Isenhart, and P.D. Mitchell, 1992. Nitrate and organic N analyses with second derivative spectroscopy, Limnology and Oceanography 907 913. Dall'Olmo, G., and A.A. Gitelson, 2005. Effect of bio optical param eter variability on the remote estimation of chlorophyll a concentration in turbid productive waters: experimental results, Applied Optics 44(3): 412 422. Daubechies, I., 1988. Orthonormal bases of compactly supported wavelets, Communications on pure and applied mathematics 41(7): 909 996. Nenad, H. Hoogenboom, and S. Peters, 1996. Remote sensing, ecological water quality modelling and in situ measurements: a case study in shallow lakes, Hydrological sciences journal 41(4): 531 547.

PAGE 163

163 D'Elia, C.F., P.A. Steudler, and N. Corwin, 1977. Determination of total nitrogen in aqueous samples using persulfate digestion, Limnology and Oceanography 760 764. Donoho, D.L., and J.M. Johnstone, 1994. Ideal spatial adaptation by wavelet shrink age, Biometrika 81(3): 425 455. Doxaran, D., R.C.N. Cherukuru, and S. Lavender, 2004. Estimation of surface reflection effects on upwelling radiance field measurements in turbid waters, Journal of Optics A: Pure and Applied Optics 6(7): 690. Eskicioglu A.M., and P.S. Fisher, 1995. Image quality measures and their performance, IEEE Transactions on Communications 43(12): 2959 2965. Fagerbakke, K.M., M. Heldal, and S. Norland, 1996. Content of carbon, nitrogen, oxygen, sulfur and phosphorus in native aq uatic and cultured bacteria, Aquatic Microbial Ecology 10(1): 15 27. Fukuda, S., and H. Hirosawa, 1999. Smoothing effect of wavelet based speckle filtering: The Haar basis case, IEEE Transactions on Geoscience and Remote Sensing 37(2): 1168 1172. Gadal lah, F., F. Csillag, and E. Smith, 2000. Destriping multisensor imagery with moment matching, International Journal of Remote Sensing 21(12): 2505 2511. Gallegos, C.L., 2005. Optical water quality of a blackwater river estuary: the Lower St. Johns River, Florida, USA, Estuarine, Coastal and Shelf Science 63(1 2): 57 72. Gege, P., and A. Albert, 2006. A tool for inverse modeling of spectral measurements in deep and shallow waters, Remote sensing of aquatic coastal ecosystem processes 81 109. Gersman, R., E. Ben Dor, M. Beyth, and D. Avigad, M. Beyth, M. Abraha, and A. Kibreab, 2006. Hyperspectral remote sensing as a tool for geological exploration examples from the northern Danakil depression, Eritrea, The Geological Survey of Israel, Report G SI/08/06, Jerusalem. Gitelson, A.A., D. Gurlin, W.J. Moses, and T. Barrow, 2009. A bio optical algorithm for the remote estimation of the chlorophyll a concentration in case 2 waters, Environmental Research Letters 4(4): 045003.

