Jason W. Rueter. A MULTIVARIATE APPROACH TO ESTABLISH HABITAT CLASSIFICATIONS IN A CORAL REEF ECOSYSTEM USING LANDSAT 7 SATELLITE IMAGERY AT TURNEFFE ATOLL, BELIZE, CENTRAL AMERICA. (Under the direction of Dr. Joseph J. Luczkovich). Department of Biology, April 2004. The use of multivariate cluster and discriminant function analysis of ground truth data and LANDSAT 7 satellite imagery (23 May 2000) was examined to classify tropical marine habitats at Tumeffe Atoll, Belize. Ground truth data were obtained using SCUBA, digital video and the global positioning system in June 2000 and July 2001. Ground trothing was done by swimming along eleven 150- to 300-meter transect lines located on the southeastern side of the atoll and videotaping bottom habitat from 1.5 m above the bottom at 30-meter intervals (126 sites). The still images were displayed on a 21cm x 29 cm flat-screen monitor, overlaid with a 20-cell grid (5.25 cm x 5.80 cm cells), and% cover of seven bottom cover classes was computed. Using hierarchical cluster analysis of water depth and habitat percent cover data, the 126 sites were grouped into three distinct habitat clusters (sand, seagrass and coral reef habitats). A discriminant analysis of the sites was performed using LANDSAT digital values in enhanced thematic mapper (ETM+) Bands 1, 2, 3, and 4 (the only ones that penetrate water) as predictors of habitat cluster membership. When predicting the habitat class of the sites used for training, coral reef areas were correctly classified 100%, seagrass areas 80 %, and sand dominated areas 77 % of the time. Using the classification function from the discriminant analysis, the LANDSAT 7 image was recoded to create a habitat map of the region surrounding Tumeffe Atoll, showing coral reef, seagrass, and sand-dominated regions. Using the recoded map, new ground truth sites were visited in 2002 to record depth and benthic classification. The recoded map had an overall accuracy of 60% and a ,au coefficient of 52.47%. Such maps can be used to identify essential fish habitat areas and study changes in habitat area over time. A MULTIVARIATE APPROACH TO ESTABLISH HABITAT CLASSIFICATIONS IN A CORAL REEF ECOSYSTEM USING LANDSAT 7 SATELLITE IMAGERY AT TURNEFFE A TOLL, BELIZE, CENTRAL AMERICA. A Thesis Presented to The Faculty of the Department of Biology East Carolina University In Partial Fulfillment of the Requirements for the Degree Master of Science in Biology By Jason Walter Rueter March 2004 A MUL TIV ARIA TE APPROACH TO ESTABLISH HABIT AT CLASSIFICATIONS IN A CORAL REEF ECOSYSTEM USING LANDSAT 7 SATELLITE IMAGERY AT TURNEFFE ATOLL, BELIZE, CENTRAL AMERICA. by Jason Walter Rueter APPROVED BY: DIRECTOR OF THESIS COMMITTEE MEMBER Gerhard W. Kalmus, Ph.D. COMMITTEE MEMBER Roger A. Rulifson, Ph. 1;). COMMITTEE MEMBER Yong Wang, Ph.D. Chair of the Department of Biology: Interim Dean of the Graduate School: Paul D. Tschetter, Ph. D. Acknowledgements I would like to sincerely thank Dr. Joseph J. Luczkovich for all that he has done for me from getting me into the Graduate program through the completion of this thesis. I would also like to thank Dr. Gerhard Kalmus and the Biology Graduate Department for their generous help in making this thesis possible. Dr. Roger Rulifson, and Dr. Yong Wang have been very generous with their support and guidance in this thesis and throughout my years in school. I would also like to thank Eden Garcia and the Institute of Marine Science at Belize University for their assistance in this project. The facilities and the people at IMS went above and beyond time and again with their support. In addition, the staff at ECU's dive safety office headed by Steve Sellers was extremely helpful in providing needed equipment and training for this thesis. I would also like to thank the numerous undergraduate and graduate students as well as Dr. Peter Sale' s Mesoamerican Barrier Reef Initiative group who helped with data collection and gave their support to me throughout this thesis. Finally, I would like to thank my parents, Curtis and Carol Rueter, for their years of support, without which none of this would have been possible. I would also like to thank my grandparents, Barbara and Charles Tillman, and my aunt and uncle, Sandra and Richard Goodman, for their financial support and belief in me. Thank you to all of you who helped, I am forever indebted. Table of Contents List of Tables ....... ......... ....................... .. ... ................................. .... ..... ......... .. .... ... ........ .. .... iii List of Figures .... .......... ... ...... ...... ...... . , .... .......... .... ... ... ....... .. .. .................. .......... ................ iv Introduction .. ... ................... ...... ........ .... .......... .... ... ....... .... .... ............. ... .......... ........ ............. 1 Background ........ .. ................ ..... ....................................... .... ...... ... ...... ..... ... ........ .. .......... 1 LANDSAT and Enhanced Thematic Mapper (ETM+) ....................................... 5 Materials and Methods ......... ....... .. ........ .. ................................ ...... ... .... ................ .. ........ .... . 8 Site Description and Remote Sensing ............................................................................. 8 Ground-truthing ............................................................................................................. 12 Habitat Classification using Optimization Clustering ....................... ... ............ ... .......... 16 Discriminant Function Methodology ......... ...... ........................................... ...... .. .. ... ..... 16 Classification of the LANDSAT 7 .... .... .. ..... ....... .. .. ................................. .. ..... ... .. .. ....... 17 Ground Truthing the Accuracy of the Discriminant Function at New Sites on Tumeffe ... , .. .. .......... ........................... .. ......... ........... ......... .... .................... .. ...... ............. 18 Results ...... ..... ........ .. ... ..... .. .......... ..................... .... ....................... .... ....... .... ............. .. ....... 21 Classification of Ground Truth Data ..................................... .. ........ ................... ...... ..... 21 Discriminant Function Analysis .................................................................................... 22 LANDSAT 7 Image Recoding ................... .............................. ... .. .... .. ..... ... ...... ..... ....... 27 Assessment of Classification Accuracy of the Discriminant Function ...... ......... ....... ... 32 Concluding Remarks .. .. .... ....... ....... ... .. ....... ... .... ..... ...... .... ..... .................... ......... ..... ... .. . 37 References .................................................................: ........ .. ........ ... ...... .. .. .. ............ ... ....... 44 Appendix A. Ground-Truthing Data for the 126 Stations .... .. ..... .. . . . .. ... . .......... .. . .48 11 Appendix B. Canonical Scores of the 78 Stations Used in the Analysis ... ... ........... . ... 58 Appendix C. CD of Digital Images of the Ground-Truthing Stations ........ . . . ........ . .... 60 List of Tables 1. The spectral characteristics and resolution of the Enhanced Thematic Mapper (ETM+) bands ......................................................................................................... 7 2. The R2 values for groups of stations ranging from 2 to 10 user-defined groups ......... 24 3. Classification error matrix of the discriminant function based on training data with the number of stations in each classification ................................................. 26 4. Error Matrix for new sites visited in July 2002 with the number of stations in each class ................... ... ......................................................................................... 35 5. Error Matrix for sites visited in July 2002 with the land and ocean classifications removed from the analysis with the number of stations in each class ...... ............ 36 6. Error Matrix for new sites visited in July 2002 without site 2 ............ .... ... ................. 42 7. Error Matrix for new sites visited in July 2002 removing sites with a depth > 11.0 m ................................................................................................................ 43 List of Figures 1. Map of Central America and the Caribbean showing the location of Belize (www.geographynetwork.com) ...................... .. .... .............. ..... .. ...... .................. ........ .. 4 2. The coastal area of Belize, showing Belize City, the barrier reef system, and Tumeffe Atoll, LANDSAT 7 scene L71019048_04820000523, shown in the true color image ofETM+ bands 1 (red), 2 (green), and 3 (blue) with the windward Cay es and the study areas used for ground-truthing ... ............................ . 10 3. Area of interest close up view of the study area, showing the windward cay es and the study areas used for ground-truthing in 2000 and 2001 shown in false color bands 1, 2, and 4 .. .... .. ......................... .. ..... ........................ ........... ......... .. 11 4. Transect locations at the two Cayes as well as the sand flat between Calabash and Soldier Caye, and Cut-finger reef south of Calabash ........... .... ... ............ ........... 14 5. The 7 bottom cover types (a) sand, (b) seagrass, (c) foliose algae, (d) hard coral and soft coral, ( e) coraline algae on dead coral, and (f) rubble ..... .. ...... .................... 15 6. The Model Maker design implemented in the reclassification of the LANDSAT image for factor 1 (a) and factor 2 (b ) ................... .. .............................. 20 7. Canonical scores plot showing the grouping of scores for each of the three habitat types .................................................................... .. ......................................... 25 8. Discriminant Function Factor 1 used to recode Landsat image (Scene ID: L71019048_04820000523; WRS Path 019, Row 048) identifying 5 habitats .......... 29 9. Discriminant Function Factor 2 used to recode Landsat image (Scene ID: L71019048 04820000523; WRS Path 019, Row 048) .... .... ... .. .. ....... .... ........ ........... 30 10. Close-up of Calabash Caye Transects in the recoded LANOSAT image ................... 31 11. Factor 1 recoded map showing the location of the July 2002 ground- truthing sites ... .. ............... ............... ........ ... .... ... ..... ........ ........... ....... ..................... ..... 34 Introduction Background Fisheries biologists would like to measure the amount of available essential habitat that fish use as spawning, feeding and shelter sites (Baird 1999, Rubec et al. 1999). Maps of the benthic habitats in coastal areas are important so that coastal planners and managers can monitor changes in habitats over time, ecologists can understand the natural habitat associations of continuously varying communities of organisms, and develop marine protected areas and coastal zone management plans (Mcfield et al. 1996). In tropical coastal zones, the main aquatic habitat type is coral reefs and their associated seagrass and mangrove habitats. The habitat maps are widely regarded to be an essential data source for coastal management planning and establishing marine protected areas (Cendrero 1989, McNeill 1994, Kenchington and Claasen 1998). In addition, ecologists want to have such maps to understand habitat distribution on large areal extents so that they can understand habitat associations of organisms, examine the influence of habitat on animal abundance and distribution patterns, and plan biological sampling programs (Aronson and LeFloc 'h 1996). In tropical coastal ecosystems, habitat patterns and animal associations must be discerned over areas that are hundreds of square kilometers. With such areal extent, it is useful to establish habitat patterns and community associations in small regions, and then extrapolate these associations to a large area plotted on a map. One approach to delimit habitat is to use satellite remote sensing to study the benthic cover available as fish habitat. This approach is especially useful in the clear 2 waters of the tropics and on coral reefs, where visible wavelengths of light penetrate the waters and the reflectance from the bottom can be measured by a satellite sensor (Lyzenga 1981, Jupp et al. 1985, Luczkovich et al. 1993). Not only do these images cover large geographic areas, but also repetitive images can be captured consistently over time (Luczkovich et al. 1991 , Holden and LeDrew 1999). Mumby et al. (1999) reported that satellite remote sensing techniques are cost-effective and accurate in studying coral reefs as compared with aerial photography data sources. In their study, the authors concluded that LANDSAT thematic mapper (TM) satellite digital data were found to be more cost effective and less time consuming to acquire and process than using airborne sensors (Mumby et al. 1999). Landsat TM is also significantly more accurate than other satellite sensors ( overall accuracy 73 % ) for