This item is only available as the following downloads:
1 CLASSIFICATION OF MA RL PRAIRIE AND MARSH VEGETATION COMMUNITI ES IN THE EVERGLADES NATIONAL PARK By ZOLTAN SZANTOI A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE R EQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2011
2 2011 Zoltan Szantoi
3 To my family
4 ACKNOWLEDGMENTS I would like to thank Dr. Scot Smith and the faculty of the School of Forest Conservation and Resources for the support and guidance during this endeavor. I also thank Dr. Francisco Escobedo Dr. Amr Abd Elrahman, Dr. Bon Dewitt, Dr. Franklin Percival and Dr. Leonard Pearlstine for serving o n my committee and providing valuable assistance and encouragement throu ghout the entire process. I would like to thank my parents and family for the support they have given for the past time, and most importantly to my wonderful wife, Joysee and son Zolta for their love and support throughout the studies at the University of Florida I would also like to thank my friends and colleagues for the great time we spent together.
5 TABLE OF CONTENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ .. 4 LIST OF TABLES ................................ ................................ ................................ ............ 7 LIST OF FIGURES ................................ ................................ ................................ .......... 9 LIST OF ABBREVIATIONS ................................ ................................ ........................... 10 ABSTRACT ................................ ................................ ................................ ................... 12 CHAPTER 1 INTRODUCTION ................................ ................................ ................................ .... 14 Objectives ................................ ................................ ................................ ............... 19 Significance of the Study ................................ ................................ ........................ 20 2 LITERATURE REVIEW ................................ ................................ .......................... 22 Remote Sensing of Wetlands ................................ ................................ ................. 22 Textural Features ................................ ................................ ................................ .... 23 Photo Interpretation ................................ ................................ ................................ 25 Automated Image Classification Techniques ................................ .......................... 26 Hypothesis ................................ ................................ ................................ .............. 29 Study Area ................................ ................................ ................................ .............. 30 3 WETLAND COMPOSITION ANALYSIS USING HIGH RESOLUTION IMAGES AND TEXTURE FEATURES ................................ ................................ .................. 32 Introduction ................................ ................................ ................................ ............. 32 Methods ................................ ................................ ................................ .................. 37 Classification Scheme ................................ ................................ ...................... 37 Study Area and Field Data Collection ................................ ............................... 38 Imagery ................................ ................................ ................................ ............ 39 Development of Texture Features ................................ ................................ .... 40 Separability Analysis ................................ ................................ ........................ 42 Parametric Classification ................................ ................................ .................. 43 Accuracy Assessment ................................ ................................ ...................... 43 Evaluation of the Classification Results ................................ ............................ 45 Results ................................ ................................ ................................ .................... 45 Semivariogram Analysis ................................ ................................ ................... 45 Spectral Separability ................................ ................................ ........................ 46 Using Spectral Bands and Normalized Difference Vegetation Index ................ 47
6 Incorporation of First Order Texture Features ................................ .................. 47 Incorporation of Second Order Texture Features ................................ ............. 48 Merging Different Window Size s ................................ ................................ ...... 48 Discussion ................................ ................................ ................................ .............. 50 4 CLASSIFICATION OF WETLAND COMMUNITIES: USING HIGH RESOLUTION AERIAL IMAGERY AND TESTING ALGORYTHMS ...................... 69 Introduction ................................ ................................ ................................ ............. 69 Methods ................................ ................................ ................................ .................. 75 Study Area ................................ ................................ ................................ ........ 75 Classification System and Training Data ................................ .......................... 76 Imagery ................................ ................................ ................................ ............ 77 Texture Features ................................ ................................ .............................. 78 Classification of Vegetation Communities ................................ ........................ 78 Decision tree ................................ ................................ .............................. 79 Artificial neural network ................................ ................................ .............. 80 Evaluation of Classification Accuracy and Algorithm Comparison .................... 81 Results ................................ ................................ ................................ .................... 82 Discuss ion ................................ ................................ ................................ .............. 85 5 CONCLUSIONS ................................ ................................ ................................ ..... 99 LIST OF REFERENCES ................................ ................................ ............................. 102 BIOGRAPHICAL SKETCH ................................ ................................ .......................... 113
7 LIST OF TABLES Table page 3 1 Major plant communities in the study area modified from Ross et al. (2003 and 2006) and Richardson et al. (2008). ................................ ............................ 62 3 2 Description of collected ground reference data. ................................ ................. 62 3 3 Aerial imagery data characteristics. ................................ ................................ .... 63 3 4 Texture feature calculation formulae. ................................ ................................ 64 3 5 Jeffries Matsushita (J M) separability analysis using the 4 spectral bands. ....... 65 3 6 Jeffries Matsushita (J M) separability analysis using the 4 spectral bands, first and second order texture features. ................................ .............................. 65 3 7 Maximum likelihood classificati on results using spectral bands and Normalized Difference Vegetation Index (NDVI) layer. ................................ ....... 65 3 8 Maximum likelihood classification results for spectral bands and first order texture features wi th different window sizes. ................................ ...................... 66 3 9 Maximum likelihood classification results for spectral bands and first order texture features with different window sizes. ................................ ...................... 66 3 10 Maximum likelihood classification results for spectral bands and second order texture features with small Grey Level Co occurrence Matrix. .................. 67 3 11 Maximum likelihood classification results for spectral bands and second order texture features with large Grey Level Co occurrence Matrix (11 by 11 and 15 by 15). ................................ ................................ ................................ ..... 67 3 12 Maximum likelihood classifica tion results for spectral bands and first order texture features with merged window sizes. ................................ ....................... 68 3 13 Maximum likelihood classification results for spectral bands and first and second order textu re features with merged window sizes. ................................ 68 4 1 Major plant communities in the research area. ................................ ................... 93 4 2 Description of collected groun d reference data. ................................ ................. 93 4 3 UltracamX imagery data characteristics. ................................ ............................ 94 4 4 Classification results based on spectral bands and the N DVI layer. ................... 94
8 4 5 Spectral bands, NDVI and first order texture features (9 pixels by 9 pixels moving window) classification results. ................................ ................................ 95 4 6 Spectral bands and NDVI and second order texture features (7 pixels by 7 pixels GLCM) classification results. ................................ ................................ .... 95 4 7 Spectral bands and NDVI and first order texture features (3 pix els by 3 pixels variance and 9 pixels by 9 pixels data range, mean, entropy moving window) classification results. ................................ ................................ ........................... 96 4 8 Spectral bands, NDVI and second order texture features merged 3 pixels by 3 pixels and 9 pixels by 9 pixels GLCM window classification results. ............... 96 4 9 Spectral bands, NDVI and first and second order texture features 3 pixels by 3 pixels and 9 pixels by 9 pixels moving window and GLCM classification results. ................................ ................................ ................................ ................ 97 4 10 Kappa confidences and significance testing based on the z score for maps using various combinations of input layers and classifiers. ................................ 98
9 LIST OF FIGURES Figure page 2 1 Study area within the Everglades National Park ................................ ................. 31 3 1 Lo cation of study area in the Everglades National Park based on a Landsat Thematic Mapper 5 image in true color composite. ................................ ........... 55 3 2 Microsoft Vexcel UltacamX mosaiced image covering the study area in the ENP. ................................ ................................ ................................ ................... 56 3 3 Spectral response curve for the studied vegetation communities in the Blue. Green, Red and Infrared spectra l bands. ................................ .......................... 57 3 4 Selected training sites for semivariance calculation. ................................ ......... 58 3 5 Semivariograms derived from four spectral bands for the selected trai ning sites. ................................ ................................ ................................ ................... 59 3 6 Land cover map based on spectral bands, N ormalized Difference Vegetation Index layer and first order texture features (3 by 3 window variance and 9 by 9 window data range, mean, entropy).. ................................ .............................. 60 3 7 Land cover map based on spectral bands, NDVI layer and second order texture features (3 by 3 and 9 by 9 window). ................................ ...................... 61 4 1 UltacamX mosaiced image of the study area i n the Everglades National Park. 90 4 2 Example images of the classification classes. ................................ .................... 91 4 3 Plant community maps of the study area i n the Everglades National Park using different classification algorithms. ................................ ............................. 92
10 LIST OF ABBREVIATION S ANN Artificial Neural Network ANOVA A nalysis of V ariance ASM A ngular S econd M oment AVHRR Advanced Very High Reso lution Radiometer BP B ack P ropagation B&W B lack and W hite CASI C ompact A irborne S pectrographic I mager CART Classification and Regression Trees CHAID Chi squared Automatic Interaction Detector CCD C harge C oupled D evice CERP Comprehensive Evergl ades Restoration Plan CIR C olor I nfrared CM Tall Sawgrass Marsh CWP Short Sawgrass P rairie DEM D igital E levation M odel DT Decision T ree EAA Everglades Agricultural Area EDEN Everglades D epth E stimation N etwork ENP Everglades National Par k EVCS Everglades Vegetation Classification System ETM Enhanced Thematic Mapper GIS G eographic I nformation S ystem GLCM Grey L evel C o occurrence M atrix
11 GPS G lobal P ositioning S ystem HARN High Accuracy Reference Network ISODATA Iterative Self Organizing Data Analysis J M Jeffries Matsushita separability measure ML Maximum L ikelihood MLP M ulti L ayer P erceptron MMU M inimum M apping U nit MWP Muhlenbergia Wet P rairie NAVD88 North American Vertical Datum of 1988 NIR N ear I nfrared RMSE R oot M ean S quare E rror QDA Q uadratic D iscriminant A nalysis QUEST Quick, Unbiased, Efficient Statistical Tree SAR Syntetic Aperture Radar SCWP Schizachyrium Wet P rairie S B Spectral Bands SFWMD South Florida Water Management District VCS Veget ation Classification System for South Florida National Parks WP W et P rairie Woody Woody V egetation
12 Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doct or of Philosophy CLASSIFICATION OF MA RL PRAIRIE AND MARSH VEGETATION COMMUNITI ES IN THE EVERGLADES NATIO NAL PARK By Zoltan Szantoi August 2011 Chair: Scot Smith Cochair: Francisco Escobedo Major: Forest Resources and Conservation Th is study develop ed a classification and discrimination methodolog y with quantifiable accuracies for wetland mapping focusing on at risk plant communities in marl prairie and marsh areas of the Everglades National Park, Florida Accurately and rapidly mapping the baseline spa tial distribution of these communities with classification algorithms using accessible and temporally frequent imagery is critical t o better assess changes in wet graminoid/sedge communities and their effects on the ecotone wildlife habitats and overall e cological change. Maximum likelihood (ML) Decision T ree (DT) and Artificial Neural N etwork (ANN) classifiers were tested on high resolution aerial imagery (30.5 cm) of the Everglades National Park, Florida. Several parameters in the classification models were tested to improve mapping accuracy such as spectral bands, normalized difference vegetation index (NDVI) and first and second order texture features derived from the near infrared band. The proper window size of the different texture features w as esti mated using the semivariogram method. The addition of NDVI and the texture features to the classification models yielded an increase in classification accuracy, especially i n differentiating between wetland plant communities. Comparison
13 among the classifie rs was based on the calculated Kappa coefficient of the developed maps. The results indicated that ANN classifier produced a statistically significantly higher accuracy (89.49%) than the DT (86.89%) or ML (84.92%) classifiers ( <0.05) For delineating and mapping wetland plant communities, first and second order texture features derived from high resolution imagery improve d the classification. Additionally, the utilization of machine learning algorithms (DT and ANN) in the class ification process further increased overall accuracy of the developed maps. F indings from this study can be used to monitor vegetation communities in an accurate and efficient manner on a more frequent basis.
14 CHAPTER 1 INTRODUCTION Wetlands are an essen tial part of the global ecosystem (Mitsch and Gosselink 2007) and provide habitat for endangered species, supply groundwater aquifers and have been reported to reduce floods (Nicholls et al. 1999; Winter 1999; Zedler 2000, 2003; Brinson and Malvrez 2002; Bullock and Acreman 2003; Hancock et al. 2005; Mitsch and Gosselink 2007) The Everglades is the only protected subtropical wetland ecosystem in the U.S and is located at the southern end of peninsular Florida covering about 10,100 km 2 from which 6108 km 2 is protected (Junk et al. 2006; Davis and Ogden 1994) The Everglade s National Park is listed in the Ramsar Convention of Wetlands as being of International Importance (Richardson 2010) as well as in the UNESCO List of the World Heritage Sites in Danger (O gden 2004) a result of groundwater interactions and flow regimes (Harvey and McCormick 2009) The park has the largest mangrove forest in the western hemisphere and is the most significant breeding terrain for tropical wa ding birds as well as containing the largest continuous stand of sawgrass in North America (Ogden 2004) Because of its size location and climate, unique flora and fauna, hydrological functions and geological history the Everglades has beco me one of the m ost studied ecosystems. According to Web of Knowledge the Everglades is the most extensively studied and globally important wetland. A literature search in April, 2011 identified 753 (66 in 2010) scientific papers for the E Botswana
15 numerous issues such as sea level ri se, sedimentation, chemical pollution and human disturbance, climate change (Titus and Richman 2001; Ogden 2004; Pearlstine et al. 2010; Larsen et al. 2011) along with endangered species it provides habitat to the Florida panther ( Puma concolor coryi ), Ame rican crocodile ( Crocodylus acutus ) and Cape Sable seaside sparrow ( Ammodramus maritimus mirabilis ) (Walker and Solecki 2004) decreased by 90 to 95 percent since the pre drainage era (Ogden 2004) and overall 68 native plant and animals found in the ENP are listed as endangered or threatened (Ogden 2004) The Biscayne aquifer which is located beneath the Everglades provides about 33% of the total public supply groundwater in Florid a and serves 4.48 million people (Marella 2009) Miami Dade C ounty itself withdr a w s all of its drinking water from the Biscayne aquifer which is more than 0.75 million cubic meters per day in 2005 (Marella 2009) In recent years, in addition to water explo itation and land reclamation for urban development the change to the hydrograph and drainage canal construction combined with intense agriculture has led to increased disturbance, eutrophication and pollution of the Everglades ecosystem (Junk et al. 2006 ) Restoring the Everglades wetlands to their natural state is underway, not just to save the flora and fauna, but to ensure valuable water storage needed to maintain the quality of life in areas where population growth and developmen t have substantially i ncreased. The major habitats and vegetation types in the Everglades are: sloughs, sawgrass marsh wet prairies and tree islands (Richardson 2010) Sloughs have the longest
16 hydroperiods and deepest water levels with submerged vegetation and are found in the south central and northeast Everglades. Sawgrass marsh is the main vegetation community in the park found throughout the area, dominated by Tall sawgrass ( Cladium jamaicense ) (Richardson 2010) however it is estimated that its extent is declining, making up only 38% of the 417,000 ha of historic sawgrass dominated areas (Davis et al. 1994) Sawgrass, which is sedge, can occur in pure dense or sparse complexes or is mixed with other sedges, grasses, small shrubs, and attached emergent or floating plants (Ri chardson 2010) The nearly total dominance of sawgrass is one of the distinguishing features of the Everglades (Kushlan 1990) Wet prairies comprise the second most abundant vegetation community, and are found mainly in the northern and central portion of the park (Gunderson and Loftus 1993) These communities typically consist of relatively low growing emergent plant species such as Rhynchospora Panicum and Eleocharis The water depth in the wet prairies is generally deeper than sawgrass marshes but shal lower than sloughs. However, while sawgrass marshes are inundated 6 9 months per year, wet prairies are usually flooded less than 6 months (Kushlan 1990) Tree islands are scattered throughout the park, dominated by a mixture of trees, shrubs and ferns. T ree islands are the highest elevation land features in the Everglades, even though they are elevated only slightly above the surrounding wetland (Willard et al. 2006) with elevations ranging from 0.2 m to 1.5 m above the sloughs and cover less than 5% of t he park (Sklar and van der Valk 2002) Wet graminoid communities in the wet prairie and sawgrass marsh mosaic cover the largest portion of the Everglades ecosystem and shelter 44% of the remaining
17 618,000 hectares of the Everglades National Park (Davis et al. 1994) Wet graminoid communities provide habitat shelter and food for their numerous inhabitants including white tailed deer ( Odocoileus virginianus ), Florida panther ( Puma concolor coryi ), bobcat ( Lynx rufus ), and raccoon ( Procyon lotor ) Cape Sable seaside sparrow ( Ammodramus maritimus mirabilis ), Wood Stork ( Mycteria americana ), cottonmouth snake ( Agkistrodon piscivorus ) soft shelled turtle ( Apalone ferox ) crayfish ( Procambarus alleni ) (USFWS 1999) The continuous degradation of the wet graminoid co mmunities adjacent to the urban rural interface and inner sections in the national park (Davis et al. 1994; Richardson 2010) warrants rapid and repeatable monitoring to better understand changes in their size and flora. These communities will be the first to change in response to restoration, hydrological changes climate change or other impacts such as fire, hurricanes and nutrient input (Gunderson 1997; Armentano et al. 2006; Zweig and Kitchens 2008) So, it is important to rapidly and regularly assess f lora and fauna changes to prevent further degradation. Armentano et al. (2006) predicted that observable change like shifts in species dominance in wet prairie and marsh vegetation may occur in a few years due to changes in hydrological regimes, where the once dominant muhly grass ( Muhlenbergia capillaries ) was replaced by sawgrass ( Cladium jamaicense ) Besides hydrological changes, nutrient availability and fire are considered the most important influences on wetland vegetation in the Everglades (Doren et al. 1997; Childers et al. 2003; Lockwood et al. 2003) Historically the Everglades was a phosphorus limited ecosystem (Richardson and Huvane 2008; Richardson 2010)
18 However, nutrient loading from the Everglades Agricultural Area (EAA) and adjacent urban a reas have significantly increased nutrient concentrations, particularly phosphorous (Davis 1994a; Stober et al. 1996; Richardson 2010) The nutrient loading resulted in increased water phosphorous content, changed periphyton communities, and has led to los s of native saw grass communities, increased organic matter in water, loss of dissolved oxygen, conversion of wet prairie plant communities to cattails, and loss of important wading bird habitats (Stober et al. 1996; Richardson 2010) Olmsted and Armentan o (1997) emphasized that accurate spatial data on vegetation distribution is key to understanding and detecting changes in ecosystem structure and ext ent at the local and landscape scale in the Everglades. Frequent h igh de nsity in situ monitoring of the pa rk would be prohibitively costly and time consuming But, an accurate map developed from aerial or satellite imagery could be used to more rapidly detect and monitor spatial and temporal changes in the wetland graminoid and assess degradation. However, identifying and distinguishing wet graminoid communities in the Everglades using aerial or satellite imagery is difficult because of the color similarity between these communities (Olmsted and Armentano 1997) Several vegetation classification sy stems (Florida Fish and Wildlife Conservation Commission, Florida Land Use and Cover Classification System, Florida Gap Analysis Project and Vegetation Classification System for South Florida National Parks ) (Duever 2006) have been developed and used to ma p the ENP One of the most detailed vegetation map for the ENP was produced in 1994 (Madden et al. 1999) using the Vegetation Classification System for South Florida National Parks The study was
19 interpre tation of 1:40,000 scale color infrared (CIR) film based vertical aerial photographs (Madden et al. 1999) In 2004 a new set of color infrared imagery was taken at the 1:24,000 scale. The imagery from the 2004 mission is currently being processed e mploying visual photo interpretation method to produce a new vegetation map for the Park. The latest aerial imagery which covers the entire ENP was done in April 2009. The South Florida Water Management District (SFWMD) requested the imagery in the framework of th e Comprehensive Everglades Restoration Plan (CERP) program which requires a periodical vegetation map for the Everglades to assess the system wide performance of the CERP (http://www.evergladesplan.org/index.aspx). Since the latest comprehensive mapping e ffort (Madden et al. 1999) was conducted, a number of technological advancements have been made in remote sensing technology. Digital cameras are replacing film based cameras, computer based image classification techniques are used instead of human photo i nterpretation, images can now be radiometrically enhanced and the spatial accuracy in the imagery has greatly increased. This project will apply these new technologies and techniques i n a pilot s tudy to (1) test the use of high resolution aerial imagery, ( 2) assess the usefulness of texture features derived from spectral data, (3) analy ze classification algorithms and (4) determine acquisition times for large scale implementation of these techniques The results of this study will assess which data sets or algorithms are better suit ed for the large scale mapping implementation in the ENP Objectives 1. To determine whether graminoid/sedge ( Sawgrass marsh, Sawgrass wet prairie, Schizachyrium wet prairie, Muhlenbergia wet prairie ) communities can be discriminated based on spectral and textural properties.