PAGE 164

164 Gitelson, A.A., G. Dall 'Olmo, W. Moses, D.C. Rundquist, T. Barrow, T.R. Fisher, D. Gurlin, and J. Holz, 2008. A simple semi analytical model for remote estimation of chlorophyll a in turbid waters: Validation, Remote Sensing of Environment 112(9): 3582 3593. Gitelson, A.A., J. F. Schalles, and C.M. Hladik 2007. Remote chlorophyll a retrieval in turbid, productive estuaries: Chesapeake Bay case study, Remote Sensing of Environment 109(4): 464 472. Gons, H.J., 1999. Optic al teledetection of chlorophyll a in turbid inland waters Environmental science & technology 33(7): 1127 1132. Gonzalez, R.C., and R.E. Woods, 2008. Digital Image Processing third ed Prentice Hall NJ, USA Goodman, J.A., Z.P. Lee, and S.L. Ustin, 2008. Influence of atmospheric and sea surface corrections on retrieval of bottom depth and reflectance using a semi analytical model: a case study in Kaneohe Bay, Hawaii, Applied Optics 47(28): F1 F11. Green, A.A., M. Berman, P. Switzer, and M.D. Craig, 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(1): 65 74. Gualtieri, J.A., and S. Chettri, 2000. Support vector machines for classification of hyperspectral data, Geoscience and Remote S ensing Symposium, 2000. Proceedings of IGARSS 2000 IEEE 2000 International 2: 813 815 Hedley, J., A. Harborne, and P. Mumby, 2005. Simple and robust removal of sun glint for mapping shallow water benthos, International Journal of Remote Sensing 26(10) : 2107 2112. Hedley, J., C. Roelfsema, and S.R. Phinn, 2009. Efficient radiative transfer model inversion for remote sensing applications, Remote Sensing of Environment 113(11): 2527 2532. Heermann, P.D., and N. Khazenie, 1992. Classification of multisp ectral remote sensing data using a back propagation neural network, IEEE Transactions on Geoscience and Remote Sensing 30(1): 81 88.

PAGE 165

165 Hochberg, E.J., S. Andrefouet, and M.R. Tyler, 2003. Sea surface correction of high spatial resolution Ikonos images to improve bottom mapping in near shore environments, IEEE Transactions on Geoscience and Remote Sensing 41(7): 1724 1729. Horn, B.K.P., and R.J. Woodham, 1979. Destriping Landsat MSS images by histogram modification, Computer Graphics and Image Processing 10(1): 69 83. Hoyer, M.V., C.A. Horsburgh, D.E. Canfield Jr, and R.W. Bachmann, 2005. Lake level and trophic state variables among a population of shallow Florida lakes and within individual lakes, Canadian Journal of Fisheries and Aquatic Sciences 62( 12): 2760 2769. Huynh Thu, Q., and M. Ghanbari, 2008. Scope of validity of PSNR in image/video quality assessment, Electronics Letters 44(13): 800 801. International Ocean Color Coordinating Group, 2000. Remote sensing of ocean colour in coastal, and o ther optically complex waters. In: Sathyendranath, S. (ed.), Reports of the International Ocean Colour Coordinating Group No. 3, IOCCG, Dartmouth, Canada. International Ocean Color Coordinating Group, 2009. Remote sensing in fisheries and aquaculture. In: Forget, M. H., V. Stuart, T. Platt (eds.), Reports of the International Ocean Colour Coordinating Group No. 8, IOCCG, Dartmouth, Canada. Isaaks, E., and R. Srivastava. 1989. An introduction to applied geostatistics Oxford University Press, New York Je nsen, J.R., 2004. Introductory digital image processing Prentice Hall, Upper Saddle River, NJ, USA. Jensen, R.R., P.J. Hardin, and G. Yu, 2009. Artificial Neural Networks and Remote Sensing, Geography Compass 3(2): 630 646. Jiao, H., Y. Zha, J. Gao, Y. Li, Y. Wei, and J. Huang, 2006. Estimation of chlorophyll a concentration in Lake Tai, China using in situ hyperspectral data, International Journal of Remote Sensing 27(19): 4267 4276. Kamarainen, A.M., R.M. Penczykowski, M.C. Van de Bogert, P.C. Hans on, and S.R. Carpenter, 2009. Phosphorus sources and demand during summer in a eutrophic lake, Aquatic Sciences Research across Boundaries 71(2): 214 227.