mapping at coarse descriptive resolution (i.e. four habitat classes; sand, coral, algae, seagrass) (Mumby et al. 1997). However it is not clear how many habitat classes are biologically relevant in such a study. Habitat classifications that are discemable from satellite imagery may not correspond to useful habitat classifications that are used differently by fishes or other mobile organisms. This study is an attempt to characterize and map available fish habitat from space along the coastal reefs of Belize for ultimate use in developing marine protected areas (MP A) and coastal zone management plans. I will attempt to derive biologically relevant habitat classes from the underlying bottom habitat (using digital video ground trothing) and relate them to Landsat ETM+ digital data to use in habitat classification over a larger area. Multivariate cluster and discriminant function analysis (DF A) of ground truth data and LANOSA T 7 satellite imagery were used to classify tropical 3 marine habitats at Tumeffe Atoll, Belize (Figure 1). The objectives of this study were (1) to obtain ground-truth data ofbenthic habitats and depth using underwater digital video and (2) to assess the similarity of the habitats statistically using optimization cluster analysis to define natural habitat classes; (3) to use Landsat 7 satellite data as a predictor of the benthic habitat classes as defined by the cluster analysis in objective (2) in a discriminant function analysis; and ( 4) to ground-truth predictions of the discriminant function analysis so that prediction accuracy of the Landsat data can be estimated for this habitat map in Belize. Belize was chosen as the study site of this project for several reasons. Belize is home to the second largest barrier reef system in the world and the only barrier reef system in the Western hemisphere. It is also a rapidly growing nation that relies heavily on the reefs for sustenance as well as tourism, but is increasingly under pressure from coastal development, coral reef bleaching events, pollution, and habitat destruction (Mcfield et al. 1996). The tourism industry accounted for 20.6 % of the Gross Domestic Product in 1998 and 19.3 % in 1999 (Belize Audubon Society). The Tumeffe Atoll island chain, where this study took place, is currently under consideration for marine protected area (MP A) status, has a history of research, and is home to several current projects including a Bonefish and Tarpon Unlimited (www.tarbone.org) project as well as the Mesoamerican Barrier Reeflnitiative (Sale et al. 2001). 4 Gulf Of Me ,ico ~ ,· ./ {l ___ _,.,,,- eche ICO f Calibbean Sea Belize + N Pacific Ocean Figure 1. Map of Central America and the Caribbean showing the location of Belize (www.geographynetwork.com). 5 LANDSAT and Enhanced Thematic Mapper (ETM+) LANDSAT 7 is a U.S. satellite used to acquire remotely sensed images of the Earth's surface. It was launched on April 15, 1999 from Vandenberg Air Force Base in California as part of the National Aeronautic and Space Administration' s (NASA) Earth Observing System (EOS)(Landsat.gsfc.nasa.gov). Landsat 7 collects 532 scenes/day from an orbit of 705 km above the Earth's surface at an inclination of 98.2° and has a revisit period (of the same location on the Earth) every 16 days. The satellite is equipped with an Enhanced Thematic Mapper+ (ETM+) . The ETM+ consists of advanced, multispectral scanners with eight bands designed to achieve higher image resolution, sharper spectral separation, improved geometric fidelity and greater radiometric accuracy and resolution than previous sensors. Landsat ETM+ data are sensed in eight spectral bands simultaneously. Bands 1-5 and 7 have a resolution of 30 x 30 m while band 6 has a resolution of 60 x 60 m. Band 8, a panchromatic band has a resolution of 15 m x 15m (Table 1) . The spectral characteristics of each band can also be seen in Table 1. Using Landsat satellite data for coral reef monitoring has many advantages. The Landsat 7 satellite imagery and images obtained from its predecessors (Landsat 4 & 5) have been investigated by a number of researchers since their launches to examine their usefulness and effectiveness in mapping and monitoring coastal and aquatic environments ( e.g., Luczkovich et al. 1993, Holden and Le Drew 1999, Mumby et al. , 1999, Andrefouet et al. 2001). This tool will become important to fisheries managers, environmentalists, and ecologists in monitoring degradation to essential fish habitats (EFH) and ensuring compliance with the Sustainable Fisheries Act (SF A). By mapping coastal habitats and 6 comparing them over time, degradation, preservation, and recovery can be noted and further analyzed. This study hopes to show that Landsat satellite data can be used to classify and monitor these large areal extents in Belize with a known level of accuracy. 