20 2. Assess the accuracy of using Back Propagation Artificial Neural Network and QUEST Decision Tree classification algorithms in classifying graminoid/sedge plant communities Significance of the Study This study dev eloped a repeatable classification and discrimination methodology with quantifiable accuracy for wetland mapping and monitoring of at risk plant communities in marl prairie and marsh areas. Based on the findings, a classification plan/scheme was developed by which these vegetation communities could be monitored in an accurate and efficient manner on a regular basis, and an accurate, current map of the graminoid communities produced for the study area in the Everglades. The recently proposed expansion of th e ENP by the State of Florida will significantly increase the area of the park (Stokstad 2008) detailed information on vegetation patterns, including wetland communities which are the major factor in evaluating alteration in w ildlife habitats and ecological change. This research is essential to better understand spatial and temporal changes in vegetation cover, monitoring of hydrologic restoration of the acquired areas, and detect change on the existing area (Olmsted and Arment ano 1997) An annually updateable vegetation database for the ENP w ill offer an optimal solution to track and monitor changes in the For example, Doren et al. (1999) calculated that the cost to develop such a database for the less fragmented and more homogenous Water Conservation Area 1, which is located approximately 70 km northeast of the ENP is about $7.20/ha ($9.66 in Employing a commercial fir m with this unit cost would roughly translate to a cost of
21 $14,000,000 ($18,800,000 in 2011 dollar terms) to map the entire ENP. In a recent study Rutchey and Godin ( 2009) estimated a cost of over $59,000,000 to map the ENP with a 10 m x10 m minimum mappin g unit (MMU) using photo interpretation techniques. Because of cost and precision factors, Rutchey and Godin (2009) recommended develop ing a vegetation map with 50 m x 50 m MMU, which would costs about $560.00 per hectare. However, the authors also mention ed that field ecologists usually require finer MMU (i.e. 2m x 2m). Therefore t o provide more accurate maps and change detection products for local scale studies, investigating plant community shifts, and when capturing spatial heterogeneity is very importa nt, a map using the original pixel size as the minimum MMU is a suitable option to reduce the loss of specificity introduced in the resampling process (Knight and Lunetta 2003)
22 CHAPTER 2 LITERATURE REVIEW Remote S ensing of W etlands Remote sensing techn iques have been used to provide rapid and comprehensive repr esentation of plant communities in marsh and wet prairie (Reed 1988; Sader et al. 1995; Smith et al. 1998; Harvey and Hill 2001; Belluco et al. 2006; Johansen et al. 2007; Dillabaugh and King 2008 ; Maxa and Bolstad 2009) However, the spectral similarity between marsh and wet prairie communities has prevented high classification accuracy (typically above 80%) using single remote sensing data sources (spectral bands) or standard classification appro aches such as Iterative Self Organizing Data Analysis ( ISODATA ) (Reed 1988; Sader et al. 1995) Additionally, significant shifts towards drier vegetation communities such as Muhlenbergia wet prairie, Schizachyrium wet prairie and tree islands in the Evergl ades has occurred since the last mapping effort due to urbanization, anthropogenic drainage activities, climate change (Duever 2005; Ogden 2005; Willard et al. 2006; Richardson and Huvane 2008; Bernhardt and Willard 2009; Pearlstine et al. 2010) These rap id ecological changes illustrate the need for more accurate and automated classifications methods. Previous studies have been done to map wetland plant communities using different remote sensing platforms and techniques (Madden et al. 1999; Belluco et al. 2006; Gilmore et al. 2008) Ozesmi and Bauer (2002) reviewed the literature on remote sensing of wetlands and found that one of the main problems is the spectral separability of specific wetland vegetation classes. Schmidt and Skidmore (2003) reported that mapping major physiognomic or forest type classes is fairly straightforward and provides good, accurate results, while differentiating between grasses or sedges was
23 difficult because of the very similar reflectance wavelengths of these communities. Gluck et al. (1996) found that, indeed, spectral separability of different wetland types is a problem due to the overlap among their spectral signatures. As a limitation of satellite remote sensing, Ozesmi and Bauer (2002) concluded that at medium spatial resolu tion, it is difficult to identify wetland types. Moreover, because most satellites are on a fixed orbit and return intervals, it is difficult to capture optimal lighting and water level conditions for wetlands. They suggested aerial photography as the pref erred image type for detailed wetland mapping. Dahl (2006) agreed that high resolution satellite imagery can be an effective remote sensing platform for wetland mapping. Textural Features Spectral features describe the average tonal variations in various b ands of the visible and infrared portion of an electromagnetic spectrum, whereas textural features contain information about the spatial distribution of tonal variations within a band (Haralick et al. 1973) When a pixel is viewed in context with its surro unding pixels, a pattern may emerge that is consistent with a cover type yet differen t from other cover types. H igh resolution imagery could provide significant feature variables based on first and second order texture analyses (Wang et al. 2004) This sor t of contextual identification can improve the accuracy of image classification and is essential for classification of high resolution images (Ge et al. 2006; Waser et al. 2008) Texture is a statistical property of the spatial distribution of image spectr al tones (Haralick et al. 1973) First order statistics quantify the distribution of spectral tone properties in the image for a given neighborhood while second order statistics describe the frequency with which one gray tone appears in a specific relatio nship to another gray tone in the image and they are calculated from the grey level co occurrence matrix
24 (GLCM). The GLCM indicates the probability of each pair of pixel values ( i j ) co occurs in a given direction and distance (Haralick et al. 1973) Thes e measures can be calculated in four directions (horizontal, vertical, left diagonal, right diagonal) on the image and are then averaged (Haralick et al. 1973) The most common and successfully used first order statistics in classification studies are the mean and standard deviation (Pearlstine et al. 2005; St Louis et al. 2006; Berberoglu et al. 2007; Ashish et al. 2009) The fourteen second order statistics that Haralick et al. (1973) defined are often statistically correlated. Overall, several studies (B araldi and Parmiggiani 1995; Wulder et al. 1998; Clausi 2002; Puissant et al. 2005) attempted to find the most appropriate second order texture features for image classification. Baraldi and Parmiggiani (1995) found that angular second moment (ASM) contra st and correlation were the least correlated among the fourteen second order texture features. Wulder et al. (1998) and entropy were useful. Although Clausi (2002) reco mmended contrast, correlation and entropy; Puissant et al. (2005) recommended homogeneity. For vegetation mapping applications both first order statistics such as the mean and standard deviation and second order statistics have been employed (Pearlstine et al. 2005; Dillabaugh and King 2008; Ashish et al. 2009) Based on previous research findings (Baraldi and Parmiggiani 1995; Wulder et al. 1998; Pearlstine et al. 2005; Puissant et al. 2005; Ashish et al. 2009) mean, standard deviation, ASM, dissimilarity homogeneity and entropy are calculated and tested for each of the digital image (red, green, blue and infrared) bands
25 Photo I nterpretation Visual photo interpretation has been used in several studies (Madden et al. 1999; Rutchey et al. 2008; Maxa and Bo lstad 2009) with good overall accuracy results. Ground collected datasets were used in these studies to evaluate the accuracy of the developed maps. However, photo interpretation is extremely time and labor consuming and highly dependent on the photo inte rpreter skills and knowledge of the area (Ozesmi and Bauer 2002) Most image classification cost s in curred in using digital aerial photography for land cover mapping is not in obtaining the imagery, but in the interpretation process, especially when more t han one photo interpreter is involved (Joria 2001) As resources become scarce and the need for rapid land cover assessment becomes more necessary, government and non government agencies are ever more interested in finding more effective and less costly me thods of obtaining land cover information (Joria 2001) Automated interpretation of digital aerial photographs has a strong potential for offering an alternative to the visual photo interpretation process because of its reliability, repeatability and proce ssing speed (Joria 2001) Puig et al. (2002) also noted that visual photo interpretation is the preferred method for analyzing land cover in low and medium resolution satellite imagery in a tropical setting, but its application is limited to high spatial r esolution imagery because of the increase d need for detail to facilitate recogni tion Baker et al. (2006) also found that comparison of multispectral image classification of wetlands has similar accuracy and greater repeatability to human interpretation. A utomated remote sensing techniques such as Maximum likelihood (ML), Decision trees (DT) and Artificial Neural Networks (ANN)
26 (Wang et al. 2004; Baker et al. 2006; Belluco et al. 2006; Wright and Gallant 2007; Sesnie et al. 2008) are reliable alternatives i n wetland inventory and mapping Automated I mage C lassification T echniques As mentioned above, a great number of remote sensing techniques have been used to differentiate land cover types using space or airborne imaging sensor data in combination with sev eral classification procedures. Maximum likelihood is regarded as a c ommon and widely used supervised classification algorithm for satellite and airborne image classification (Lu and Weng 2007) It is based on normal distribution assumption where the algor ithm uses the mean and variance of the collected training data to estimate the probability that a particular pixel belongs to a category or class. New methods, such as decision trees and neural networks, have been introduced in the last decade (Tso and Ma ther 2009) Decision trees and ANNs become especially popular because they are non parametric models thus make no assumption about the statistical distribution of the data (Richards and Jia 2006) Ozesmi and Bauer (2002) also noted that when using automat ed techniques such as ML, DT, ANN methods, ancillary data such as image texture information (Haralick et al. 1973) can be incorporated into the analysis in addition to multispectral data, and in most cases, can improve classification outcomes (Ghedira et al. 2000; Wang et al. 2004; Fuller 2005; Wright and Gallant 2007; Dillabaugh and King 2008; Sesnie et al. 2008) Wang et al. (2004) for example, incorporated first and second order texture features into his ML classification model and his results indicated that texture increased the classifier accuracy. Sesnie et al. (2008) achieved 12% higher overall accuracy when DEM and climate related ancillary variables were used in addition to multispectral bands in a classification tree analysis. Dillabaugh and King (2008) incorporated Normalized
27 Difference Vegetation Index (NDVI) into their model and found it useful in improving classification accuracy relative to using solely spectral band combinations. Wright and Gallant (2007) tried to discriminate five p alustrine wetland types. When image texture features, DEM and ancillary GIS data (DEM, habitat and cover type) were added to the multispectral data, the average overall error rate dropped incremen tally. However, Michishita et al. (2008) compared decision tree class ifier with ancillary (e.g. DEM and land surface temperature and other spectral indices) data against the ML classifier at the wetland plant family level and found that ML performed slightly better. Michishita et al. (2008) a rgue that the lower accuracy was due to the low number of training pixels used for most of the studied vegetation classes. In general, we can surmise that first and second order image texture features can increase image based wetlands classification. Artificial Neural N etworks has becom e an important part of remote sensing image classification because they can handle complex and very large datasets efficiently and in many cases, produce more accurate results than traditional classifiers (e.g. ML or non parametric parall ele piped classifie rs (Moody et al. 1996) Artificial Neural Networks technology has drawn significant interest in recent years and are being used in landcover mapping (Paola and Schowengerdt 1995; Berberoglu et al. 2007; Ashish et al. 2009) frequently, but literature on the ir application and use in wetland mapping is limited (Ghedira et al. 2000; Mas 2004; Fuller 2005; Dillabaugh and King 2008) So, t hese studies indicate that ANN can be used to improve vegetation discrimination with various success rates
28 In a recent land u se classification study Ashish et al. (2009) employed probabilistic ANN using 1 meter resolution multispectral image and second order texture features (Haralick et al. 1973) in the classification process. The overall accuracy when image texture was involve d was 89%. Berberoglu et al. (2007) tested several texture features for landcover mapping with ML and ANN classifiers. Artificial Neural Networks outperformed the ML classifier in the overall classification accuracy without texture data using Landsat spect ral bands. The authors also reported that ANN performed better when texture data was incorporated into the classification; however the overall accuracy slightly decreased for agricultural ( citrus plantation, first crop corn, second crop corn, cotton, soil, soya, water and urban) and semi natural ( bulrush, cotton, dune vegetation, salty plain, dune, soil, wet soil, wetland and water) landcovers. Ghedira et al. (2000) attempted to distinguish between wetland categories using the ANN back propagation algorithm on texture information of RadarSat SAR (Canadian Space Agency) data. When texture was included in the analysis, 99.8 percent of the pixels were classified, and less than 3 percent of the total number of pixels were classified as a mixed class (shrubby and woody wetland). Mas (2004) used Landsat imagery to map land cover classes in a tropical, coastal area by employing ANN and ancillary data such as elevation and soil data. Model output was compared with output from spectral classification and a significant increase in accuracy of the land use classification (from 67% to 79%) was detected. Multispectral IKONOS imagery was used by Fuller's (2005) to detect invasive M elaleuca trees ( Melaleuca quinquenervia ), sawgrass and 3 other classes in South Florida. Back propagation ANN was employed as well as spectral bands, texture layer and normalized difference vegetation index as input layers. The
29 author achieved over 85% overall accuracy, however a visual inspection revealed that a frequently misclassified as Melaleuca class. However, Dillabaugh and King (2008) attempted to map wetland composition using Ba ck Propagation ANN based on high resolution satellite imagery (IKONOS) and concluded that ML classifier performed better than the artificial neural network classifier. Most of these studies above utilized satellite or hyperspectral imagery with variant spatial resolution. The delineation of wetlands as single landcover was frequently the objective of these studies, rather than deli neation of vegetation classes or species within wetlands. Therefore, t he purpose of this study is to determine the applicability of multiple classification algorithms such as decision trees and neural networks to graminoid comm unities using high resolution aerial imagery and texture features Hypothese s 1. Classification accuracy of grass /sedge communities will be statistically significant using texture features at an alpha of 0.05 level. 2. Machine learning i mage classification metho ds can discriminate graminoid communities more accurately than the traditional Maximum Likelihood s tatistical classifier algorithm at a n alpha of 0.05 level The classification accuracy of the developed maps was assessed using confusion matrices, where use presented. Kappa coefficients (Foody 2004) to evaluate the difference between two thematic maps were also calculated. To compare the kappa coefficients z tests were performed to test the hypo thesis whether the results of two classifiers differ significantly (Donner et al. 2000)
30 Study A rea The study site is located in the eastern portion of the ENP as shown in Figure 2 1 (Miami Dade County). The area was selected because of its diversity of land cover types including spatially mixed marsh and marl prairie but homogeneous communities as well. The area w as also selected because recent baseline data (Ross et al. 2003; Ross, Sah, et al. 2006; Ross, Mitchell Bruker, et al. 2006) of these vegetatio n communities exists. The terrain of the study area is flat. Elevation varies between 0 and 0.6 meter
31 Figure 2 1. Study area within the E verglades National P ark Florida USA
32 CHAPTER 3 WETLAND COMPOSITION ANALYSIS USING HIGH RESOLUTION IMAGES AND TE XTURE FEATURES Introduction contain about 12% of the global carbon pool and play an important role in the global nutrient and carbon cycle s (Erwin 2008) One of the biggest concerns in regard to climate change effe cts is the status of wetlands regarding nutrient (nitrate, iron and sulfate) dynamics and matter fluxes (Paul et al. 2006; Erwin 2008) The Everglades are being subject ed to numerous detrimental effects such as sea level rise, se dimentation, chemical poll ution, human disturbance, and climate change effects (Titus and Richman 2001; Ogden 2004; Pearlstine et al. 2010; Larsen et al. 2011) These are affecting endangered species such as th ose of the Florida panther ( Puma concolor coryi) American cro codile ( Crocodylus acutus ) and Cape Sable seaside sparrow ( Ammodramus maritimus mirabilis ) (Walker and Solecki 2004) In addition, the past decade, the growth of the Miami Fort Lauderdale metropolitan area in the last decade intense agriculture activities in the Lake Okeechobee area and human induced changes in the hydrological system have caused significant disturbance in the Everglades ecological system (Junk et al. 2006; Marella 2009; Richardson 2010) Restoring the Everglades wetlands to their natural state is in progress to not only to save the flora and fauna, but to ensure valuable water storage and improve living conditions in areas with changing population growth and development (Richardson 2010) Wet graminoid communities ( e.g. marl prairie s and marsh es ) cover the largest portion of the Everglades ecosystem (Davis et al. 1994) and provide living area, shelter and food for their animals These communities suffer continuous degradation from being
33 close to the urban rural interface, as well as in som e developed inner sections of the Everglades National Park (ENP). C hanges in graminoid communit y distribution could be used to provide advanced warning or indicators of ecological disturbance and more severe ecological shifts due to climate and/ or land use changes (Goodin and Henebry 1997; Pearlstine et al. 2010) Monitoring these graminoid communities could also serve as advanced warning against expected sea level rise and a tool to better assess and evaluate restoration efforts. Olmsted and Armentano (199 7) stressed the disturbance monitoring concept and confirmed that accurate spatial data on vegetation distributions is key to understanding and detecting changes in ecosystem function at the local and landscape levels. In fact the discrimination of wet gr aminoid communities using remote sensing technologies has been the most frequently cited challenge in past attempts at land cover classification in many areas worldwide including the Everglades (Ozesmi and Bauer 2002) Remote sensing techniques have been u sed to provide rapid and comprehensive discrimination of plant communities, including marsh and wet prairie (Smith et al. 1998; Belluco et al. 2006) However, the spectral similarity and inter digitations between these two communities prevented high classi fication accurac ies using a single remote sensing data source (spectral bands only) or standard classification approaches such as unsupervised or parallelepiped classification Although vegetation maps for the ENP have been reported to have low classificat ion accuracies, there is an ongoing shift toward drier vegetation communities (Armentano et al. 2006) Accuracy assessment results of data collected through direct visual interpretation of 1:40,000 scale vertical Color Infrared (CIR) aerial photographs ta ken in the mid (Welch et al. 1995) in