PAGE 166

166 Keiner, L.E., and X.H. Yan, 1998. A neural network model for estimating sea surface chlorophyll and sediments from thematic mapper imagery, Remote Sensing of Environment 66(2): 153 165. Klonowski, W.M., P.R. Fearns, and M.J. Lynch, 2007. Retrieving key benthic cover types and bathymetry from hyperspectral imagery, Jo urnal of Applied Remote Sensing 1( 1): 011505. Kruse, F., A. Lefkoff, J. Boardman, K. Heidebrecht, A. Shapiro, P. Barloon, and A. Goetz, 1993. The spectral image processing system (SIPS) interactive visualization and analysis of imaging spectrometer data, Remote Sensing of Environment 44(2 3): 145 163. Kutser T., Vahtmae E., and Praks J., 2009. A sun glint correction method for hyperspectral imagery containing areas with non negligible water leaving NIR signal, Remote Sensing of Environment 113(10): 2267 2274. Lee, J.B., A.S. Woody att, and M. Berman, 1990. Enhancement of high spectral resolution remote sensing data by a noise adjusted principal components transform, IEEE Transactions on Geoscience and Remote Sensing 28(3): 295 304. Lee, J.S., 1983. Digital image smoothing and the sigma filter, Computer Vision, Graphics, and Image Processing 24(2): 255 269. Lee, Z., K.L. Carder, C.D. Mobley, R.G. Steward, and J.S. Patch, 1998. Hyperspectral remote sensing for shallow waters. I. A semianalytical model, Applied Optics 37(27): 6329 6338. Lee, Z., K.L. Carder, R.F. Chen, and T.G. Peacock, 2001. Properties of the water column and bottom derived from Airborne Visible Infrared Imaging Spectrometer (AVIRIS) data, Journal of Geophysical Research 106(C6): 11639 11651 Lee, Z.P., and C. H u, 2006. Global distribution of Case 1 waters: An analysis from SeaWiFS measurements, Remote Sensing of Environment 101(2): 270 276. Lillesand, T.M., and R.W. Kiefer, 1987. Remote sensing and image interpretation John Wiley & Sons, New York Liu, J.G., and G.L.K. Morgan, 2006. FFT selective and adaptive filtering for removal of systematic noise in ETM imageodesy images, IEEE Transactions on Geoscience and Remote Sensing 44(12): 3716 3724.

PAGE 167

167 Liu, X., B. Zhang, L.R. Gao, and D.M. Chen, 2009. A maximum no ise fraction transform with improved noise estimation for hyperspectral images, Science in China Series F: Information Sciences 52(9): 1578 1587. Lu, X., R. Liu, J. Liu, and S. Liang, 2007. Removal of noise by wavelet method to generate high quality temp oral data of terrestrial MODIS products, Photogrammetric Engineering and Remote Sensing 73(10): 1129 1139 Lyzenga, D.R., 1978. Passive remote sensing techniques for mapping water depth and bottom features, Applied Optics 17(3): 379 383. Lyzenga, D.R., 1981. Remote sensing of bottom reflectance and water attenuation parameters in shallow water using aircraft and Landsat data, International Journal of Remote Sensing 2(1): 71 82. Lyzenga, D.R., 1985. Shallow water bathymetry using combined lidar and pa ssive multispectral scanner data, International Journal of Remote Sensing 6(1): 115 125. Lyzenga, D.R., N.P. Malinas, and F.J. Tanis, 2006. Multispectral bathymetry using a simple physically based algorithm, IEEE Transactions on Geoscience and Remote Sen sing 44(8): 2251 2259. Mallat, S.G., 1989. A theory for multiresolution signal decomposition: the wavelet representation, IEEE Transactions on Pattern Analysis and Machine Intelligence 11(7): 674 693. Martens, H.A., and P. Dardenne, 1998. Validation an d verification of regression in small data sets, Chemometrics and Intelligent Laboratory Systems 44(1): 99 121. Mazumder, A., and K.E. Havens, 1998. Nutrient chlorophyll Secchi relationships under contrasting grazer communities of temperate versus subtro pical lakes, Canadian Journal of Fisheries and Aquatic Sciences 55(7): 1652 1662. McIntyre, M.L., D.F. Naar, K.L. Carder, B.T. Donahue, and D.J. Mallinson, 2006. Coastal bathymetry from hyperspectral remote sensing data: comparisons with high resolution multibeam bathymetry, Marine Geophysical Research 27(2): 129 136. Meaden, G.J., and J.M. Kapetsky, 1992, Geographical information systems and remote sensing in inland fisheries and aquaculture FAO Fisheries Technical Paper Roma, Italy, 318: 286p.