7 Table 1. The spectral characteristics and resolution of the Enhanced Thematic Mapper (ETM+) bands. Thematic Mapper Band Spectral Range (µm) Resolution (m) Water penetration 1 0.45 to 0.515 30 X 30 Yes 2 0.525 to 0.605 30 X 30 Yes 3 0.63 to 0.690 30 X 30 Yes 4 0.75 to 0.90 30 X 30 Some 5 1.55 to 1.75 30 X 30 No 6 10.40 to 12.5 60x 60 No 7 2.09 to 2.35 30 X 30 No Pan (8) 0.52 to 0.90 15 X 15 Yes Materials and Methods Site Description and Remote Sensing Belize is located just south of Mexico on the Caribbean Sea. Tumeffe Atoll is the largest of three offshore atolls in Belize and consists of more than 200 cayes. Specific sites that were visited in June 2000 and July 2001 for ground-truthing were near Soldier (87°47'43.4" Wand 17°19' 16.1" N) and Calabash Cayes (17°16' 55.2" Wand 17°16'55.2" N), which are islands in the shallow lagoonal waters along the windward edge of Tumeffe Atoll (Figure 2). Soldier Caye is located to the north and is separated from Calabash Caye by a wide channel with a sandy bottom (the sand flat area, 17°17'47.6" Wand 87°48'32.9" N ), which was also used for ground-truthing. Cut Finger Reef (17°16'02.0" Wand 87°49.09.6" N), located at the southern end of Calabash Caye, was a fourth ground-truthing site (Figure 3). At each of these four study sites, transects ranging in size from 150-300 meters were surveyed using SCUBA and a digital video camera. Depth of transects ranged from 0.6 - 11_.0 m and was measured by a diver positioned on the bottom with a Suunto Vyper dive computer. These sites were chosen after a preliminary examination of the satellite imagery indicated that bottom cover changed dramatically over the length of these transects, and thus provided the maximum range of bottom reflectance as detected by the Landsat sensor. Specific transect lines depended upon accessibility to the sites as well as wave and weather conditions. In addition, travel to distant parts of the atoll in 2000 and 2001 was limited due to boat and fuel resources. 9 The image used in this study is a LANDSAT 7 image of the coastal region of Belize acquired on 23 May 2000 (Figure 2; Scene ID: L71019048_04820000523; WRS Path 019, Row 048). 10 Figure 2. The coastal area of Belize, showing Belize City, the barrier reef system, and Turneffe Atoll, LANDSAT 7 scene L7l019048_04820000523 , shown in the true color image of ETM+ bands 1 (red), 2 (green), and 3 (blue) with the windward Cayes and the study areas used for ground-truthing. 11 Figure 3. Area of interest close up view of the study area, showing the windward ca yes and the study areas used for ground-truthing in 2000 and 2001 shown in false color bands 1, 2, and 4. 12 Ground-truthing Ground-truthing data were gathered in June 2000, July 2001 and 2002. The June 2000 data were used as a preliminary assessment of the feasibility of the project. This data was used in conjunction with the July 2001 data to construct a habitat classification model using multivariate statistical analyses. Ground truth images of the sea bottom were obtained by swimming along 11 transect lines located at the four study sites and videotaping bottom habitat from 1.5 m above the bottom at 30 m intervals using a Sony PC 100 digital video camera in a Mako underwater housing (Figure 4, 126 locations). Positions along each transect were determined using a Garmin GPS 12 receiver held at the water surface above each location. Accuracy of the GPS receiver was ± 15m RMS at the time because the Selective Availability was off during the period of this study. The Selective Availability is a tool used by the department of Defense to alter the accuracy of GPS receivers by up to 100m. Geo-referencing of the image itself was done using ERDAS and known land references at the University of Belize campus and the dock at the research center on Calabash Caye and Carry Bow Caye. Still images were taken from the digital video for each 30-m location along a transect, which placed each image in a separate LANDSAT