34 the ENP found a 25% error in the classification of Tall sawgrass communities, 33% error for Short sawgrass and 60% error for muhly grass/marl prairie Spatial resolution is one of the most significant factors th at influence s image classification accuracy. Yang (2007) demonstrated the significance of high spatial resolution (2m) scanned, ortho rectified color aerial images in accurately classifying riparian zone vegetation when compared to Satellite for the Observ ation of the Earth (SPOT) 4 and Landsat 7 Enhanced Thematic Mapper (ETM + ) imagery. The overall accuracy of the high resolution aerial imagery results was 81% while Spot 4 and Landsat 7 provided 63% and 53% overall accuracy. Maheu Giroux and Blois (2005) used scanned black and white (B&W) and color aerial photos (0.33m ground reso lution) to discriminate c ommon r eed ( Phragmites australis ) an invasive plant from other types of vegetation They achieved an overall accuracy between 71% 87% for panchromatic i mage analysis a nd 77% 88% for color imagery H owever when both results were compared it was evident that the difference s w ere not statistically significant. Dillabaugh and King (2008) used the spectral bands of IKONOS imagery (4 meter resolution) to disc riminate wetland vegetation types. The overall accuracy of their classification increased when the Normalized Difference Vegetation Index (NDVI) was included in the analysis. High resolution imagery provides better discrimination of significant features be yond their spectral content represented by their first and second order texture features (Haralick et al. 1973; Wang et al. 2004) This sort of contextual identification can improve the accuracy of image classification of high resolution images (Ge et al. 2006; Waser et al. 2008) Moreover, in digital image processing and remote sensing
35 data analyses, the spatial neighborhood of a pixel may provide even more information than the pixel itself Spectral features describe the average t onal variations in various bands of the visible and infrared portions of the electromagnetic spectrum, whereas textural features contain information on the spatial distribution of tonal variations within a band (Haralick et al. 1973). When a pixel is viewe d in context with its surrounding pixels, a pattern may emerge that is consistent with a landcover type and different from other types thereby facilitating discrimination (Pearlstine et al. 2005) The most commonly used first order textural statistics use d for most classification studies are the range, mean standard deviation/variance, skew and entropy (Coburn and Roberts 2004; Pearlstine et al. 2005; St Louis et al. 2006; Ashish et al. 2009) In Coburn and Roberts' (2004) study, the authors found that in clusion of variance significantly increased the classification accuracy of landcover classes by 9%. Pearlstine et al. (2005) concluded that the mean and standard deviation were very effective in discriminating Brazilian pepper ( Schinus terebinthifolius ) fr om other type of vegetation. St Louis et al. (2006) used first order texture measures to predict bird species richness and reported that range and standard deviation were significant predictors. Ashish et al. (2009) also observed that mean and standard dev iation increases the overall mapping accuracy. Baraldi and Parmiggiani (1995) found that angular second moment (ASM), contrast and correlation provided the best results when testing it on an Advanced Very High Resolution Radiometer (AVHRR) using cluster an alysis. Wulder et al. (1998) using compact airborne spectrographic imager (CASI)
36 imagery, and found that homogeneity, contrast, dissimilarity, and entropy were useful. Clausi (2002) recommended contrast, correlation and e ntropy based on Synthetic Aperture Radar (SAR) imagery and maximum likelihood classifier while Puissant et al. (2005) recommended the use of homogeneity texture feature using SPOT 5 imagery Since textural features are scale dependent, determining the opt imal window size to extract these features is important (Puissant et al. 2005; Tso and Mather 2009) In general, a larger window size performs better and is needed for higher resolution imagery and f or spatially heterogeneous classes. On the other hand a small window size yielded higher accuracy for spatially homogeneous classes (Ashish et al. 2009) However, if the window size is too small relative to the examined texture feature, the real property of the texture feature will not be accurately reflected ( Tso and Mather 2009) A 3x3 window size is generally preferred for first order texture feature extraction (Berberoglu et al. 2007) Different window sizes have been used to extract second order texture statistics (Wang et al. 2004; Puissant et al. 2005; Ge et al. 2006; Ashish et al. 2009) Finding the most appropriate window size f or the appropriate scale al. 2006) for better achiev ing the highest classification accuracy is difficult Semivariograms have been extensively used in remote sensin g studies to calculate optimal neighborhood size 2007; Balaguer et al. 2010) Franklin et al. (1996) successfully used semivariogram range to estimate the optimal window size by calculating the semivari ance for different vegetation classes in an image They concluded that the semivariogram range can be a useful guide for selecting the appropriate texture window size. Semivari ance is typically
37 expressed as a semivariogram curve that depicts its v alues at different lag distances In a typical semivariogram, the semivariance values increase with the increase of the lag distances up to a certain distance (range) where the values are no longer correlated with the lag distance and thus provide a measur e of the size or scale of the largest and most dominant elements in the scene (Curran 1988) Franklin et al. (1996) also concluded that the semivariogram range can be useful as a guideline for selecting the appropriate second order texture window size. Thi develop a method for utiliz ing spectral and textural characteristics of high resolution digital aerial imagery to better discriminate between marl prairie and marsh plant communities in the ENP These plant communities have been report ed as difficult to classify due to their spectral similarity (Olmsted and Armentano 1997) Thus, the main hypothesi s of this research is to determine if first and second order textural features lead to significant improvement in the classification accuracy Specifically, this study used multiple combinations of spectral and textural features as additional dimensions to improve classification accuracy Multiple window sizes for first and second order texture features were also tested to identify optimal para meters for achiev ing the highest classification accuracy Methods Classification Scheme Several vegetation classification schemes (Florida Gap Analysis Project, Florida Land Use and Cover Classification System, Florida Fish and Wildlife Conservation Commi ssion, Multi Species Recovery Project, and Vegetation Classification System for South Florida National Parks ) (Duever 2006) have been used to classify the ENP plant communities This study employed a modified version of the Vegetation Classification
38 System for South Florida National Parks (VCS) (Rutchey et al. 2006) This system was developed by the South Florida Natural Resources Center, the Center for Remote Sensing and Mapping Science at the University of Georgia, Big Cypress National Preserve and the So uth Florida Water Management District (Rutchey et al. 2006) The major vegetation types in this system are hierarchically arranged plant communities which are defined by the dominant vegetation species. However, the prairies and type needs additional classes to better separate graminoid wet prairie and marsh communities as sho wn in Table 3 1 Most marsh types were dominated by fewer than 10 species (Kushlan 1990; Craft et al. 1995; Olmsted and Armentano 1997) while w et prairies ha ve much higher species richness (Olmsted et al. 1980) Table 3 1 also contains the major plant species in the study site (Ross et al. 2003; Ross, Sah, et al. 2006) Study A rea and F ield D ata C ollection The 8.5 km study area was flat and had elevation ran ging from 0 to 0.6m above sea level and is located in the south eastern portion of the ENP (252509.09 to 252415.16 N and 804641.46 to 804541.00 W) as shown in Figure 3 1. The study area was selected because of its existing land cover, which includes an intermix of distinct marsh and marl prairie communities. The study area has been previously restoration efforts (Ross et al. 2003) This study utilized a datas et collected in the spring of 2008 by scientists from Florida International University (Dr. Mike Ross and Michael Kline personal communication ) as well as data from existing plots that were revisited and measured in the spring of 2009. These data sets wer e collected using Trimble Pathfinder GPS units (accuracy < 0.5m) and included information about the plant
39 communities (Table 3 1) water levels, and soil depths. Additional ground reference real time kinematic GPS unit (TOPCON Hiperlite, minimum accuracy of < 0.1m) at the time of image acquisition. In general (Tso and Mather 2009) suggest that for a statistical classifier, at least 30 times the number of feature classes (i.e. number of band s) is needed to determine the number of training data pixels per class. The training dataset collected as GPS points were extrapolated into polygons using visual image interpretation to include a sufficient number of pixels in the analysis (Table 3 2). Bas ed on the proposed five class classification scheme (see Table 3 1), the number of feature classes ( 15 spectral and texture features) and Tso and Mather (2009) suggestion regarding training data size, a minimum of 450 pixels were selected (Table 3 2) as a training set for each class. Each training and accuracy polygon contained a minimum of 297 pixels. T he training area s for each class were larger than the minimally determined pixel number (450) to acquire statistically significant number of pixels for each class as well as to get good representation of the class spectral reflectance values. Two thirds of the ground truthing data were used to train the classifier while the remaining third was set aside to test classification accuracy Imagery The image s use camera in the spring of 2009. The main characteristics of the imagery are summarized in Table 3 3. Natural color and color infrared imagery were collected over the ENP. For the UltracamX cam era, a separate sensor was used to collect each spectral band, with the optical path passing through a filter, lens assembly, and a C harge C oupled D evice
40 (CCD) array. These bands are collected when the image gets triggered during the survey flight. All ima ges were free of smoke, clouds and cloud shadows. The images were acquired while the solar elevation angle was greater than 30 above the horizon. They were collected in the spring during spring leaf on conditions. The images were geometrically and rad iometrically corrected and pan sharpened (UltracamX system data level 3). No orthorectification was performed; however, geometric distortions due to relief change were ignored because of the minimal elevation relief found in the study research area (Wolf a nd Dewitt 2000) Geo referencing of the imagery was accomplished via direct referencing using post processed airborne global positioning systems (GPS) and inertial measurement unit data supplier, Aerial Cartographics of America, Inc. Orlando FL The horiz ontal positioning accuracy of the imagery was 3.00m with a 95% confidence interval (Root Mean Square Error, RMSE = 1.76m given that positional accuracy = 1.7308 RMSE). Imagery was in the State Plane Coordinate System, Florida East Zone 0901, North Americ an Datum of 1983/High Accuracy Reference Network (HARN). Imagery was mosaiced to construct a single image covering the entire study area and clipped to save file space and to eliminate the portion of the images beyond the area of interest (Figure 3 2) De velopment of T exture F eatures First order statistics ( i.e. data range, mean, variance and entropy) were used to quantify the distribution properties of the image spectral tone for a given neighborhood (Haralick et al. 1973). S econd order statistics (i.e. m ean, variance, homogeneity, dissimilarity, entropy and angular second moment) describe d the frequency with which one gray tone appears in specific relationship to another gray tone in the image These
41 features were calculated from the grey level co occurre nce matrix (GLCM) extracted from an image neighborhood (moving window). The GLCM indicates the probability that each pair of pixel values co occur s in a given direction and for certain lag distance in the image (Haralick et al. 1973) Texture measures wer e calculated in four directions (horizontal, vertical, left diagonal, right diagonal) and averaged ( ITT ENVI software ) Out of the fourteen second order statistic texture features that Haralick et al. (1973) defined, many were correlated. Thus, the mean, v ariance, homogeneity, dissimilarity, entropy and ASM were calculated and tested for the near infrared bands. The spectral bands of the imagery were highly correlated (r>0.95), thus only one band (near infrared) was used to calculate the texture features. T hese second order textures measures were considered to be the most relevant texture measures for image classification as well (Baraldi and Parmiggiani 1995) The first and second order texture features are calculated based on formula e shown in Table 3 4 In addition, image statistics w ere used to explore statistical relationships between texture layers. In this study the most feasible window size was estimated for each second order statistics feature using a semivariogram approach (Carr and de Miranda 1998) Determining the spatial extent (i.e. window size) for extracting the GLCM is important for texture analysis so one of the most common methods for determining window size is based on evaluating the spatial autocorrelation between image pixels. Since semi variogram depicts the relationship between semivariance values (half of the mean of the sum of the square of the difference between pixel values at certain spatial lag distance) and the corresponding lag distances, the semivariance S 2 between a pixel
42 value z(x) and its neighboring pixel value z(x+h) at a lag distance h were computed as (Carr and de Miranda 1998) : Where m is equal to the number of value pairs in which the separation distance is equal to h The most dominant grass comm unities in the study area were the Short sawgrass wet prairie (CWP) and Muhlenbergia wet prairie (MWP) (Ross et al. 2003; Ross, Mitchell Bruker, et al. 2006) The typical diameter of these communities were about 10 meters, thus 30 pixels were included as t he range in which semivariance values were calculated for each vegetation class from the training polygons in 8 directions (N, S, W, E, NE, SW, SE, NW). Separability A nalysis The spectral signatures five classes were examined visually and statistically. Visual inspection of these signatures indicated strong similarit ies, except for the Woody vegetation class (Figure 3 3 ) and can be attributed to within species phenological differences, the relative similarity between involved species and th e spatial intermix of plant communities. Examples of such community mix were noted in the existence of Short and Tall sawgrass vegetation within the sawgrass communities and the spatial intermix between the Muhlenbergia wet prairie and Schizachyrium wet pr airie classes. The Jeffries Matsushita (J M) separability measure (Richards and Jia 2006) was used to compute the statistical distance between the spectral/textural signatures of each pair of classification classes as represented by their training sets and results were used to determine the optimal classification classes based on the maximum separability
43 among bands. The J M separability measure is asymptotic to a value of 2.0. Classes with J M values close 2.0 imply high separability and thereby providing potentially good classification accuracy Parametric C lassification In this study, m aximum l ikelihood classifier (ML) classification of different combination of the image spectral information (spectral bands and first and second order texture features ) was tested Maximum likelihood classifier is a statistical classifier that assumes normal distribution for the values of each class training set and can be described by the mean and covariance matrix. Maximum likelihood uses training areas to compute the distribution statistics from the study areas and use the s e statistics to calculate the probability of each pixel belong ing to a class distribution of values for each spectral band. Maximum likelihood can be calculated by the following discriminant function ( for each class: Where =class, = data vector, p ( i ) = probability that class i occurs in the image and is assumed the same for all classes, i = covariance matrix of class = mean vector of class and is the transposition function The discriminant is calculated for each class and the class with the highest value is selected (Paola and Schowengerdt 1995) Accuracy A ssess ment using one third
44 of the training data set ( Table 3 2 ) User's accuracy as defined in this study measure s the probability that a pixel on the map accurately identifies the classes. Accuracy was calculated by dividing the total number of correctly classified pixels with the total number of all pixels assigned to that particular class in the map, thus reflecting the measure of commission error. Producer's accuracy measures the probability that a pixel observed in the field is correctly classified in the map and reflects the error of omission, where the correctly classified pixels in a class are divided by total number of pixels of the particular class measured in the field. Overall accuracy is the percentage of correctly classifi ed points from the field data and is computed by dividing the sum of the correctly classified pixels by the s um of all pixels in the error matrix (Congalton 1991; Tso and Mather 2009) accuracies of the maps were designated by five standard categories : very high accuracy ( over 90 % ), high accuracy (80 % 90% ), accept able accuracy (70 % 80% ), low accuracy (50 % 70 % ), and poor accuracy ( less than 50 % ). Additionally, the Kappa coefficient was also used to test whether the generated classification is significantly better than a randomly generated map. Kappa coefficient ran ge d between 0 and 1, where a 1 indicates total agreement (perfect classification). Kappa coefficient was calculated as follows:
45 Where is the number of ro ws in the matrix, is the number of observations in row and column and are the marginal totals of row and column respectively, and is the number of observations (Congalton 1991; Foody 2004) Evaluation of the C la ssification R esults Kappa coefficients (Foody 2004) were calculated to evaluate the difference between two thematic maps To compare the kappa coefficients, z tests were performed to test the hypothesis whether the results of two classifiers are significan tly different (Donner et al. 2000) at a 5% level of significance if |z| >1.96 (Congalton 1991) The z tests were based on the following formula: where corresponds to the estimated covariance between and (Donner et al. 2000) Results Semivariogram A nalysis The spatial autocorrel ation of each of the 5 vegetation classes were analyzed to determine optimal window size for texture feature extraction. Spherical s emivariograms were calculated using pixel values of th e training data sets (Figure 3 4 ) for each spectral band to estimate t he most appropriate window size for the texture features (Figure 3 5). The semivariogram curves of the vegetation classes showed different behavior s in a particular spectral band H owever, the semivariogram trends were similar for the class es in a differen t spectral band. The semivariogram of the Schizachyrium wet prairie semivariogram with very low semivariance values. The Muhlenbergia WP class had low semivariance values; however its semivariogram was
46 and had rising trend until a pproximately 3 pixels and was flat with a minor hump at around 11 pixels. The Tall sawgrass marsh had medium semivariance values compared to the other classes with an increasing semivariogram in the first 4 5 pixels, but flatten ed afterward until 11 pixels and was similar to the Schizachyrium WP class where a minor increase was noticeable. Short sawgrass WP had a shape comparable to the MWP and CM classes with a rising trend in the first 4 5 pixels and low internal variability wi th a small increase after 10 pixels. The Woody vegetation class had the highest semivariance values and a semivariogram shape that is similar to the MWP, CM and CWP classes sho w ing a clear increasing trend t o 4 pixel s and low spatial variability afterword. Using these semivariogram curves t he most appropriate window size for extracting the texture features was manually identified. Since most of the classes had very similar patterns with increasing semivariance values up to 4 5 pixels of lag distance follow ed by a flattened curve, with small variations up to a lag distance of 11 pixels, a set of window sizes ranging between 3 and 15 pixels were tested for the first and second order texture features Spectral S eparability Spectral separability analysis using the J M index revealed that using spectral bands alone would not provide the necessary separability to discriminate among classes (Table 3 5). Only the W oody vegetation class could be classified accurately (J M > 1.98) if only the four spectral bands were used in the classification. Classes with similar species in the community such as the Short sawgrass wet prairie versus Tall sawgrass marsh had as low J M value as 0.38. Similarly, Short sawgrass wet prairie versus Muhlenbergia wet prairie had a 0.39 J M v alue.