PAGE 168

168 Melack, J.M., and M. Gastil, 1994. Comparison of spectral feature algorithms for remote sensing of chlorophyll in eutrophic lakes, Geoscience and Remote Sensing Symposium, 1994. IGARSS'94. Surface and Atmospheric Remote Sensing: Technologies, Data Analysis and Interpretation, International 4: 2363 2365. Melgani, F., and L. Bruzzone, 2004. Classification of hyperspectral remote sensing images with support vector machines, IEEE Transactions on Geoscience and Remote Sensing 42(8): 1778 1790. Menzel, D.W., and N. Corwin, 1965. The measurement of total phosphorus in seawater based on the liberation of organically bound fractions by persulfate oxidation, Limnology and Oceanography 10(2): 280 282. Mittenzwey, K.H., S. Ullrich, A. Gitelson, and K. Kondratiev, 1992. Determination of chlorophyll a of inland waters on the basis of spectral reflectance, Limnology and Oceanography 147 149. Mobley, C.D., 1999. Estimation of the remote sensing reflectance from above surface measurements, Applied Optics 38(36): 7442 7455. Mobley, C.D., L.K. Sundman, C.O. Davis, J.H. Bowles, T.V. Downes, R.A. Leathers, M.J. Montes, W.P. Bissett, D.D.R. Kohler, and R.P. Reid, 2005. Interpretation of hyperspectral remote sensing imagery by spectrum matching and look up tables, Applied Optics 44(17): 3576 3592. Morel, A., 1988. Optical modeling of the upper ocean in relation to its biogenous matter content (case I waters), Journal of Geophysical Research 93(10): 749 710. Morel, A., and L. Prieur, 1977. Analysis of variations in ocean color, Limnology and Oceanography 709 722. Mueller, J.L., R.W. Austin, E.R. Firestone, and J.G. Acker, 1995, SeaWiFS Technical Report Series: Ocean Optics Protocols for Sea WiFS Validation NASA Technical Memorandum 104566 NASA Goddard Space Fligh t Center, Greenbelt, Maryland. Mnch, B., P. Trtik, F. Marone, and M. Stampanoni, 2009. Stripe and ring artifact removal with combined wavelet Fourier filtering, Optical Express 17(10): 8567 8591. Murphy, J., and J.P. Riley, 1962. A modified single solu tion method for the determination of phosphorus in natural waters. Anal. Chim. Acta 27: 12 26.