pixel. Images were acquired from the video using a Sony® Ilink ™ video system and DV gate capture software for still pictures®. The still images were displayed on a 21cm x 29 cm flat-screen monitor, overlaid with a 20-cell grid (5.25 cm x 5.80 cm cells), and each cell assigned to one of seven bottom cover types (Figure 5). The cover types were sandy bottoms, seagrass meadows, macro-algae, hard coral, soft coral, dead coral covered with coraline algae, or small rubble fragmented from the main rock 13 formations. Percentage cover of each bottom cover type in each still video image was computed by counting cells assigned to each bottom cover type using the formula: % cover=(# cells of bottom cover type / 20) x 100. Those cells that contained more than one bottom cover type were judged by eye to determine the dominant benthic cover type in that cell. The bottom cover types (Figure 5) were the only ones observed in the shallow water areas around Calabash Caye. The organisms composing the seven cover types were not resolved to lower taxonomic levels to allow for rapid video analysis. Lower taxonomic levels of the benthic organisms may have been possible from the video, but was unlikely to produce a different clustering of habitats at a level that would be discemable in Landsat imagery, so was beyond the scope of this analysis. Video images for 90 of the 126 ground truth stations can be viewed in the accompanying CD (Appendix 3). The remaining 36 stations were collected in the preliminary study in July 2000, and the images have been lost, however the ground-truth data is still accounted for. 14 ~igure 4. Transect locations at the two Cayes as well as the sand flat between Calabash and Soldier Caye, and Cut-finger reef south of Calabash. 15 Figure 5. The 7 bottom cover types (a) sand, (b) seagrass, (c) foliose algae, (d) hard coral and soft coral, (e ) coraline algae on dead coral, and (f) rubble. 16 Habitat Classification using Optimization Clustering Cluster analysis is a multivariate statistical procedure that can find natural groups in a similarity matrix. Using an optimization cluster analysis (UCINET verion 5.78, Borgatti et al., 1999), the 126 locations were grouped based on similarity (Pearson correlation) of water depth and percentage bottom cover data obtained from video of the seven bottom types. The Pearson correlation was computed as a similarity measure among stations; the water depth and percentage bottom cover were used as variables to compute the correlations. A 126 x 126 matrix with a Pearson correlation coefficient in each cell of the matrix was created. The optimization clustering strategy is divisive, and divides the station matrix into two or more user-defined groups, in which similarity is greatest within a group, as judged by the average similarity within groups. This cluster analysis was repeated for group sizes 2 up to 10. Group sizes > 10 were not considered, because there were only 7 bottom cover types visible from the video. An R2 measure (a measure of within to between group variation) was examined after each division. Discriminant Function Methodology Discriminant Function Analysis (DF A) is a multivariate approach that allows prediction of group membership based on a training set of independent variables. In this study, group membership in the habitat classes determined from the groud-truthing video surveys was predicted from Landsat digital pixel values derived from each site visited (the training set). A DF A derives linear combinations of the independent variables (Landsat bands) that will discriminate between the a priori defined groups (habitat types) 17 in such a way that the misclassification error rates are minimized by maximizing the between group variance relative to the within-group variance. A subset of the ground truth sites were used as a training set for a discriminant function; only 78 sites were strongly associated with definable bottom classes. These aggregated bottom classes included sites with varying degrees of percent cover of the bottom cover classes. The coral reef habitat class for example, included the soft coral, hard coral, live rock, and rubble bottom cover types identified previously. The next step was to predict the habitat classes (cluster groups) from the Landsat data (training set) using a discriminant function. A discriminant function analysis was derived using data from Landsat bands 1, 2, 