47 When first and second order textural features were included in the separability analysis, the discrimination accuracy increased (Table 3 6). Most of the class comparisons revealed high separability. Only the Short sawgrass wet prairie and the Muhlenb ergia wet prairie had a lower J M value (1.40). These results indicate the potential of obtaining high classification accuracy by introducing texture features as additional bands in the classification process Using S pectral B ands and N ormalized Differenc e Vegetation Index The ML classifier was implemented ( ITT ENVI software ) using : (1) the 4 spectral bands only and (2) the spectral bands in combination with the normalized difference vegetation index (NDVI) layer. The classification results ( Table 3 7 ) usi ng the 4 spectral bands resulted in an acceptable o verall accuracy ( 73.8 the C WP and SCWP classes while t he Woody vegetation class had very high When the NDVI layer was included in the ML c la ssification the overall accuracy of the map increased to 75.4%, which is only a 1.6% improvement in overall accuracy, compared to the accuracy obtained using only spectral bands Incorporation of F irst O rder T exture F eatures The results using the textura l features in addition to the spectral bands are shown in Table s 3 8 and 3 9 Experimenting with different window size used to extract first order texture revealed varying accurac ies with the highest overall accuracy ( 74.08 %) was achieved with a 9 by 9 win dow size ( Figure 3 6 ) The poorest overall accuracy (5. 01 %) was achieved for the 5 by 5 window size, which was associated with the very low number of pixels correctly recognized in the CM class. The classified image produced using the 4 spectral bands and their texture features ( 9 by 9 window) had high
48 and acceptable accurac ies in detecting CWP and MWP classes (81.85% and 75.74%, respectively) high for the CM and SC WP classes. When employing larger window size (11 by 11 and 15 by 15) a decreasing trend in overall accuracy was noticeable, this is associated mainly with the poor Incorporation of S econd O rder T exture F eatures Maximum Likelihood classification res ults obtained using spectral and second order texture features (using different window sizes) are presented in Tables 3 10 and 3 11. The highest overall accuracy was achieved when a 7 by 7 window was employed (76.89%); however, there was no significant dif ference between these results ( > 0.05 ) and the ones achieved using 5 by 5 window size. However, some of the vegetation window sizes For example, the SCWP class had a very hig (89.90%) when using the 3 by 3 window and the CM class had very accuracy (90.41%) when a 9 by 9 window was employed. When larger windows (e.g. 11 by 11 or 15 by 15) were utilized, the overall accuracy decrease d (72.72% and 67.61%, respectively). Merging D ifferent W indow S izes The spectral bands and NDVI combination provided a 75.40% overall accuracy (Table 3 7). When texture features from different window sizes were also incorporated in addition to the spectral bands an d the NDVI the accuracy of the map significantly increased. Spectral bands, NDVI and first order texture features with the highest overall accuracies (3 by 3 and 9 by 9 pixels windows); were stacked and tested and produced an 81.73% overall accuracy map (T able 3 12) This is significantly better than previous
49 accuracies obtained using only the spectral bands and NDVI or the spectral bands and racy improved or slightly decreased in t he case of CM and SCWP classes. To further reduce the number of bands in the analysis, image statistics were examined for the 3 by 3 and 9 by 9 window first order texture features. The statistics revealed very high c orrelation ( r> 0.99) between the 3 by 3 window variance and the 9 by 9 window variance feature s while the data range, mean and entropy features for both wind ow sizes were less correlated ( r= 0.85). Because of the correlation between the variance features, t he 3 by 3 pixels window variance and the 9 by 9 pixels window data range, mean and entropy features were merged with the spectral bands and the NDVI layer and classified. The classification results listed in Table 3 12 shows that the overall accuracy incre ased to 84.01%. Second order texture features were also tested in combination with the image spectral bands and the NDVI layer. Instead of using the 5 by 5 or 7 by 7 window texture features, which produced the best overall accuracy, the 3 by 3 and 9 by 9 f eatures were used due to their higher producer SCWP and CM communities (Table 3 10). The overall accuracy using the 3 by 3 and 9 by 9 second order texture features was 84.92%, with very high and high for the CM and SCWP classes (94.36% and 87.96%, respectively) and acceptable and low values for CWP (75.35%) and MWP (65.89%) classes. To reduce the number of features, image statistics were reviewed similar ly to the first order texture features. The revi ew revealed that the variance feature from the 3 by 3 window was highly correlated
50 ( r >0.94) with the variance feature of the 9 by 9 window variance. Thus, the variance, from the 3 by 3 GLCM and the mean, homogeneity, dissimilarity, entropy and ASM from the 9 by 9 GLCM were used in the classification with the spectral bands and the NDVI layer. The overall accuracy of the classification was 84.33% (Table 3 13), slightly lower than the results obtained using all 3 by 3 and 9 by 9 second order texture features. Finally, both first and second order texture features (3 by 3 and 9 by 9 window) were utilized with the spectral bands and the NDVI layer. The overall accuracy of the classification was slightly lower (83.97%) compared to the previous results using the fi rst and second order texture features separately (Table 3 13) Discussion One of the objectives of this research was to test the performance of the ML classification of wetland communities when texture features were introduced. The results showed that inco rporating texture features in the classification yielded better classification accuracy than using the image spectral bands alone However, when extracted first order texture features using only one window size were included in the analysis the overall ac curacy decreased slightly The incorporation of the second order texture features extracted from a single window size slightly increased the achieved overall accuracy. The low overall accuracy associated with using the spectral bands can be explained by t he very similar spectral properties of some of the classes as indicated by the J M separability analysis. For example, the Tall sawgrass marsh (CM) Short sawgrass (CWP) and the Muhlenbergia wet prairie (MWP) classes w ere poorly detected. The Schizachyrium wet prairie (SCWP) class was incorrectly classified 59.78% of times in the field, had better detection accuracy (76.15 %). The inclusion of
51 the NDVI layer increased the overall accuracy by 1.6% as well as the detectability of the CM, MWP and SCWP classes. When different window sizes were merged together with the spectral bands and the NDVI layers, the overall accuracy and the accuracy of individual vegetation classes < 0.05). This indicates that the additional characteristics of the texture features developed from multiple window size provide useful information, which can improve the accuracy of the automated classification algorithm. Results from similar studies indicated similar findings when spectral bands and different window siz es were included in the analysis (Wulder et al. 1998; Johansen et al. 2007; Dillabaugh and King 2008; Barbosa 2010) Johansen et al. (2007) as well as Barbosa (2010) achieved 78.95% and over 80% overall accuracy when classifying wetland vegetation, respect ively, when using both spectral bands and texture features ( contrast, dissimilarity, homogeneity, ASM Dillabaugh and King (2008) reached 88% overall accuracy when classifying broad wetland classes (marsh, floating aquatic, emergent aquatic, forest, open w ater). The window size was a n important parameter when deriving texture measures. The window size is associated with the scale of the image content as expressed by individual objects. In this study the objects were individual plant community elements that can be differentiated from a certain size d atial resolution. The communities included not only larger contiguous areas ( CM class ), but also smaller discontinuous patches such as the Schizachyrium wet prairie. M oving windows between 3 by 3 and 15 by 15 pixels were tested considering the smallest vegetation patches in the area image resolution and the range property of the
52 semivariogram The results indicated that in the case of the CM class, which can cover larger areas, smal accuracy of both classes increased, in the case of CM class from 12.56% (3 by 3 window) to 70. 33% (9 by 9 window), while the W oody vegetation was similar regardless of the size of the moving window. The CWP class classification accuracy showed some variations with different window sizes but remained above 63% except when a 7 by 7 window s s accuracy decreased to 42.5%). The S CWP class showed a window size was increased. When small window size (3 by 3) was tested, the SCWP with an overall accuracy of 70.46%. Larger window sizes (5 by 5 and 7 by 7) produced almost identical overall accuracy (76.35% vs. 76.89%) ; however, the accuracy of the SCWP class started to decrease, when 9 by 9 or larger window size was used This increase in w indow size (9 by 9 and more) was associated with an increase in the CM class accuracy, which reached a 90.41% Short sawgrass and the MWP class accuracy were steady in all window sizes indicating insensitivity to window size Overall a ccuracy declined significantly when window size was increased above 9 by 9 for all classes indicating that larger window sizes are not appropriate in detecting small size vegetation communities in the ENP This highlights the importance of determining the best window size (or combination of windows), where optimal classification accuracy of different vegetation communities can be achieved
53 Texture features obtained from different window sizes that best facilitated the delineation of the vegetation communit ies were merged and ex amined Since the Woody vegetation CWP, MWP classes were less sensitive to window size changes and the 3 by 3 and 9 by 9 window size s helped in best delineat ing the SCWP and CM class es, these window sizes were used in the analysis wi th the spectral bands and the NDVI layer When 3 by 3 and 9 by 9 first order texture features were used, the overall accuracy increased significantly 0.05) compared to either using the spectral bands only (73.80% vs. 84.01%) or using texture features f rom a single window size (76.89% vs. 84.01%) S window size was used, its variance feature was merged with the 9 by 9 window range, mean and entropy features as well as the spectral band s and the NDVI layer. The overall accuracy of this combination increased significantly to 84.01% ( 0.05 ) compared to the spectral bands, NDVI, and all available (3 by 3 and 9 by 9) first order texture features (81.73%) This implies that variance feature is a very important feature for d ifferentiating among classes and furthermore when decreasing the number of layers in the classification and the results are statistically different ( 0.05 ) ; the computation time was reduced Second order texture 3 by 3 and 9 by 9 features were also merged with the spectral bands and NDVI and this resulted in a n 84.92 % overall accuracy S imilar to the first order texture features, the 3 by 3 variance and the 9 by 9 mean, homogeneity, dissimilarity, entropy and ASM were tested and resulted in a slightly decreased overall accuracy ( 84.33% ) Th is study results indicate that a combination of first and second order texture features and different moving window sizes in addition to the image spectral bands and
54 NDVI layer perform significantly better ( 0.05 ) than all other data combination s I n general, first order texture features are easier to compute and the classifier tend s to be faster because of the less number of features (data range, mean, variance and entropy) compared to the second order text ure features which are more computationally exhaustive and have more features to include in the classification. The best window size depends on the spatial distribution of a particular vegetation class F or example the SCWP class was differentiated very well with smaller window size, while the CM class needed a larger window size. Th is observation agree s with previous work (Ouma et al. 2008) wh ich showed that the most appropriate texture feature and window size ( s ) to classify a vegetation community depen d on the textural pattern of th is particular vegetation community. The semivariogram analysis was found to be useful for estimating the spatial autocorrelation within the vegetation communities, and hence defining a range of window sizes to extract texture features. However, it did not specify the preferred window size for each class. A l arge window size can depict the spatial patterns of large patches communities ; however when delineat ing vegetation communities of smaller patches, larger window sizes can c apture additional communities and increase the overall error rate. To correct for this, w indow s should be small enough for communities of small spatial extent so as to keep the varia bility low and to maximize the potential for class separability
55 Fig ure 3 1. Location of study area in the Everglades National Park based on a Landsat T hematic M apper 5 image in true color composite. The images to the right show Short s awgrass wet prairie with some trees (top) M uhlenbergia wet prairie (middle) and S chizac hyrium wet prairie (bottom).
56 Figure 3 2. Microsoft Vexcel UltacamX m osaiced image covering the study area in the ENP
57 Figure 3 3 Spectral response curve for the studied vegetation communities in the Blue Green, Red and Infrared spectral bands DN Digital Numbers. CM Tall sawgrass marsh CWP Short sawgrass wet prairie MWP Muhlenbergia wet prairie, SCWP S chi zachyrium wet prairie, Woody W oody vegetation 600 800 1000 1200 1400 1600 1800 2000 Blue Green Red InfraRed DN MWP Woody SCWP CM CWP
58 A. B. C. D. E. Figure 3 4 Selected training sites for semivariance cal culation. (A) S chizachyrium wet prairie (B) Muhlenbergia wet prairie (C) Tall sawgrass marsh (D) Short sawgrass wet prairie (E) Woody vegetation
59 Figure 3 5 S emivariograms derived from four spectral bands for the selected train ing sites. CM Tall sawgrass marsh CWP Short sawgrass wet prairie MWP Muhlenbergia wet prairie, SCWP S chizachyrium wet prairie, Woody Woody vegetation 0 200000 400000 600000 800000 1 3 5 7 9 11 13 15 580 700 nm (Red) Semi variance lag (pixels) 0 200000 400000 600000 800000 1000000 1 3 5 7 9 11 13 15 480 630 nm (Green) Semi variance lag (pixels) 0 50000 100000 150000 200000 250000 300000 1 3 5 7 9 11 13 15 410 540 nm (Blue) Semi variance lag (pixels) 0 100000 200000 300000 400000 1 3 5 7 9 11 13 15 690 1000 nm (NIR) Semi variance lag (pixels)
60 Figure 3 6 Land cover map based on spectral bands, N ormalized Difference Vegetation Ind ex layer and first order texture features ( 3 by 3 window variance and 9 by 9 window data range, mean, entropy) CM Tall sawgrass marsh CWP Short sawgrass wet prairie MWP Muhlenbergia wet prairie, SCWP S chizachyrium wet prairie, Woody Woody vege tation
61 Figure 3 7. Land cover map based on spectral bands, NDVI layer and second order texture features (3 by 3 and 9 by 9 window). CM Tall sawgrass marsh CWP Short sawgrass wet prairie MWP Muhlenbergia wet prairie, SCWP S chizachyrium wet pr airie, Woody Woody vegetation
62 Table 3 1. Major plant communities in the study area modified from Ross et al. (2003 and 2006 ) and Richardson et al. (2008) Class name Dominant plant species Tall s awgrass m arsh (CM) Dominated by dense, t all stands of Cladium jamaicense, some Eleocharis cellulosa and Typha domengensis Pers. Short s awgrass wet prairie (CWP) Moderate stands of Cladium jamaicense (<22%) and some Rhynchospora spp., Schoenus nigricans, May also have Schizachyrium rhizomatum Schizachyrium wet prairie (SCWP) Schizachyrium rhizomatum and some sparse/moderate stands of s hort Cladium jamaicense Muhlenbergia wet prairie (MWP) Mix of Muhlenbergia capillaries, Cladium jamaicense Schizachyrium rhizomatum, and s ome Schoenus nigricans Woody vegeta tion (Woody) Dominated by Salix caroliniana, Persea borbonia, Magnolia virginiana Table 3 2. Description of collected ground refer ence data Classes Number of points collected Number of pixels selected for training set Number of pixels selected for accur acy testing Average number of pixels in a training and accuracy polygon CM 48 12536 3237 328 CWP 60 14733 4730 324 SCWP 36 12782 3856 462 MWP 56 12940 3737 297 Woody 56 13439 4685 323 CM Tall sawgrass marsh CWP Short sawgrass wet prairie MWP Muhlenbergia wet prairie, SCWP Schizachyrium wet prairie, Woody Woody vegetation.
63 Table 3 3. Aerial i magery d ata c haracteristics Sensor Microsoft Vexcel UltracamX Acquisition date March 28 April 10, 2009 Flight altitude 4115 90 m eter Spectral resolution Red ( anometer ) Green ( nanometer ) Blue ( nanometer ) Near Infrared ( nanometer ) Spatial resolution 0. 305 meter Data Spectral Bands Derived Texture Features Radiometric resolution 14 bit Area covered by one image 4400 m x 2870 Overlap 60% forward lap 40% side lap
64 Table 3 4. Texture feature calculation formula e. Feature Formula First order statistics Data range Mean ( ) Variance Entropy Second order statistics ** Mean Variance Homogeneity Dissimilarity Entropy Angular second moment Where = grey level of the pixel, = number of pixels in a window and = is the probability that the difference between two adjacent pixels is equal to (Irons and Petersen 1981; Culbert et al. 2009) ** Where is the th entry in the co occurrence matrix (Haralick et al. 1973)
65 Table 3 5. Jeffries Mats ushita (J M) separability analysis using the 4 spectral bands Classes J M value CWP and CM 0.38 CWP and MWP 0.39 MWP and CM 0.66 SCWP and CM 0.84 CWP and SCWP 1.02 MWP and SCWP 1.20 MWP and Woody 1.98 CM and Woody 1.99 CWP and Woody 1.99 SCWP and Woody 1. 99 CM Tall sawgrass marsh CWP Short sawgrass wet prairie MWP Muhlenbergia wet prairie, SCWP Schizachyrium wet prairie, Woody Woody vegetation. Table 3 6. Jeffries Mats ushita (J M) separability analysis using the 4 spectral bands, first and se cond order texture features Classes J M value CWP and MWP 1.40 CWP and CM 1.89 SCWP and CM 1.92 MWP and CM 1.95 CWP and SCWP 1.98 MWP and SCWP 1.99 CM Tall sawgrass marsh CWP Short sawgrass wet prairie MWP Muhlenbergia wet prairie, SCWP S chizachyrium wet prairie, Woody Woody vegetation. Table 3 7. Maximum likelihood classification results using spectral bands and N ormalized D ifference V egetation I ndex (NDVI) layer Class Spectral bands Spectral bands and N DVI layer P A U A P A U A CM 52. 46 73.49 63.80 77.65 CWP 63.57 62.17 52.98 74.19 MWP 70.85 74.20 71.31 71.60 SCWP 76.15 59.78 83.69 58.44 Woody 100.00 100.00 100.00 100.00 Overall accuracy (%) 73.8070 75.4016 Kappa Coefficient 0.6716 0.6915 PA accuracy (%) UA accuracy (%) CM Tall sawgrass marsh CWP Short sawgrass prairie MWP Muhlenbergia wet prairie, SCWP Schizachyrium wet prairie, Woody Woody vegetation. Kappa value closer to 1 means better map accuracy.