PAGE 169

169 Nagao, M., and T. Matsuyama, 1979. Edge preserving smoothing, Compute r graphics and image processing 9(4): 394 407. Oimoen, M.J., 2000. An effective filter for removal of production artifacts in US Geological Survey 7.5 minute digital elevation models, Proceedings of the Fourteenth International Conference on Applied Geologic Remote Sensing Las Vegas, Nevada, 6 8. Pal, M., and G.M. Foody, 2010. Feature selecti on for classification of hyperspectral data by SVM, IEEE Transactions on Geoscience and Remote Sensing 48(5): 2297 2307. Pande Chhetri, R., and A. Abd Elrahman, 2011. De striping hyperspectral imagery using wavelet transform and adaptive frequency domain filtering, International Journal of Ph otogrammetry and Remote Sensing 66: 620 636. Peng, C.K., S. Havlin, J. Hausdorff, J. Mietus, H. Stanley, and A. Goldberger, 1995. Fractal mechanisms and heart rate dynamics: long range correlations and their breakdo wn with disease, Journal of electrocardiology 28: 59 65. Richards, J., and X. Jia 1999 Remote Sensing Digital Image Analysis: An Introduction 5 th ed. Springer Rodgers, J.H. 2008, Algal toxins in pond aquaculture Southern Regional Aquaculture Center (U.S.), 8p Ruddick, K.G., H.J. Gons, M. Rijkeboer, and G. Tilstone, 2001. Optical remote sensing of chlorophyll a in case 2 waters by use of an adaptive two band algorithm with optimal error properties, Applied Optics 40(21): 3575 3585. Rundqui st, D.C., J.F. Schalles, and J.S. Peake, 1995. The response of volume reflectance to manipulated algal concentrations above bright and dark bottoms at various depths in an experimental pool, Geocarto International 10(4): 5 14. S. Haykin, 1994. Neural Net works: A Comprehensive Foundation, Macmillan College Publishing Company : Englewood Cliffs. Sartory, D., and J. Grobbelaar, 1984. Extraction of chlorophyll a from freshwater phytoplankton for spectrophotometric analysis, Hydrobiologia 114(3): 177 187.

PAGE 170

170 Sc halles, J.F., 2006. Optical remote sensing techniques to estimate phytoplankton chlorophyll a concentrations in coastal waters with varying suspended matter and CDOM concentrations In: L. Richardson and E. Ledrew, eds., Remote sensing of aqua tic coastal e cosystem processes, Springer, Dordrecht, Netherlands, 27 79. Schalles, J.F., A.A. Gitelson, Y.Z. Yacobi, and A.E. Kroenke, 1998. Estimation of chlorophyll a from time series measurements of high spectral resolution reflectance in an eutrophic lake, Journa l of Phycology 34(2): 383 390. Schmidt, K., and A. Skidmore, 2003. Spectral discrimination of vegetation types in a coastal wetland, Remote Sensing of Environment 85(1): 92 108. Schowengerdt, R.A., 2007. Remote Sensing: Models and Methods for Image Pro cessing, third ed. Academic Press pp. 229 354. Schweizer, D., R. Armstrong, and J. Posada, 2005. Remote sensing characterization of benthic habitats and submerged vegetation biomass in Los Roques Archipelago National Park, Venezuela, International Journal of Remote Sensing 26(12): 2657 2667. Simal, J., M. Lage, and I. Iglesias, 1985. Second derivative ultraviolet spectroscopy and sulfamic acid method for determination of nitrates in water, Journal of the Association of Official Analytical Chemists JANCA 68: 962 964 Singh, A., 1985. Postlaunch corrections for Thematic Mapper 5 (TM 5) radiometry in the Thematic Mapper image processing System (TIPS), Photogrammetric Engineering and Remote Sensing 51: 1385 1390. Srinivasan, R., M. Cannon, and J. White, 1988. Landsat data destriping using power spectral filtering, Optical Engineering 27(11): 939 943. Stumpf, R.P., K. Holderied, and M. Sinclair, 2003. Determination of water depth with high resolution satellite imagery over variable bottom types, Limnolo gy and Oceanography, 547 556. Sudduth, K.A., G.S. Jang, R.N. Lerch, and E.J. Sadler, 2005. Estimating Water Quality with Airborne and Ground Based Hyperspectral Sensing Proceedings of the ASAE Annual International Meeting, No. 052006 Sun, D., Y. Li, an d Q. Wang, 2009. A unified model for remotely estimating chlorophyll a in Lake Taihu, China, based on SVM and in situ hyperspectral data, IEEE Transactions on Geoscience and Remote Sensing 47(8): 2957 2965.