3, and 4 at the 78 sites identified as representative of habitat type classes. Classification oft he LANDSAT 7 Using a conditional statement with parameter values acquired from the canonical scores (Appendix 2) given by the discriminant function analysis, the ETM+ images were classified into one of five image classes (land, sand, seagrass, coral reef, and deep ocean). Figure 6 shows the steps involved with this classification. The land and deep ocean bottom types were determined by analyzing areas of the map whose scores fell outside of those determined to be sand, seagrass, or reef and identifying known geographical features . To determine boundaries for those habitats whose scores overlap, the averaging method of the canonical scores was used as described by Dillon and Goldstein (1984). 18 Ground Truthing the Accuracy oft he Discriminant Function at New Sites on Turnejfe In July 2002, additional ground-truth data were gathered to assess the accuracy of the habitat classification model at sites away from the ones used in the training set. Sites were chosen at random within new quadrants that contained a diversity of bottom reflectance patterns. Five areas were selected (1590 m x 1200m) along the coastal areas of Tumeffe Atoll. Sites for the new quadrants were areas of interest to the Mesoamerican Barrier Reef Systems project currently under way in the western Caribbean and were located on both the windward and leeward sides of the Atoll (Sale et al.) Individual sites within a quadrant were determined by Microsoft Excel random number generator and were located with the Garmin GPS 12 receiver. Each site was classified by determining bottom cover and depth measurements using SCUBA and snorkeling and was videotaped for further analysis. The percentage accuracy of the predicted habitat types versus the observed habitat type was calculated using user accuracy, classification accuracy, and the Tau coefficient (Mumby et al. , 1997). Tau ('r) is calculated from: T P0 -P, 1 Lm =-'--- , where P, =- . n; *X; 1-P N 2 l= I r Po = overall accuracy; m = number of habitats, i = ith habitat, N = total number of sites, ni = row total for habitat i, and xi = diagonal value for habitat (i.e., number of correct assignments for habitat i). User accuracy is the ability of the discriminant function to accurately predict habitat classification membership and is defined as (number of sites in each habitat class observed/ number of sites in each habitat class expected) x 100. Classification accuracy 19 is a measure of how well sites are classified for each habitat type and is defined as (number of sites classified correctly expected/ total number of sites of all classes observed) x 100. In other words user accuracy examines predicted versus observed where classification accuracy examines observed versus predicted. The Tau coefficient is a statistic that is readily interpretable, permits hypothesis testing, and accounts for chance agreement within the matrix (Ma and Redmond, 1995). A -r of 0.80 indicates that 80% more pixels were classified correctly than would be expected by chance alone (Mumby et al. 1997). 20 n1_composite_b1_b4_20000523 DISCRMINANT a) n5_ discriminant_recode_map_b1 _ 4_fctr1 CONDITIONAL n1 _ composite_b 1_ b4_20000523 DISCRIMINANT b) n5_ discriminant_recode_map_b 1_ 4_fctr2 CONDITIONAL Figure 6. The Model Maker design implemented in the reclassification of the LANOSAT image for factor 1 (a) and factor 2 (b). Results Classification ofG round Truth Data The clustering algorithm, when applied to the ground-truth data, indicated that the main biological classes of habitat were (1) coral reef, (2) seagrass, and (3) sandy bottom. The ground-truth data on percent bottom cover and depth provided a basis for arriving at these habitat classes, after analyses with the optimization clustering strategy, while varying the user-defined group sizes. When changing group size from 2 to 8, it is apparent that as the number of groups was increased, the R2 reached a peak then declined (Table 2). The R2 values, which are a measure of the degree to which within-group to between-group variance was maximized for each group size, peaked at R2 = 0.758 for 4 and 5 groups, but was slightly lower at other group sizes, with 6 groups having R2 = 0.757, and 3 and 7 groups having R2 = 0.751 (Table 2). Thus, optimal number of groups appeared to be between 3 and 7. The four-group solution