66 Table 3 8. Maximum likelihood classificat ion results for spectral bands and first order texture features with different window sizes. Class S B + first order texture 3x3 S B + first order texture 5x5 S B + first order texture 7x7 S B + first order texture 9x9 PA UA PA UA PA UA PA UA CM 12.56 93.61 5.01 100.00 63.93 91.94 70.33 79.92 CWP 79.30 53.92 62.93 58.98 42.50 74.39 81.85 57.03 MWP 70.09 82.53 80.64 44.04 86.03 44.65 75.74 62.63 SCWP 78.78 53.96 61.46 65.03 56.29 74.37 39.75 81.96 Woody 100.00 100.00 100.00 100.00 99.97 100.00 99.92 100.0 0 Overall accuracy (%) 70.1313 64.2338 70.7938 74.0884 Kappa Coefficient 0.6247 0.5501 0.6336 0.6758 SB Spectral bands, PA accuracy (%) UA (%) CM Tall sawgrass marsh CWP Short sawgrass wet prairie MWP Muhlenbergi a wet prairie, SCWP Schizachyrium wet prairie, Woody Woody vegetation. Table 3 9. Maximum likelihood classification results for spectral band s and first order texture features with different window sizes. Class S B + first order texture 11 by 11 S B + f irst order texture 1 5x 1 5 P A U A P A U A CM 68.88 76.21 70.85 66.29 CWP 65.03 69.89 65.94 61.93 MWP 84.87 48.88 83.50 52.24 SCWP 36.03 85.07 26.30 90.92 Woody 99.82 100.00 99.69 100.00 Overall accuracy (%) 71.6379 69.8851 Kappa Coefficient 0.6448 0.62 33 SB Spectral bands, PA accuracy (%) UA (%) CM Tall sawgrass marsh CWP Short sawgrass wet prairie MWP Muhlenbergia wet prairie, SCWP Schizachyrium wet prairie, Woody Woody vegetation.
67 Table 3 10. Maximum lik elihood classification results for spectral bands and second order texture features with small Grey Level Co occurrence Matrix Class S B + second order texture 3x3 S B + second order texture 5x5 S B + second order texture 7x7 S B + second order texture 9x9 P A U A P A U A P A U A P A U A CM 36.12 66.24 68.82 72.04 85.65 71.26 90.41 66.06 CWP 59.20 80.72 68.53 80.21 66.77 76.20 69.31 70.44 MWP 54.76 64.70 56.81 65.50 59.71 61.97 61.67 60.75 SCWP 89.90 55.81 81.22 68.98 70.24 76.05 51.60 78.42 Woody 99.96 91.62 1 00.00 91.60 100.00 93.93 100.00 93.55 Overall accuracy (%) 70.4646 76.3511 76.8963 74.3284 Kappa Coefficient 0.6264 0.7026 0.7107 0.6796 SB Spectral bands, PA accuracy (%) UA (%) CM Tall sawgrass marsh CWP Short saw grass wet prairie MWP Muhlenbergia wet prairie, SCWP Schizachyrium wet prairie, Woody Woody vegetation. Table 3 11. Maximum likelihood classification results for spectral bands and second order texture features with large Grey Level Co occurrence M atrix (11 by 11 and 15 by 15) Class S B + second order texture 11x11 S B + second order texture 15x15 P A U A P A U A CM 89.99 65.19 72.09 62.90 CWP 71.18 68.05 72.59 50.67 MWP 63.76 57.51 66.16 55.02 SCWP 41.70 80.93 29.65 86.61 Woody 99.86 92.92 99.7 9 99.29 Overall accuracy (%) 72.7244 67.6122 Kappa Coefficient 0.6601 0.5966 SB Spectral bands, PA accuracy (%) UA (%) CM Tall sawgrass marsh CWP Short sawgrass wet prairie MWP Muhlenbergia wet prairie, SCWP Sc hizachyrium wet prairie, Woody Woody vegetation.
68 Table 3 12. Maximum likelihood classification results for spectral bands and first order texture features with merged window sizes Class S B + first order texture 3x3 and 9x9 S B + f irst order textural fe atures 3 by 3 variance and 9 by 9 data range, mean, entropy P A U A P A U A CM 88.14 75.53 90.31 79.23 CWP 62.63 87.42 70.22 84.39 MWP 72.71 65.94 72.77 78.79 SCWP 88.14 75.53 85.08 76.45 Woody 100.00 100.00 100.00 100.00 Overall accuracy (%) 81.73 29 84.0192 Kappa Coefficient 0.7712 0.8000 SB Spectral bands, PA accuracy (%) UA (%) CM Tall sawgrass marsh CWP Short sawgrass wet prairie MWP Muhlenbergia wet prairie, SCWP Schizachyrium wet prairie, Woody Wo ody vegetation. Table 3 13 Maximum likelihood classification results for spectral bands and first and second order texture features with merged window sizes Class S B + second order texture 3x3 and 9x9 S B + second order texture 3x3 variance and 9x9 mean, homogeneity, dissimilarity, entropy and ASM S B + 1st and 2nd order 3 by 3 and 9 by 9 window P A U A P A U A P A U A CM 94.36 86.60 87.57 91.90 95.85 84.41 CWP 75.35 81.75 68.21 83.31 71.83 81.98 MWP 65.89 77.93 68.43 73.28 69.21 75.75 SCWP 87.96 76.54 93. 48 73.08 83.78 77.24 Woody 99.97 100.00 100.00 100.00 100.00 100.00 Overall accuracy (%) 84.9279 84.3358 83.9782 Kappa Coefficient 0.8114 0.8040 0.7993 SB Spectral bands, PA accuracy (%) UA (%) CM Tall sawgrass marsh CWP Short sawgrass wet prairie MWP Muhlenbergia wet prairie, SCWP Schizachyrium wet prairie, Woody Woody vegetation.
69 CHAPTER 4 CLASSIFICATION OF WE TLAND COMMUNITIES : USING HIGH RESOLUTION AERIAL IMAGERY AND TESTING ALGORYTHMS Introduction Wetla nds are an essential part of the global ecosystem since they provide habitat for endangered species, supply groundwater aquifers with water and can reduce floods (Mitsch and Gosselink 2007) In recent years, the growth of urban areas combined with intense agriculture and drainage canals has detrimentally disturb ed and pollute d the Everglades ecosystem (Junk et al. 2006; Marella 2009; Richardson 2010) T he Everglades r estor ation program is in progress (Comprehensive Everglades Restoration Plan CERP), and i ts main task is to develop a method to better capture fresh water and convey it to areas where water is scarce before it enters the Atlantic Ocean and Florida B ay. The transferred water will be used to revitalize the ecosystem, help agriculture and provid e an additional drinking water source for south Florida. Freshwater sawgrass marsh and wet prairies cover about 44% of the Everglades (Davis et al. 1994) Sawgrass marshes are characterized by an abundance of Cladium jamaic e nse or are sometimes mixed with other species such as Eleocharis cellulosa and Typha domengensis (Richardson 2010) Wet prairies consist of four major types of grasses : Schizachyrium spp. Muhlenbergia spp. Schoenus spp. and Cladiu m spp. (Olmsted and Armentano 1997) and are named afte r their dominant species. However, wet prairies also support a high diversity of other plant species (Davis et al. 2005) Sawgrass marshes and wet prairies are contin uously changing and shifting due to several factors such as longer dry seasons, loss of or ganic soils, increased fire frequency spread of Woody vegetation and non native and exotic species (Gunderson 1997; Childers et al. 2003; Zweig and Kitchens 2008; Larsen et al. 2011) Continuous
70 and frequent monitoring of these areas is necessary to bette r understand changes in their size distribution and composition (Olmsted and Armentano 1997) According to Armentano et al. (2006) t hese grass communities will be the first to change in response to the CERP restoration process sea level rise, climate cha nge or inputs of chemical fertilizers or pollutants. In addition, it is important to predict changes to flora and fauna habitat to more effectively prevent further degradation of these communities (Armentano et al. 2006) predicted that obs ervable change i n prairie and marsh vegetation in the Everglades National Park (ENP) may occur every few years. Accurate and timely mapping of wetlands is a major interest of many remote sensing profession als, land managers and decision makers (Ozesmi and Bauer 2002) How ever, most of the existing rese arches consist of delineating and discriminating wetlands from other type of land cover using medium or high resolution imagery (Baker et al. 2006; Nielsen et al. 2008) and differentiating between major plant communities such a s grassland vs. forested wetlands (Kindscher et al. 1997; Johansen et al. 2007; Maxa and Bolstad 2009; Bwangoy et al. 2010) Very few studies have attempted to map vegetation communities of small extent (Harvey and Hill 2001; Gilmore et al. 2008; Barbosa 2 010) because of the lack of availability of high resolution satellite and aerial imagery. R ecent effort to classify and map ve getation communities in the ENP was interpretation of 1:40,000 scale color inf rared ( CIR ) film based vertical aerial photographs taken in the mid (Madden et al. 1999) T he ENP conducted a field survey in 2007 to assess the accuracy of
71 Madden's et al. (1999) mapping effort and found significant changes and differences in the p resent vegetation regime compared to the existing map. Ozesmi and Bauer (2002) reviewed most of the literature on satellite remote sensing and found that the main problem in wetland remote sensing is the spec tral separability of vegetation. Schmidt and Sk idmore (2003) also recognize d that mapping major physiognomic or forest type classes was fairly straightforward with good accuracy results, but differentiation among grasses or sedges was difficult. The authors also concluded in their review that identif ic ation of wetland types at a medium spatial resolution was complicated and because of the temporal characteristics of most satellites, it was difficult to capture optimal lighting and water level conditions As a result, high resolution aerial photography is preferred for more detailed and accurate wetland mapping Visual photo interpretation has been used in several wetland mapping studies (Madden et al. 1999; Rutchey et al. 2008; Maxa and Bolstad 2009) with good overall accuracy results H owever it is ext remely costly in terms of time and labor, and moreover, highly depends on the photo interpreter skills and knowledge of the area (Ozesmi and Bauer 2002) Baker et al. (2006) conversely found that multispectral image classification of wetlands has similar accuracy and is more repeatab le relative to human photo interpretation. Studies have shown that automated remote sensing techniques such as Maximum Likelihood (ML), Decision T rees (DT) and Artificial Neural Networks (ANN) (Wang et al. 2004; Baker et al. 2006; Belluco et al. 2006; Wright and Gallant 2007; Sesnie et al. 2008) are effective alternative s to photo interpretation for ma p p ing wetlands.
72 Decision trees and ANNs to a lesser extent have become popular recently because they make no assumption about the statistical distribution of the data (Richards and Jia 2006) Decision tree classifiers have been used to successfully partition the input data into more numerous and homogeneous subsets by producing optimal rules which minimize the error rates in the branches of the tree (Friedl and Br odley 1997; Pal and Mather 2003 ; Yang et al. 2003) In addition, DTs provide a more comprehensive understanding of relationships between objects at different scales of observation or at different levels of detail. The two methods for conducting a decision based classification are manual (Swain and Hauska 1977) and utilization of existing data mining algorithms that are based on different statistical methods that organize incoming data ( i.e. Chi squared Automatic Interactio n Detector CHAID (Kass 1980) Classification and Regression Trees CART (Breiman et al. 1984) C5.0 (Quinlan 1993) Quick, Unbiased, Efficient Statistical Tree QUEST (Loh and Shih 1997) Most algorithms are available commercially as standalone applica tion s or built in to statistical software products (CHAID, CART, C5.0). The QUEST algorithm is easily accessible and free of charge. Artificial Neural Networks have been reported to perform more accurately than other techniques (Goel et al. 2003; Mas and Flores 200 8) for remotely sensed imagery classification especially when the f eature space is complex or the data sets have different statistical distributions Thus, when additional data beside band information is incorporated (Benediktsson et al. 1993) ANN might p erform better than statistical classification techniques. Moreover, ANNs perform supervised classification using less training data than the ML because the rules of recognition of a category are based on
73 the characteristics not only of the training data of this particular category class but also of the other classes (Paola and Schowengerdt 1995) The most commonly used ANN in land cover classification studies is the multi layer perceptron (MLP) neural network based on back propagation (BP) algorithm (Benedi ktsson et al. 1993; Paola and Schowengerdt 1995; Ghedira et al. 2000; Lloyd et al. 2004; Mas 2004; Fuller 2005; Berberoglu et al. 2007) (Ozesmi and Bauer 2002) noted that in most cases, improved classification was obtained when using automated techniques such as ML, DT, ANN methods, and incorporating image texture information and multispectral data in the analysis (Ghedira et al. 2000; Wang et al. 2004; Fuller 2005; Wright and Gallant 2007) Wang et al. (2004) incorporated first and second order texture f eatures into their ML classification model, and their when using a 21 by 21 window size. Wright and Gallant (2007) tried to discriminate five p alustrine wetland types and when image texture features, DEM s and other ancillary GIS data were added to the multispectral data : the overall average error rate was reduced incrementally. However, Michishita et al. (2008) compared a decision tree classifier with ancillary (i.e. DEM and la nd surface temperature and other spectral indices) data against the ML classifier at the wetland plant family level and found that ML performed slightly better than the DT. Artificial n eural network s are also used in landcover mapping (Paola and Schowenger dt 1995; Berberoglu et al. 2007; Ashish et al. 2009) frequently, but literature on wetland mapping using ANNs is limited (Ghedira et al. 2000; Mas 2004; Fuller 2005)
74 T hese previous studies demonstrated that ANN can be used to improve vegetation discrimin ation. However, accuracy r esults involving texture features in the analysis var y. I n a recent land use classification study Ashish et al. (2009) employed probabilistic ANN where 1 meter resolution multispectral image and second order texture features (Har alick et al. 1973) were included in the classification process. The overall accuracy, when image texture was in cluded in the analysis was 89%. Berberoglu et al. (2007) tested several texture features for landcover mapping with ML and ANN classifiers and fo und that ANN outperformed the ML classifier without texture data in the overall classification accuracy when using Landsat spectral bands. The ANN also performed better when texture data was incorporated into the classification; however the overall accurac y slightly decreased when classifying agricultural and semi natural land covers. Ghedira et al. (2000) attempted to distinguish between wetland categories using ANN and texture information with RadarSat data and concluded that an ANN back propagation algor ithm is an accurate tool for image classification. When texture was included in their analysis, the classification process improved significantly. Mas (2004) used Landsat imagery to map land cover classes in a tropical coastal area and employed ANN and anc illary data such as elevation and soil data into his model. When the output from that model was compared with an output from spectral classification, a significant increase of accuracy of land use classification (from 67% to 79%) was detected. Finally, mul tispectral IKONOS imagery was used in Fuller's (2005) study to detect invasive Melaleuca trees, sawgrass and 3 other vegetation classes in South Florida. Back propagation ANN was employed using input layers consisting of the spectral bands as well as a te xture layer and normalized difference vegetation index.
75 The author achieved over 85% in overall accuracy, however visual inspection revealed Overall, m ost of the previ ously mentioned studies were generally based on satellite or hyperspectral imagery and use d high to low spatial resolution data. Additionally, delineation of wetlands from other land cover types was frequently the main objectives, rather than delineatin g w etland and grassland c lasses or species. As a result, t h ere is a need to develop a new vegetation community map for the ENP using rapid and cost effective mapping techniques. Thus the specific objectives of this study are to: 1. Determine whether decision tre e and artificial neural network technologi es can be applied accurately to delineate among the dominant graminoid/sedge communities in the ENP using high resolution aerial imagery and texture data. 2. Develop a methodology that can be utilized in change detec tion studies in the Everglades or other wetlands using high resolution imagery and texture data. Results can be used to rapidly monitor and assess changes in wetland vegetation communities. Methods Study Area The 8.5 km study area is located within the Ev erglades National Park at 25 25 09.09 to 252415.16 N and 804641.46 to 804541.00 W (Figure 4 1) and is composed of mostly graminoid and sedge communities with some woody vegetation present as well. The most important species are sawgrass ( Clad ium jamaicense ), b eaksedge ( Rhynchospora spp), F lorida little bluestem ( Schizachyrium rhizomatum ), muhly g rass ( Muhlenbergia capillaries ), black bogrush ( Schoenus nigricans ) and occasional woody vegetation such as coastal p lain willow ( Salix caroliniana ), r edbay ( Persea borbonia ), and s weetbay ( Magnolia virginiana ) (Ross,
76 Mitchell Bruker, et al. 2006) The study area is very flat, local elevation ranges between 0 and 0.6 meter, and is part of a long term ecological study (Ross et al. 2003; Ross, Mitchell Br uker, et al. 2006) where permanent transects have been established to document and measure species distribution, soil depth and hydrological parameters across the site. Cl assification S ystem and Training Data Many Florida wide vegetation classification sys tems such as the Everglades Vegetation Classification System, Florida Fish and Wildlife Conservation Commission, Florida Gap Analysis Project, Florida Land Use and Cover Classification System and Multi Species Recovery Project (Rutchey et al. 2006) exist However, it appears that most of them do not have the preferred level of detail required by the ENP for (Rutchey et al. 2006) Everglades Vegetation Classification System (EVCS) is the most detailed, containing eight major vegetation types: forest, scrub, savanna, prairies and marshes, shrublands, exotics, additional class headings, and special modifiers (Rutchey et al. 2006) study area and was used a s a foundation for developing classification classes listed in Table 4 1 and Figure 4 2. Training data w ere collected in 2009 short ly after the image acquisition date using real time kinematic GPS unit (TOPCON Hiperlite, minimum positional accuracy of 0.1m ). A total of 86675 pixels, of which 66430 were used for training and 20245 for accuracy assessment were selected based on 256 field plots measure d during May, 2009 The collected point data was randomly distributed in the study area to avoid spatial autoc orrelation Mather (2004) recommended that the number of training data pixels per classificat ion class should be at least 10 30 times the number of feature
77 classes used in stud ies with statistical classifiers. Others suggested that DT and ANN classifiers a re capable of classify ing image data accurately with smaller training set (Hepner et al. 1990; Foody e t al. 1995; Pal and Mather 2003 ) The training and accuracy testing dataset details are presented in Table 4 2. Imagery A Microsoft Vexcel UltracamX (UCX ) digital fixed array camera was utili zed to acquire imagery over the ENP. Microsoft Vexcel UltracamX is a multihead large format sensor and it is capable of provid ing multispectral imagery by collecting red, green, blue (RGB) and near infrared (NIR) chan nels using individual cameras for each band. The imagery was acquired with a solar elevation angle of greater than 30 above the horizon and is cloud and shadow free. The acquisition was done in April 2009 when the wetland communities are in their spring bloom ing conditions. The imagery was geometrically and radiometrically corrected, pan sharpened (U CX system data level 3), but not orthorectified; however, the geometric distortions due to relief change can be ignored given the minimal elevation relief of the study area (Wolf and Dewitt 2000) The horizontal positioning accuracy of the imagery is 3.00m at the 95% confidence interval (RMSE = 1.76m given that positional accuracy = 1.7308 RMSE). Imagery was delivered in the State Plane Coordinate System, Flo rida East Zone 0901, North American Datum of 1983/High Accuracy Reference Network (HARN) and v ertical information is referenced to the North American Vertical Datum of 1988 (NAVD88). Imagery was mosaiced to construct a single image covering the entire stud y area and clipped to save file space and to eliminate the portion of the images beyond the interested area (Figure 4 1). The main characteristics of the imagery are summarized in Table 4 3.