PAGE 171

171 Switzer, P., and A.A. Green, 1984. Min/max autocorrelation factors for multivariate spatial imagery, Technical Report No. 6, Department of Statistics, Stanford University, p.10 Tabinda, A.B., and M. Ayub, 2010. Effect of high phosphate fertilization rate on pond phosphate concentrations, chloroph yll a, and fish growth in carp polyculture, Aquaculture International 18(3): 285 301. Thiemann, S., and H. Kaufmann, 2000. Determination of chlorophyll content and trophic state of lakes using field spectrometer and IRS 1C satellite data in the Mecklenbu rg Lake District, Germany, Remote Sensing of Environment 73(2): 227 235. Tobler, W.R., 1970. A computer movie simulating urban growth in the Detroit region, Economic geography 46: 234 240. Torres, J., and S.O. Infante, 2001. Wavelet analysis for the el imination of striping noise in satellite images, Optical Engineering 40: 1309. V an der Meer, F., 2006. The effectiveness of spectral similarity measures for the analysis of hyperspectral imagery, International journal of applied earth observation and geo information 8(1): 3 17. Varotsos, C., J. Ondov, and M. Efstathiou, 2005. Scaling properties of air pollution in Athens, Greece and Baltimore, Maryland, Atmospheric Environment 39(22): 4041 4047. Wang, L., J.J. Qu, X. Xiong, X. Hao, Y. Xie, and N. Che 2006. A new method for retrieving band 6 of Aqua MODIS, Geoscience and Remote Sensing Letters, IEEE 3(2): 267 270. Wang, Z., and Y. Fu, 2007. Frequency domain regularized deconvolution for images with stripe noise, Fourth International Conference on Image and Graphics ICIG 2007 110 115. Wegener, M., 1990. Destriping multiple sensor imagery by improved histogram matching, Internat ional Journal of Remote Sensing 11(5): 859 875. Weinreb, M., R. Xie, J. Lienesch, and D. Crosby, 1989. Destriping GOES images by matching empirical distribution functions, Remote Sensing of Environment 29(2): 185 195.

PAGE 172

172 Wollin, K.M., 1987. Nitrate determination in surface waters as an example of the application of UV derivative spectro scopy to environmental analysis, Acta Hydrochimica et Hydrobiologica 15(5): 459 469. Xu, J.P., F. Li, B. Zhang, K.S. Song, Z.M. Wang, D.W. Liu, and G.X. Zhang, 2009. Estimation of chlorophyll a concentration using field spectral data: a case study in inland Case II waters, North China, Envi ronmental monitoring and assessment 158(1): 105 116. Ye, X., K. Sakai, L.O. Garciano, S.I. Asada, and A. Sasao, 2006. Estimation of citrus yield from airborne hyperspectral images using a neural network model, Ecological Modelling 198(3): 426 432. Zim ba, P.V., and A. Gitelson, 2006. Remote estimation of chlorophyll concentration in hyper eutrophic aquatic systems: Model tuning and accuracy optimization, Aquaculture 256(1): 272 286.

PAGE 173

173 BIOGRAPHICAL SKETCH Roshan Pande Chhetri received b degre e in c ivil e ngineering in 1996 and MS in Urban Planning in 2003 from Tribhuvan University Nepal. He earned his MS in GIS Planning from Eastern Michigan University Michigan in 200 7 He worked in a position of GIS Supervisor for Geospatial Systems P. Ltd., a commercial GIS and p hotogrammetry company in Nepal, from 1997 to 2005. He also worked as a GIS System Analyst in Alachua County Department of Growth Management in 2008 before joining University of Florida. In 2009, he joi ned PhD program in Geomatics concentration in the University of Florida School of Forest Resources and Conservation and worked as a Graduate Research Assistant u nder the supervision of Dr. Amr Abd Elrahman H e was involved in remote sensing research projec ts and focused his research in h yperspectral imagery. He has published three journal articles and presented his research works at ASPRS conferences numerous times. He is a recipient of C ertificate of O utstanding A cademic A chievement from UF College of Agr icultural and Life Science and 2012 O utstanding Doctoral Student of the Y ear from School o f Forest Resourc es and Conservation.