78 Texture Features Spectral features describe the average tonal var iations in various bands of the visible and infrared portion of an electromagnetic spectrum, whereas textural features contain information about the spatial distribution of tonal variations within a band (Haralick et al. 1973) First order texture features are derived using first order statistics of local areas. First order statistics (grey le vel difference method) estimate the probability density function for differences between neighboring pixels (Carr 1999) using a moving pixel window technique. U sing th e m oving window technique in remote sensing results in a new raster file created from a particular spectral band by using the values of surrounding pixels within a designated area (e.g. a window size of 5 row by 5 column has 25 pixels) to compute a new val ue of the middle pixel. After each the procedure again for every pixel in the area of interest. In this study, every pixel had a data range, mean, variance and entropy so, first or der texture value calculated by a window that was moved in the row and column dimensions, one pixel at a time. S econd order statistics describe the frequency with which one gray tone appears in specific rel ationship to another gray tone i n the image and th ey are calculated from the grey level co occurrence matrix (GLCM) which indicates the probability that each pair of pixel values ( i j ) co occur in a given direction and distance (Haralick et al. 1973) These measures are calculated in four directions (hor izontal, vertical, left diagonal, right diagonal) on the image and averaged (Haralick et al. 1973) Classification of Vegetation Communities Three classification techniques were tested Maximum likelihood classification algorithm was used as a base method since it is the most widely used and accepted
79 statistical classification technique (Richards and Jia 2006) Decision tree and ANN classifiers were also employed to develop a high accuracy classification map of the study area. Decision tree In this study, the QUEST algorithm was employed to automatically generate decision tree rules. Classification and Regression Trees (CART) algorithm proposed by Breiman et al. (1984) is the most commonly used the remote sensing, however recent studies (Kim and Loh 2001; S esnie et al. 2008) indicated that QUEST has several advantages over recursive tree construction methods such as CART. Thus, the QUEST algorithm was selected because it is computationally simple and fast, has negligible variable selection bias and produces binary rules (Loh and Shih 1997; Kim and Loh 2001) Kim and Loh (2001) also indicated that when selecting among splitting rules CART is biased in favor of variables that have a large number of possible splits. The QUEST algorithm selects the split variable based on the computed p values from an analysis of variance (ANOVA) F test from the training samples The variable with the smallest p value will be selected as the split predictor for the node. Once, the split predictor is established, t he s plitting poin t is determined by using the quadratic discriminant analysis (QDA). The QDA algorithm calculates the sample mean and variance for the th class ( =1, 2) first. Based on the calculated means and variances the QDA splits the X axis into three intervals ( ), ( ) and ( and selects as split point ( th e one solution between and which is closest to the sample mean for each class (Loh and Shih 1997)
80 Artificial n eural n etwork Artificial neural networks are models that attempt to process information. They try to replicate the functionality and decision making processes of the human brain and acquire the knowledge through learning process es and store the gain ed information in synaptic weights (Haykin 1999) M ulti layer p erceptron BP neural network consist of : (1) input layer nodes where external inputs can be fed into the system, (2) one or more hidden layers of computation nodes and (3) one output layer which produces the results of the classification. In the training ph ase the external input signals consisting of individual band pixel values and texture features were passed through the network via the input layer in a feed forward manner on a layer by layer basis and an output was produced as the actual response of the n etwork. In this forward pass process all synaptic weights were assigned randomly and are fixed. In the last phase of the forward pass process, the result is subtracted from the desired results, which was established by the user (i.e. tall sawgrass marsh) o nce the actual result was generated in the output node. The sum of errors ( E ) is calculated in the following way: where a kd is the actual target value, o kd is the predicted class value; associated with the k th output unit and training example d (Mitchell 1997; Larose 2004) To minimize the error between the actual and the predicted values the output layer sends back the error rate and tr ies to adjust the weights. Once the error rate was calculated ( ), the weights can be updated by :
81 Where w ij represents the weight associated with the i th input to node j and T o calculate the a positi ve constant is introduced ( ) called the learning rate, ranging between 0 and 1 (usually very close to 0) it help s move the network weights toward an optimal solution (Mitchell 1997) and x ij represents the i th input to node j An additional constant, momen tum learning ( is added to the weight update rule ( which is used to speed up and stabilize convergence (Mitchell 1997) By employing the equation, the weight update on the n th iteration partially depend s on the update that occurred duri ng the previous iteration. Normally influence of previous adjustments (Mitchell 1997) In this study the learning rate was set to 0. 2 and the momentum learning to 0. 3 (Berberoglu et al. 2007) The following stage, called backward pass consists of the output error being propagated back to the hidden layer. The synaptic weights are then adjusted to make the final output nearer to the expected output (i.e. training classes) (Haykin 1999; L arose 2004) Once the training session is done, the network saves the calculated weights for each input training set. The model use s these weights to classify the new input layers. Evaluation of Classification Accuracy and Algorithm Comparison The results of the three employed classification algorithms were evaluated using a confusion matrix. The confusion matrix was calculated based on part of the field measured dataset (Table 4 2), which was split into a training and accuracy assessment dataset. The accur acy dataset (about 1/3 of the collected ground reference data) was
82 measures the accuracy of the classified image by computing the number of pixels which the output map in the field, calculating the number of pixels which were incorrectly committed to an incorrect class by the classifier T accuracies of the developed maps were allocated by five categories : very high accuracy ( > 90 % ), high accuracy (80 % 90% ), acceptable accuracy (70 % 80% ), low accuracy (50 % 70 % ), and poor accuracy ( < 50 % ). The Kappa coefficient estimates the reduction in error generated by the classifier versus the error of a random classifier. Where stand for the proportion (Canty 2010) The k appa coefficient range s between 0 and 1, where 1 indicates a perfect classification Z tests were calculated to test that the differ sign ificantly (Donner et al. 2000) The z tests were based on the following formula: where corresponds to the estimated covariance between and (Donner et al. 2000) To determine whether there is a difference between 2 kappa coefficients the null hypothesis of no significance would be rejected at 5% level of significance if |z| >1.96 (Congalton 1991) Results Based on the results of Chapter 3, the most promising textural features and window sizes were chosen to validate the classification algorithms accuracy for wetland
83 graminoid/sedge communities mapping. The tested spectral bands and textural features were the following: spectral bands (red green blue near infrared), NDVI, first order texture features (data range, mean, variance and entropy) calculated by using the 9 by 9 moving window, second order texture features (mean, variance, homogeneity, dissimilarity, entropy and ASM) calculated by using the 7 by 7 GLCM, merged first order texture features (3 by 3 variance and 9 by 9 data range, mean and entropy), merged 3 by 3 and 9 by 9 GLCM second order texture features, and merged first and second order texture features (3 by 3 and 9 by 9). When spectral bands and the NDVI layer were tested with the different classifiers, the results were in a comparable range among all the tested algorithms (Table 4 4). The artificial neural network classifier (implemented using ITT ENVI v. 4.4) yielded the best overall accuracy of the 3 algorithm s when using only the spectral bands for the classification, however, the result produced by the ML classifier was statistically not different ( The QUEST algorithm ( implemented using Rule G en extension v. 1.02 in ITT ENVI v. 4. 4 ) produced a lower overall accuracy with a significantly lower ( 0.05) kappa coeffic ient (0.63). When the NDVI layer was added to the spectral bands, the overall accuracy values improved in the case of ML and ANN to 75.40% and First and second order texture layers were also included in the analysis to further increase the accuracy of the maps. According to Chapter 3, the most promising texture features and window sizes to increase the map accuracy are the data range, mean, variance and entropy in the first or der texture measures, while the mean, variance, homogeneity, dissimilarity, entropy and angular second moment performed well in the
84 second order texture features. The most appropriate window size to discriminate the graminoid and s edge communities are the 9 by 9 pixels moving window in the firs t order texture and the 7 by 7 pixels GLCM in the second order texture when the texture measures are used without merge. When merging diff erent window sizes, the 3 by 3 pixels and 9 by 9 pixels with the first order te xture moving window sizes as well as for the second order texture GLCM are the most suitable (Chapter 3). Moreover, when merging the most ap propriate window sizes (3 by 3 pixels and 9 by 9 pixels) in the first order texture features the variance is calcula ted with the 3 by 3 pixels moving window size while the others (data range, mean and entropy) are calculated with the 9 by 9 pixels moving window size (Chapter 3). The map accuracy significantly increased compared to the ML classifier result ( 5 and 4 10), w hen the QUEST algorithm (DT) and the back propagation ANN were employed using the 9 by 9 pixels moving window siz e to develop the vegetation map. The QUEST algorithm yielded an 84.65% overall map accuracy, with very high det ection values for the CM (91.59%) MWP (87.57%) and Woody (100%) classes. The back propagation ANN algorithm further increased the map accuracy to 88.64%, with an excellent detection rate for the CM (97.87%) and Woody (100%) classes and with high detection rate for the MWP (83.44%) and SCWP (81.5%) classes. The comparison of the Kappa coefficients of the QUEST and the BP ANN algorithm also revealed that the BP ANN algorithm delivered a significantly more accurate map ( 0.05) (Table 4 10). When the second order tex ture features using the 7 by 7 pixels GLCM were tested, the QUEST algorithm yielded a similar overall accuracy value (76.70%) to the
85 ML classifier (76.89%). The ANN classifier yielded the highest overall accur acy with 81.29%, which is significantly higher than the other classifiers ( 0.05) (Table 4 10). The first tested dataset when the different window sizes were merged was the spectral bands, NDVI and first order texture moving window size (3 by 3 pixels var iance and 9 by 9 pixels data range, mean and entropy) data (Table 4 7). There was no statistical difference in the overall accuracy of the map, produced by the ML and DT classifiers, while the ANN classifier produced a significa ntly higher overall accuracy (89.49%) 0.05) compared the ML and DT classifier (Table 4 10) Spectral bands, NDVI and second order texture features (3 by 3 pixels and 9 by 9 pixels GLCM) were also merged and tested (Table 4 8). The QUEST and ANN algorithms yielded a significantly higher ove rall accuracy (86.89% and 87.95%, respectively) than the ML classifier. The QUEST algorithm delivered very high detection 80% of all classes. The four spectral bands, NDVI and first and second order texture featur es calculated using the 3 by 3 pixels and 9 by 9 pixels moving window and GLCM were merged ( n= 45 layers) and tested. The overall accuracy of the DT and ANN classifiers were higher than the ML classifier, ho wever, K z tests revealed that the difference was not significantly different (Tables 4 9 and 4 10). Discussion This study was based on the assumption that DT and ANN classifiers can improve classification accuracy for graminoid/sedge wetland com munities compared to the M L classifier (Pal and Mather 2003 ; Berberoglu et al. 2007) Moreover, the inclusion of
86 textural features beside spectral bands was considered an additional factor in the premise of achieving higher accuracy maps. R esults demonstra ted that ANN BP classifier produced significantly more accurate 0.05) maps in most of the examined cases where ML or QUEST algorithms were used However, when only spectral bands and NDVI layer w ere included in the analysis, the ANN and ML algorithms pr oduced maps with comparable accuracy (74.53% vs. 75.40%) while the QUEST algorithm underperformed compared to the ANN and ML classifiers. In fact, the QUEST DT algorithm produced less accurate maps in all but one trial in this study. The best result was pr oduced by the QUEST algorithm using the spectral bands, NDVI layer and the merged 3 by 3 and 9 by 9 window second order texture fe atures. However, even with this high overall accuracy, the result ing maps were not statistically more accurate than those of p roduced by the ML or the ANN BP classifiers. The QUEST algorithm significantly improved the accuracy of the vegetation map only once versus the ML classifier, when the spectral bands NDVI, and the first order texture features (9 by 9 window) were included Similarly to Sesnie et al. ( 2008) the QUEST algorithm produced low overall accuracy when only the spectral bands wer e included in the mapping (71%) and further decreased when only the NDVI layer was added (69%), but results increased when textural inform ation was incorporated. Sesnie et al. (2008) achieved 93% overall accuracy when using 12 spectral and geospatial layers, however, they used low spatial resolution imagery (Landsat TM), and also noted that overall accuracy result increased when less heterog eneous land cover was mapped. In this study, the highest overall accuracy for the QUEST algorithm was 86.89%, which is lower than Sesnie et al. ( 2008) achieved H owever, the ir wetland
87 c ategories were less detailed consisting only of palm swamps, forested s wamps and herbaceous swamps Artificial neural network classifier consistently produced high accuracy results for the study area. Yet, when only the spectral bands and NDVI layer w ere included in the classification, the output was not significantly better than the ML classifier. However, as texture features were added in the classification, the produced vegetation map s accuracies increased significantly in all trials but one, compared to the ML and DT classifiers. Also, the highest overall accuracy (89.49%) was achieved when using the ANN with spectral bands, NDVI layer and merged first order texture features (3 by 3 pixel window variance and 9 by 9 pixel window data range, mean, entropy). The second highest overall accuracy was produced with the combination of spectral bands, NDVI layer and a first order texture features 9 by 9 moving window (88.64%). Th is set of data contained the same number of layers (9), however, the pre processing of the data was simpler and faster, as only one moving window size was ca lculated and merged with the spectral bands and NDVI layer. The consistently high accuracy results can be a ttributed obtain accurate outputs for new input data tha t is different from the training data (Zhang et al. 2003) Berberoglu et al. (2004) also noted that the generalization capability of the ANN is a key advantage against the ML classifier, especially in areas where the spectral difference among vegetation co ver attributes is small. results reveal that the ANN algorithm detect ed the classes with the highest accuracy. For example, ANN curacy for all
88 classes when used to classif y the spectral bands, NDVI layer and first order texture features (3 by 3 window variance and 9 by 9 window data range, mean and entropy). The W oody vegetation class was always detected easily and accurately whic h is understandable based on the very different spectral signature from the other classes. is evident when individual class accuracies are reviewed. The ANN classifier generated increased to 82.06% when the 3 by 3 window size first order texture variance was included beside the 9 by 9 window size. This can be due to the heterogeneity of CWP class, where several plant speci es compose the community. The MWP and SCWP classes also showed similar trends, in the ML and DT classification but they were more consi stent with the ANN classifier. When the second order texture features (7 by 7 window) were tested together with the spectral bands and classification algorithms, which implie s that the 7 by 7 window second order texture measures did not characterized the communities accurately enough. However, similarly to the first order texture windows, when the 3 by 3 and 9 by 9 GLCM window second order texture measures were merged with the spectral bands and NDVI layer, the classification results increased with all the employed classification algorithms, but still they were lower compared to the merged first order texture features (3 by 3 and 9 by 9 window), spectral bands and NDVI layer. M oreover, when the second order texture features were merged with the spectral bands, NDVI layer and first order texture
89 decreased similarly to the overall accuracy, for the communities that used the DT and the ANN classifier compared to the merged layers of spectral bands, NDVI layer and first order texture measures.
90 Figure 4 1. UltacamX mosaiced image of the study area in the Everglades National Park The scale bar refers to the image.
91 A. B. C. D. E. Figure 4 2. Example images of the classification classes. A. Tall sawgrass marsh B. Short sawgrass wet prairie C. Schizachyrium wet prairie D. Muhlenbergia wet prairie E. W oody vegetation
92 Figure 4 3. Plant community maps of the stu dy area in the Everglades National Park using different classification algorithms. Spectral bands, NDVI layer and second order texture features (3 by 3 and 9 by 9 window sizes) for the m aximum likelihood and d ecision tree classifiers while spectral bands, NDVI layer and first order texture features (3 by 3 window variance and 9 by 9 window data range, mean, entropy ) for the artificial neural network classifier were used. CM Tall sawgrass marsh CWP Short sawgrass wet prairie MWP Muhlenbergia wet pra irie, SCWP S chizachyrium wet prairie, Woody Woody vegetation
93 Table 4 1. Major plant communities in the research area. Class name Dominant plant species Tall s awgrass m arsh (CM) Cladium jamaicense, some Eleocharis cellulosa and Typha domengensis Pers Short s awgrass wet prairie (CWP) Cladium jamaicense, some Rhynchospora spp., Schoenus nigricans, Schizachyrium rhizomatum Schizachyrium wet prairie (SCWP) Schizachyrium rhizomatum Cladium jamaicense Muhlenbergia wet prairie (MWP) Muhlenbergia capilla ries, Schizachyrium rhizomatum, Schoenus nigricans, Cladium jamaicense Woody vegetation (Woody) Salix caroliniana, Persea borbonia, Magnolia virginiana Table 4 2. Description of colle cted ground reference data Classes Numbe r of points collected Number of pixels were selected for training Number of pixels selected for accuracy testing Average number of pixels in a training/accuracy polygon CM 48 12536 3237 328 CWP 60 14733 4730 324 SCWP 36 12782 3856 462 MWP 56 12940 3737 297 Woody 56 13439 4685 32 3 CM Tall sawgrass marsh CWP Short sawgrass wet prairie MWP Muhlenbergia wet prairie, SCWP Schizachyrium wet prairie, Woody Woody vegetation.
94 Table 4 3. UltracamX i magery d ata c haracteristics Sensor Microsoft Vexcel UltracamX Acquisition d ate March 28 April 10, 2009 Flight altitude 4115 90 meter Spectral resolution Spatial resolution 0.305 meter Data Spectral Ban ds Derived Texture Features Radiometric resolution 14 bit Area covered by one image 4400 m x 2870 Overlap 60% forward lap 40% side lap Table 4 4. Classification results based on spectral bands and the NDVI layer Spectral bands Spectral bands and N DVI layer Maximum likelihood Overall accuracy (%) 73.8070 75.4016 Decision Tree Overall accuracy (%) 71.1348 68.8957 Artificial Neural Network Overall accuracy (%) 74.4250 74.5316
95 Table 4 5. Spectral bands, NDVI and first order texture features ( 9 pixels by 9 pixels moving window) classification results. Class ML DT ANN PA UA PA UA PA UA CM 70.33 79.92 91.59 91.78 97.87 89.49 CWP 81.85 57.03 62.06 67.32 70.31 73.48 MWP 75.74 62.63 87.57 66.15 83.44 77.22 SCWP 39.75 81.96 69.81 97.79 81.50 95.25 Woody 99.92 100.00 100.00 97.13 100.00 99.17 Overall accuracy (%) 74.0884 84.6514 88.6412 PA (%) UA (%) ML maximum likelihood classifier, DT decision tree classifier, ANN multi layer perceptron back p ropagation artificial neural network classifier, CM Tall sawgrass marsh CWP Short sawgrass wet prairie MWP Muhlenbergia wet prairie, SCWP Schizachyrium wet prairie, Woody Woody vegetation. Table 4 6. Spectral bands and NDVI and second order te xture features (7 pixels by 7 pixels GLCM) classification results. Class ML DT ANN PA UA PA UA PA UA CM 85.65 71.26 57.37 80.93 69.32 80.43 CWP 66.77 76.20 62.60 54.66 60.09 81.21 MWP 59.71 61.97 89.86 59.46 89.26 61.92 SCWP 70.24 76.05 68.86 95.7 3 75.09 89.75 Woody 100.00 93.93 94.58 94.31 99.24 97.80 Overall accuracy (%) 76.8963 76.7023 81.2926 PA (%) UA (%) ML maximum likelihood classifier, DT decision tree classifier, ANN multi layer perceptron b ack propagation artificial neural network classifier, CM Tall sawgrass marsh CWP Short sawgrass wet prairie MWP Muhlenbergia wet prairie, SCWP Schizachyrium wet prairie, Woody Woody vegetation.
96 Table 4 7. Spectral bands and NDVI and first or der texture features (3 pixels by 3 pixels variance and 9 pixels by 9 pixels data range, mean, entropy moving window) classification results. Class ML DT ANN PA UA PA UA PA UA CM 90.31 79.23 92.20 91.37 97.37 95.39 CWP 70.22 84.39 59.28 66.90 82.06 74.94 MWP 72.77 78.79 86.97 63.93 82.84 75.93 SCWP 85.08 76.45 66.56 96.66 80.51 94.99 Woody 100.00 100.00 100.00 97.13 99.88 99.56 Overall accuracy (%) 84.0192 8 3 6031 89.4962 PA (%) UA (%) ML maximum likelih ood classifier, DT decision tree classifier, ANN multi layer perceptron back propagation artificial neural network classifier, CM Tall sawgrass marsh CWP Short sawgrass wet prairie MWP Muhlenbergia wet prairie, SCWP Schizachyrium wet prairie, Woody Woody vegetation. Table 4 8. Spectral bands, NDVI and second order texture features merged 3 pixels by 3 pixels and 9 pixels by 9 pixels GLCM window classification results. Class ML DT ANN PA UA PA UA PA UA CM 94.36 86.60 95.04 92.46 87.04 91.58 CWP 75.35 81.75 75.07 66.85 86.55 79.75 MWP 65.89 77.93 87.02 74.12 83.69 71.62 SCWP 87.96 76.54 70.62 97.81 80.14 94.82 Woody 99.97 100.00 100.00 97.13 99.64 99.36 Overall accuracy (%) 84.9279 86.8906 87.9593 PA (%) UA (%) ML maximum likelihood classifier, DT decision tree classifier, ANN multi layer perceptron back propagation artificial neural network classifier, CM Tall sawgrass marsh CWP Short sawgrass wet prairie MWP Muhlenbergia wet pr airie, SCWP Schizachyrium wet prairie, Woody Woody vegetation.
97 Table 4 9. Spectral bands, NDVI and first and second order texture features 3 pixels by 3 pixels and 9 pixels by 9 pixels moving window and GLCM classification results. Class ML DT ANN PA UA PA UA PA UA CM 95.85 84.41 94.73 90.00 87.95 87.24 CWP 71.83 81.98 56.68 61.42 76.59 76.18 MWP 69.21 75.75 89.31 70.79 82.99 70.90 SCWP 83.78 77.24 70.17 97.31 77.35 96.40 Woody 100.00 100.00 100.00 97.13 100.00 97.17 Overall accuracy (%) 83 .9782 85.1094 86.3308 PA (%) UA (%) ML maximum likelihood classifier, DT decision tree classifier, ANN multi layer perceptron back propagation artificial neural network classifier, CM Tall sawgrass marsh CW P Short sawgrass wet prairie MWP Muhlenbergia wet prairie, SCWP Schizachyrium wet prairie, Woody Woody vegetation.
98 Table 4 10 Kappa confidences and significance testing based on the z score for maps using various combinations of input layers a nd classifiers. Comparison of Kappa Coefficients (Standard error) Classifier Classifier Classifier ML DT ML ANN DT ANN Model SB 0.6716** (0.0042) 0.6303 (0.0072) 0.6716 0.6740 (0.0069) 0.6303 0.6740** SB+NDVI 0.6915** (0.0041) 0.6013 (0.0073) 0.69 15 0.6755 (0.0069) 0.6013 0.6755** SB+NDVI+FO 9x 9 0.6758 (0.0042) 0.8055 ** (0.0046) 0.6758 0.8560 ** (0.0041) 0.8055 0.8560 ** SB+NDVI+SO 7x 7 0.7107 (0.0036) 0.7056 (0.0054) 0.7107 0.7622 ** (0.0050) 0.7056 0.7622 ** SB+NDVI+FO 3x3 variance, 9x 9 data range, mean, entropy 0.8000 (0.0035) 0. 7921 (0.0047) 0.8000 0.8671 ** (0.0039) 0.8055 0.8671 ** SB+NDVI+SO 3x 3 a nd 9x 9 0.8114 (0.0034) 0.8343 (0.0043) 0.8114 0.8476 ** (0.0042) 0.8343 0.8476 SB+NDVI+FO 3 x 3 and 9 x 9+SO 3 x 3 and 9 x 9 0.7993 (0.0035) 0.8113 (0.0046) 0. 7993 0.8268 (0.0044) 0.8113 0.8268 SB spectral bands, NDVI normalized difference vegetation index, FO first order texture features, SO second order texture features, 3 x 3 3 by 3 pixels window size, 9 x 9 9 by 9 pixels window size, ML maximum li kelihood classifier, DT decision tree classifier, ANN neural network classifier. ** s ignificant ly different based on the z test. Differences are significant at the 95% confidence level (|z| >1.96) (Congalton 1991) No asterisk means not statistically different.
99 CHAPTER 5 CONCLUSIONS This study has shown that the spatial heterogeneity of different vegetation cover in a wetland environment can be mapped accurately using high resolution aerial imagery. Beside spectral bands, inclusion of an NDVI layer derived texture features and classification algorithms improved the accuracy of the derived map. The results of this study support the use of first order texture features in high accuracy mapping of very complex, heterogeneous wetlands plant communities. Results also show that first order texture features have a higher impact on the increased mapping accuracy when using ANN than the second order texture features. Second comm unities as the first order texture features, but the overall accuracy tended to be lower. Also, it is simpler to calculate first order texture features than employing the GLCM for calculating second order texture features. When first and second order textu re features and NDVI were included, beside spectral bands, into the classification; overall accuracy increased. A l s o, the moving window size is a very important consideration where the The applied window size must depend on the communities of interest When the semivariogram method was used to estimate the most appropriate window size, its results generally indicated the proper window size. However, trials with other window sizes revea led that the window size estimated using the semivariogram did not necessary generate the highest accuracy map Thus, a careful selection is very important when choosing the right window sizes for texture derivation.
100 In this study t wo plant community clas ses reacted very differently to changes. The Tall sawgrass class (CM) had very low accuracy values when a small window size has been used to calculate first order texture features while the Schizachyrium wet prairie (SCWP) class had high prod moving window size was gradually increased, the accuracy of detecting CM class increased and decreased in the case of the SCWP class. Thus, when more complex classes are involved in the classification, smaller window size might be more appropriate. The classification algorithms produced different results. The highest classification accuracy was achieved with the ANN classifier (89.49%), which is very close the 90% accuracy standard that is expected by the RECOVER monitoring and as sessment plan (RECOVER 2004) The DT classifier, when texture measures were included produced similar or higher overall accuracies than the ML classifier H owever its computing time was significantly higher. When only spectral bands were tested, the ML cl assifier had similar results as the ANN or DT classifiers, but overall, maps based on the spectral bands and NDVI layer only had low overall accuracies. Future research should focus on the inclusion of ancillary data such as soil depth, high resolution dig ital elevation model s and hydrological variables to further improve classification accuracies. Additionally, the developed methodology allows for not only the monitoring of the investigated plant communities, but other communities of interest could be incl uded as well such as the native but spreading cattail ( Typha spp.) or monitoring of invasive plant species such as m elaleuca ( Melaleuca quinquenervia ) or Brazilian pepper ( Schinus terebinthifolious )
101 H igh spatial resolution imagery is an important factor when classifying plant communities. This is especially important since s ome of the plant communities in this study occup ied small localized patches, which can not be delineate d with lower resolution imagery. The developed methodology is repeatable, thus i t can be used with a new set of imagery for change detection studies. The thematic vegetation map can also serve to better analyze spatial patterns of the studied vegetation communities, identify habitats of interest and help determine wetland areas that a re susceptible to change.
102 LIST OF REFERENCES Armentano TV, Sah JP, Ross MS, Jones DT, Cooley HC, Smith CS (2006) Rapid responses of vegetation to hydrological changes in Taylor Slough, Everglades National Park, Florida, U SA. Hydrobiologia 569: 293 309 Ashish D, McClendon RW, Hoog enboom G (2009) Land use classification of multispectral aerial images using artificial neural networks. International Journal o f Remote Sensing 30: 1989 2004 Baker C, Lawrence R, Montagne C, Patten D (2006) Mapping wetlands and riparian areas using Land sat ETM+ imagery and decision tree bas ed models. Wetlands 26: 465 474 Balaguer A, Ruiz LA, Hermosilla T, Recio JA (2010) Definition of a comprehensive set of texture semivariogram features and their evaluation for object oriented image classification. Com pu ters & Ge osciences 36: 231 240 Baraldi A, Parmiggiani F (1995) An Investigation of the Textural Characteristics Associated with Gray Level Cooccurrence Matrix Statistical Parameters. IEEE Transactions on Geoscience and Remote Sensing 33: 293 305 Barbo sa I (2010) Mapping Wetland Environments in the Brazilian Savannah From High Resolution IKONOS Image Data. In: Wagner W, Szkely B (eds) ISPRS TC VII Symposium pp. 62 67 Belluco E, Camuffo M, Ferrari S, Modenese L, Silvestri S, Marani A, Marani M (2006) Mapping salt marsh vegetation by multispectral and hyperspectral remote sensing. Remote se nsing of environment 105: 54 67 Benediktsson JA, Swain PH, Ersoy OK (1993) Conjugate gradient neural networks in classification of multisource and very high dimensi onal remote sensing data. International Journal o f Remote Sensing 14: 2883 2903 Berberoglu S, Yilmaz KT, Ozkan C (2004) Mapping and monitoring of coastal wetlands of Cukurova Delta in the Eastern Mediterranean region. Biodiversity and Conservation 13: 61 5 633 Berberoglu S, Curran PJ, Lloyd CD, Atkinson PM (2007) Texture classification of Mediterranean land cover. International Journal of Applied Earth Observatio n and Geoinformation 9: 322 334 Bernhardt CE, Willard DA (2009) Response of the Everglades ridge and slough landscape to climate variability and 20th century water management. Ecolog i cal Applications 19: 1723 1738 Breiman L, Friedman J, Stone C, Olshen R (1984) Classification and regression t rees. Chapman & Hall, New York N Y USA
103 Brinson MM, Malv rez AI (2002) Temperate freshwater wetlands: types, status, and threats. Environ mental Conservation 29: 115 133 Bullock A, Acreman M (2003) The role of wetlands in the hydrological cycle. Hydrology and E arth System Sciences 7: 358 389 Bwangoy JRB, Hanse n MC, Roy DP, Grandi GD, Justice CO (2010) Wetland mapping in the Congo Basin using optical and radar remotely sensed data and derived topographical indices. Remote Se nsing of Environment 114: 73 86 Canty MJ (2010) Image Analysis, Classification, and Chan ge Detection in Remote Sensing: With Algorithms for ENVI/IDL, Second Edition CRC Press, Boca Raton, F L, USA Carr J (1999) Classification of digital image texture using variograms. In: Atkinson PM, Tate NJ (eds) Advances in remote sensing and GIS analysis. Wiley, West Sussex, UK, pp. 135 146 Carr JR, de Miranda FP (1998) The Semivariogram in Comparison to the Co Occurrence Matrix forClassification of Image Texture. IEEE Transactions on Geoscience a nd Remote Sensing 36: 1945 1953 Childers DLD, Jones RF, No e R, Rugge GB, Scinto M, Leonard J (2003) Decadal change in vegetation and soil phosphorus pattern across the Everglades landscape. Journal of En vironmental Quality 32: 344 362 Clausi DA (2002) An analysis of co occurrence texture statistics as a function of grey level quantization. Canadian Jour nal of remote sensing 28: 45 62 Coburn CA, Roberts ACB (2004) A multiscale texture analysis procedure for improved forest stand classification. International Journal o f Remote Sensing 25: 4287 4308 Congalton RG (1991) A review of assessing the accuracy of classifications of remotely sensed data. Remote s ensing of environment 37: 35 46 Craft C, Vymazal J, Richardson C (1995) Response of everglades plant communities to nitrogen and phosphorus additions. Wetlands 1 5: 258 271 Curran PJ (1988) The semivariogram in remote sensing: an introduction. Remote Sen sing of Environment 24: 493 507 Dahl TE (2006) Status and trends of wetlands in the conterminous United States 1998 to 2004. U.S. Department of the Interior; Fish and Wildlife Service, Washington, D.C. USA
104 Davis SM (1994) Phosphorus inputs and vegetation sensitivity in the Everglades. In: Davis SM, Ogden JC (eds) Everglades: The Ecosystem and Its Restoration. St. Lucie Press, Delray Beach, FL, USA, pp. 357 378 Davis S M, Ogden JC (eds) (1994) Everglades: The Ecosystem and Its Restoration. St. Lucie Press, Delray Beach FL USA Davis SM, Gaiser EE, Loftus WF, Huffman AE (2005) Southern marl prairies conceptual ecologi cal model. Wetlands 25: 821 831 Davis SM, Gunderson LH, Pa rk WA, Richardson JR, Mattson JE (1994) Landscape dimension, composition, and function in a changing Everglades ecosystem. In: Davis SM, Ogden JC (eds) Everglades: The Ecosystem a n d Its Restoration., pp. 419 444 i automatic choice of scale dependent features for satellite SAR image classification. Pattern recognition letters 27: 244 251 Dillabaugh K, King D (2008) Riparian marshland composition and biomass mapping using Ikonos imagery. Canadian Journal of Remote Sensing 34: 143 158 Donner A, Shoukri MM, Klar N, Bartfay E (2000) Testing the equality of two dependent kappa statistics. Sta tistics in Medicine 19: 373 387 Doren RF, Armentano TV, Whiteaker LD, Jones RD (1997) Marsh vegetation patterns and soil phospho rus gradients in the Everglades ecosys tem. Aquatic Botany 56: 145 163 Doren RF, Rutchey K, Welch R (1999) The Everglades: a perspective on the requirements and applications for vegetation map and database products. Photogrammetric Engineering and Remote S ensing 65: 155 161 Duever M (2006) South Florida Vegetation Classification Scheme. Crosswalks. http:/ /ufdc.ufl.edu/UF00066304/00001 A ccessed 13 Jun 2011 Duever MJ (2005) Big Cypress regional ecosystem conceptual ecologi cal model. Wetlands 25: 843 853 Er win KL (2008) Wetlands and global climate change: the role of wetland restoration in a changing world. Wetlands E cology and Management 17: 71 84 Foody GM (2004) Thematic map comparison: evaluating the statistical significance of differences in classificat ion accuracy. Photogrammetric Engineering and Remote Sensing 70: 627 634
105 Foody GM, McCulloch MB, Yates WB (1995) The effect of training set size and composition on artificial neural network classification. International Journal of Remote Sensing 16: 1707 1723 Franklin SE, Wulder MA, Lavigne MB (1996) Automated derivation of geographic window sizes for use in remote sensing digital image texture analysis. Compu ters & Geosciences 22: 665 673 Friedl MA, Brodley CE (1997) Decision tree classification o f land cover from remotely sensed data. Remote Sen sing of Environment 61: 399 409 Fuller DO (2005) Remote detection of invasive Melaleuca trees (Melaleuca quinquenervia) in South Florida with multispectral IKONOS imagery. International Journal of Remote S ensing 26: 1057 1063 Ge S, Carruthers R, Gong P, Herrera A (2006) Texture Analysis for Mapping Tamarix parviflora Using Aerial Photographs along the Cache Creek, California. Environmental Monito ring and Assessment 114: 65 83 Ghedira H, Bernier M, Ouard a TBMJ (2000) Application of neural networks for wetland classification in RADARSAT SAR imagery. In: Geoscience and Remote Sensing Symposium, 2000. Proceedings. IGARS S 2000. IEEE 2000 International pp. 675 677 Gilmore MS, Wilson EH, Barrett N, Civco DL, Prisloe S, Hurd JD, Chadwick C (2008) Integrating multi temporal spectral and structural information to map wetland vegetation in a lower Connecticut River tidal marsh. Remote Sensin g of Environment 112: 4048 4060 Gluck M, Rempel R, Uhlig P (1996) An eva luation of remote sensing for regional wetland mapping applications. Forest Research Report, Ontario Forest Research Institute, Sault Ste Marie, Ontario, Canada Goel PK, Prasher SO, Patel RM, Landry JA, Bonnell RB, Viau AA (2003) Classification of hyperspe ctral data by decision trees and artificial neural networks to identify weed stress and nitrogen status of corn. Computers and Elect ronics in Agriculture 39: 67 93 Goodin DG, Henebry GM (1997) A technique for monitoring ecological disturbance in tallgrass prairie using seasonal NDVI trajectories and a discriminant function mixture model. Remote Sens ing of Environment 61: 270 278 Gunderson LH (1997) Vegetation of the Everglades: determinants of community composition. In: Davis SM, Ogden JC (eds) Everglade s: The Ecosystem and Its Restoration. CRC Press, Baca Raton, FL, pp. 323 340
106 Gunderson LH, Loftus WF (1993) The Everglades. In: Martin WH, Boyce SG, Echternacht AC (eds) Biodiversity of the Southeastern United States. John Wiley and Sons, New York, NY, USA. pp 199 255 Hancock PJ, Boulton AJ, Humphreys WF (2005) Aquifers and hyporheic zones: towards an ecological understanding of groundwater. Hydrogeology Journal 13: 98 111 Haralick RM, Dinstein I, Shanmugam K (1973) Textural features for image classification IEEE Transactions on systems, man, and cybernetics 3: 610 621 water chemistry, and biological communities of a floodplain ecosystem, Everglades, South Florida, USA. Hydroge ology Journa l 17: 185 201 Harvey KR, Hill GJE (2001) Vegetation mapping of a tropical freshwater swamp in the Northern Territory, Australia: A comparison of aerial photography, Landsat TM and SPOT satellite imagery. International Journal of Remote Sensing 22: 2911 2925 Haykin S (1999) Neural networks : a comprehensive foundation. Pr entice Hall, Upper Saddle River N J USA Hepner G, Logan T, Ritter N, Bryant N (1990) Artificial Neural Network Classification Using a Minimal Training Set: Comparison to Conventional Supervised Classification. Photogrammetric Engineering and Remote Sensing 56: 469 473 Johansen K, Coops NC, Gergel SE, Stange Y (2007) Application of high spatial resolution satellite imagery for riparian and forest ecosystem classification. Remote se nsing of Environment 110: 29 44 Joria P (2001) A Review of Present Applications of Remote Sensing Technology for the Mapping of Large River Floodplains and Wetland Cover Types. U.S. Geological Survey, Upper Midwest Environmental Sciences Center, La Crosse, W I, USA Journal Citation Reports Thomson Reuters Web of Knowledge. http://wokinfo com/ A ccessed 23 Apr 2011 Junk WJ, Brown M, Campbell IC, Finlayson M, Gopal B, Ramberg L, Warner BG (2006) The comparative biodiversity of seven globally import ant wetlands: a synthesis Aquatic Sciences 68: 400 414 Kass GV (1980) An Exploratory Technique for Investigating Large Quantities of Categorical Data. Applied Statistics 29: 119 127 Kim H, Loh WY (2001) Classification trees with unbiased multiway split s. Journal of the American Stat istical Association 96: 589 6 04
107 Kindscher K, Fraser A, Jakubauskas ME, Debinski DM (1997) Identifying wetland meadows in Grand Teton National Park using remote sensing and average wetland values. Wetlands Ecology and Managem ent 5: 265 273 Knight JF, Lunetta RS (2003) An experimental assessment of minimum mapping unit size. IEEE Transactions on Geoscience an d Remote Sensing 41: 2132 2134 Kushlan JA (1990) Freshwater marshes. In: Myers RL, Ewel JJ (eds) Ecosystems of Florida University of Central Florida Press, Orlando, FL, pp. 324 363 Larose DT (2004) Discovering Knowledge in Data: An Intr oduction to Data Mining. John Wiley & Sons Ltd, Chichester, UK Larsen L, Aumen N, Bernhardt C, Engel V, Givnish T, Hagerthey S, Harvey J, Leonard L, McCormick P, Mcvoy C, Noe G, Nungesser M, Rutchey K, Sklar F, Troxler T, Volin J, Willard D (2011) Recent and Historic Drivers of Landscape Change in the Everglades Ridge Slough, and Tree Island Mosaic. Critical Reviews in Environmental Scie nce and Technology 41: 344 381 Lloyd CD, Berberoglu S, Curran PJ, Atkinson PM (2004) A comparison of texture measures for the per field classification of Mediterranean land cover. In ternational Journal of Remote Sensing 25: 3943 3965 Lockwood JL, Ross MS, Sah JP (2003) Smoke on the water: the interplay of fire and water flow on Everglades restoration. Frontiers in Ecology and the Environment 1: 462 468 Loh WY, Shih YS (1997) Split s election methods for classification tree s. Statistica sinica 7: 815 840 Lu D, Weng Q (2007) A survey of image classification methods and techniques for improving classification performance. International Journal of Remote Sensing 28: 823 870 Madden M, Jo nes D, Vilchek L (1999) Photointerpretation Key for theEverglades Vegetation Classification System. Photogrammetric Engineeri ng & Remote Sensing 65: 171 177 Maheu Giroux M, Blois S (2005) Mapping the invasive species Phragmites australis in linear wetland corrid ors. Aquatic Botany 83: 310 320 Marella RL (2009) Water Withdrawals, Use, and Trends in Florida, 2005. Scientific Investigations Report, Florida Department of Environmental Protection, U.S. Department of the Interior, U.S. Geological Survey, Reston V A, USA Mas JF (2004) Mapping land use/cover in a tropical coastal area using satellite sensor data, GIS and artificial neural networks. Estuarine, Coasta l and Shelf Science 59: 219 230
108 Mas JF, Flores JJ (2008) The application of artificial neural netw orks to the analysis of remotely sensed data. International Journal of Remote Sensing 29: 617 663 Mather P (2004) Computer processing of remotely sensed images: an introduction. John Wiley & Sons, Chichester West Sussex Englan d Maxa M, Bolsta d P (2009) Mapping northern wetlands with high resolution satellite images and LiDAR. Wetlands 29: 248 260 Michishita R, Xu B, Gong p (2008) A decision tree classifier for the monitoring of wetland vegetation using Aster data in the Poyang lake region, C hina. In: The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Beijing, China, p. 8 Mitchell TM (1997) Machine Learning. McGraw Hill Inc. New York, NY, USA Mitsch W, Gosselink J ( 2007) Wetlands. Wiley, Hoboken N J USA Moody A, Gopal S, Strahler AH (1996) Artificial neural network response to mixed pixels in coarse resolution satellite data. Remote Sen sing of Environment 58: 329 343 Nicholls RJ, Hoozemans FMJ, Marchand M (1999) Increasing flood risk and wetland losses due to global se a level rise: regional and global analyses. Global Environmental Change 9: S69 S87 Nielsen EM, Prince SD, Koeln GT (2008) Wetland change mapping for the US mid Atlantic region using an outlier detection technique. Remote Sensing of Environment 112: 4061 4 074 Ogden JC (2005) Everglades ridge and slough conceptual ecologi cal model. Wetlands 25: 810 820 Ogden L (2004) Ecosystems or Landscapes, Public Participation in Environmental Decision Making: A Case Study in the Florida Everglades. In: Knowledge of La ndscapes to Landscaping Action. Cemegref, Bordeaux, France, p. 10 Olmsted IC, Armentano TV (1997) Vegetation of Shark Slough, Everglades National Park. Technical Report, South Florida Natural Resources Center, Everglades National Park Olmsted IC, Loope L L, Rinze RE (1980) A survey and baseline analysis of aspects of the vegetation of Taylor Slough, Everglades National Park. Everglades National Park South Florida Rese arch Center, Homestead, F L, USA Ouma Y, Tetuko J, Tateishi R (2008) Analysis of co occurr ence and discrete wavelet transform textures for differentiation of forest and non forest vegetation in very high resolution optical sensor imagery. International Journal of Remote Sensing 29: 3417 345 6
10 9 Ozesmi SL, Bauer ME (2002) Satellite remote sensing of wetlands. Wetlands Eco logy and Management 10: 381 402 Pal M, Mather PM (2003 ) An assessment of the effectiveness of decision tree methods for land cover classification. Remote sen sing of environment 86: 554 565 Paola JD, Schowengerdt RA (1995) A Detai led Comparison of Backpropagation Neural Network and Maximum Likelihood Classifiers for Urban Land Use Classification. IEEE Transactions on Geoscience and Remote Sensing 33: 981 997 Paul S, Ksel K, Alewell C (2006) Reduction processes in forest wetlands: tracking down heterogeneity of source/sink functions with a combination of methods. Soil Biology and Biochemistry 38: 1028 1039 Pearlstine L, Portier KM, Smith SE (2005) Textural discrimination of an invasive plant, Schinus terebinthifolius, from low alt itude aerial digital imagery. Photogrammetric Engineeri ng & Remote Sensing 71: 289 298 Pearlstine LG, Pearlstine EV, Aumen NG (2010) A review of the ecological consequences and management implications of climate change for the Everglades. Journal of the North American Benthological Society 29: 1510 1526 Puig J, Hyman G, Bolanos S (2002) Digital Classification vs. Visual Interpretation : a case study in humid tropical forests of the Peruvian Amazon. In: XXIX International Symposium on Remote Sensing of Envi ronment. Buenos Aires, Argentina, p. 5 Puissant A, Hirsch J, Weber C (2005) The utility of texture analysis to improve per pixel classification for high to very high spatial resolution imagery. International Journa l of Remote Sensing 26: 733 745 Quinlan J (1993) C4.5: programs for machine learning. Morgan Kaufmann Publishers, San Mateo C A, USA RECOVER (2004) CERP monitoring and assessment plan: Part1 monitoring and supporting research. United States Army Corps of Engineers, Jacksonville, FL, USA and South Florida Water Management Dis trict, West Palm Beach, FL, USA Reed P (1988) National list of plant species that occur in wetlands: national summary. Biological Report, National Ecology Research Center, St. Petersburg, FL USA Richardson CJ, Huvane JK (2008 ) Ecological status of the Everglades: environmental and human factors that control the peatland complex on the landscape Everglades Experiments: 13 58 Ecology and Manageme nt 18: 517 542
110 Richards J, Jia X (2006) Remote sensing digital image analysis an introduction. Springer, Berlin Germany Ross MS, Sah JP, Ruiz PL, Jones DT, Cooley H, Travieso R, Tobias F, Snyder JR, Hagyari D (2006) Effect of hydrologic restoration on the habitat of the Cape Sable seaside sparrow Ross MS, Reed DL, Sah JP, Ruiz PL, Lewin MT (2003) Vegetation: environment relationships and water management in Shark Slough, Everglades National Park. Wetlands Eco logy and Management 11: 291 303 Ross MS, Mitchell Bruk er S, Sah JP, Stothoff S, Ruiz PL, Reed DL, Jayachandran K, Coultas CL (2006) Interaction of hydrology and nutrient limitation in the Ridge and Slough landscape of the southern Evergl ades. Hydrobiologia 569: 37 59 Rutchey K, Schall T, Sklar F (2008) Deve lopment of vegetation maps for assessing Everglades restoration progress. Wetlands 28: 806 816 Rutchey K, Schall T, Doren R, Atkinson A, Ross M, Jones D, Madden M, Vilchek L, Bradley K, Snyder J, Burch J, Pernas T, Witcher B, Pyne M, White R, Smith T, Pat terson M, Gann G (2006) Vegetation Classification for South Florida Natural Areas. Open File Report, United States Geological Survey, Saint Petersburg, F L Rutchey K, Godin J (2009) Determining an appropriate minimum mapping unit in vegetation mapping for e cosystem restoration: a case study from the Everglades, USA. L andscape Ecology 24: 1351 1362 Sader SA, Ahl D, Liou WS (1995) Accuracy of landsat TM and GIS rule based methods for forest we tland classification in Maine Remote Sensing of Environment 53 : 133 144 Schmidt KS, Skidmore AK (2003) Spectral discrimination of vegetation types in a coastal wetland. Remote Se nsing o f Environment 85: 92 108 Sesnie SE, Gessler PE, Finegan B, Thessler S (2008) Integrating Landsat TM and SRTM DEM derived variables with decision trees for habitat classification and change detection in complex neotropical environments. Remote Sensin g of Environment 112: 2145 2159 Sklar FH, van der Valk A (2002) What we know and should know about tree islands. In: van der Valk A, Skla r FH (eds) Tree Islands of the Everglades. Kluwer Academic Publishers, Dordrecht, The Netherlands, pp. 499 522 Smith GM, Spencer T, Murray AL, French JR (1998) Assessing seasonal vegetation change in coastal wetlands with airborne remote sensing: an outli ne methodology. Mang roves and Salt Marshes 2: 15 28
111 St Louis V, Pidgeon AM, Radeloff VC, Hawbaker TJ, Clayton MK (2006) High resolution image texture as a predictor of bird species richness. Remote sens ing of environment 105: 299 312 Stober QJ, Scheidt D Jones R, Thornton K, Ambrose R, France D (1996) South Florida Ecosystem Assessment Interim Report. United States Environmental Protection Agency Stokstad E (2008) Florida: Big Land Purchase Triggers Review of Plans to Rest ore Everglades. Science 321: 22 Swain PH, Hauska H (1977) The decision tree classifier: design and potential. IEEE Transactions on Geoscience E ectronics GE 15: 142 147 Titus JG, Richman C (2001) Maps of lands vulnerable to sea level rise: modeled elevations along the US Atlantic and G ulf coast s. Climate Research 18: 205 228 Tso B, Mather PM (2009) Classification methods for remotely sensed data. CRC Press, Boca Raton FL, USA USFWS (1999) Freshwater marshes and wet prairies. South Florida multi species recovery plan Ecological communities. h ttp://www.fws.gov/verobeach/images/pdflibr ary/marshes%20wet%20prairies.pd f A ccessed 23 Apr 2011 Walker R, Solecki W (2004) Theorizing Land Cover and Land Use Change: The Case of the Florida Everglades and Its Degradation. Annals of the Association of Amer ican Geographers 94: 311 328 Wang L, Sousa WP, Gong P, Biging GS (2004) Comparison of IKONOS and QuickBird images for mapping mangrove species on the Caribbean coast of Panama. Remote Sen sing of Environment 91: 432 440 Waser LT, Baltsavias E, Ecker K, Eisenbeiss H, Feldmeyer Christe E, Ginzler C, Kchler M, Zhang L (2008) Assessing changes of forest area and shrub encroachment in a mire ecosystem using digital surface models and CIR aerial images. Remote Sensin g of Environment 112: 1956 1968 Welch R, R emillard M, Doren RF (1995) GIS Database Development for South & Remote Sensing 61: 1371 1381 Willard DA, Bernhardt CE, Holmes CW, Landacre B, Marot M (2006) Response of Everglades tree is lands to environmental change. Ec ological Monographs 76: 565 583 Winter TC (1999) Relation of streams, lakes, and wetlands to groundwater flow systems Hydrogeology Journal 7: 28 45
112 Wolf P, Dewitt B (2000) Elements of photogrammetry: with applications in GIS. McGraw Hill, Boston MA, USA Wright C, Gallant A (2007) Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental data. Remote Sens ing of Environment 107: 582 605 Wulder MA, LeDr ew EF, Franklin SE, Lavigne MB (1998) Aerial Image Texture Information in the Estimation of Northern Deciduous and Mixed Wood Forest Leaf Area Index (LAI). Remote S ensing of Environment 64: 64 76 Yang CC, Prasher SO, Enright P, Madramootoo C, Burgess M, G oel PK, Callum I (2003) Application of decision tree technology for image classification using remote sensing data. Agr icultural Systems 76: 1101 1117 Yang X (2007) Integrated use of remote sensing and geographic information systems in riparian vegetation delineation and mapping. International Journal of Remote Sensing 28: 353 370 Zedler JB (2000) Progress in wetland restorat ion ecology. Nature 15: 402 407 Zedler JB (2003) Wetlands at your service: reducing impacts of agriculture at the watershed scale. Frontiers in Ecolo gy and the Environment 1: 65 72 Zhang S, Liu HX, Gao DT, Wang W (2003) Surveying the methods of improving ANN generalization capability. In: Machine Learning and Cybernetics, 200 3 International Conference pp. 1259 1263 Zweig CL, K itchens WM (2008) Effects of landscape gradients on wetland vegetation communities: information for large scale restoration. Wetlands 28: 1086 1096
113 BIOGRAPHICAL SKETCH Zoltan Szantoi, son of Zoltan Szantoi and Erzsebet Pap, was born in 1977 in Szentes, Hungary. He and his broth er, Robert Szantoi were reared o n a small tomato and cauliflower farm in Csanytelek, Hungary. In June 1995, he graduated from Janos Bartha Horticultural High School. He attended Istvan Barsony Environmental Technician School in Csongrad Hungary and graduated in 1997. He then entered Samuel Tessedik College, Szarvas, Hungary in September, 1997 and graduated with a Bachelor of Science degree in e nvironmental a gricultural e ngineering in June 2001. During 2001 to 2002, he worked as an intern at the Heritage Seedling Inc. in Salem, Oregon. Working in the industry for one year after the internship he entered the School of Forestry and Wildlife, Au burn University, in January 2004 and earned a Master of Science degree in 2006. In 2007 he was accepted as a PhD. graduate student at the University of Florida and graduated in 2011