had one group with just three stations, and the 5 and 6 group solutions had groups with one station (singletons). In cluster analysis, such small groups are undesirable. The"three-group solution was chosen because it had the greatest increase in R 2 and had the greatest number of stations per group with no singletons as group size was increased from 2 to 8. The list of stations in the 3-group solution is shown in Appendix 1. The stations within each group in the three-group solution were characterized as to the average [± one standard error of the mean (s.e.m.)] percent cover of each bottom type and depth. These three groups are hereafter called sand, seagrass and coral reef biological habitats. The 64 stations in group 1 (sand) had an average of 66.0 (± 2.49) % 22 cover of sand bottom, with an average depth of2.60 (± 0.22) m. The 11 stations in group 2 (seagrass) had mostly seagrass (mean= 82 ± 8.76 % cover) with an average depth of 1.8 (± 0.52) m. The 51 stations of group 3 (coral-reef) had mostly coraline-algae-covered dead reef (mean = 51.3 ± 3.17 % cover), but also significant amounts of hard coral (21.8 ± 2.29 % cover) and soft coral (21.7 ± 2.25 % cover) covering the bottom. This group also occurred in the deepest water (mean= 4.99 ± 0.26 m). Thus, after grouping the 126 locations along the 11 transects at the four ground-truth sites in the cluster analysis, three main habitat classes were identified as sand, seagrass, or coral reef. Discriminant Function Analysis For the discriminant function analysis, a subset of the ground-truth stations was selected as being good representatives of these three habitat types. These 78 stations were used to define the training set, because of their high percent cover for each dominant bottom cover type. The remaining locations were omitted from later analysis, as they represent transaction habitats between the three primary habitat type clusters. These 78 stations with clearly identifiable habitat classifications (sand, seagrass, and reef) were then used as a training set in a discriminant function analysis to predict image classes from LANDSAT 7 digital values in ETM+ bands 1, 2, 3, and 4. The results produced a canonical discriminant function (CFFl and CFF2) with two factors: CFFl = 0.014(TM 1) + 0.039(TM 2) + 0.023(TM 3) + 0.487(TM 4) - 26.187. CFF2 = 0.084(TM 1) - 0.027(TM 2)- 0.026(TM 3) - 0.836(TM 4) + 21.488 where TM x is ETM+ band x. 23 The large coefficients in these functions for ETM+ band 4 are due to the fact that this band (near infrared) has low digital values due to the lack of water penetration by infrared wavelengths. The canonical scores of each site in the 78-member training set were computed using these equations (Appendix 2), and plotted using the habitat classes derived from cluster analysis as plotting symbols for each station (Figure 7). The canonical scores of stations mostly fall into three distinct clouds of points showing good discrimination among habitats. However, there are some stations that plot outside the 95% confidence ellipse in the other classes. The canonical scores can then be turned into an a posteriori probability that gives the likelihood of the object belonging to each of the habitats. In terms of user accuracy, which is the ability of the DF A to accurately predict habitat classification membership, Landsat digital values from coral reefs were correctly classified 100%, from seagrass pixels 80 %, and from sand pixels 77 % of the time (Table 3). In terms of classification accuracy, the observed sand areas were placed into the - correct classification 96% of the time, the seagrass locations were correctly classified 73% of the time, and coral reef locations were correctly classified 88% of the time. Overall accuracy of the discriminant function was 88.5% and the Tau ('t) coefficient, which corrects the overall accuracy because some classifications are likely to be correct by chance alone, was slightly lower at 81 .25%. Thus, the discriminant function gives an overall highly accurate classification of habitats for the training set. 24 Table 2. The R2 values for groups of stations ranging from 2 to 10 user-defined groups. User-defined group number R2 Comments 2 0.698 2 large groups 3 0.751 Smallest has 11 stations 4 0.758 Group 4 has only 3 stations 5 0.758 Group 5 is a singleton 6 0.757 Group 6 is a singleton 7 0.751 3 singleton groups; 1 two- station group 8 0.747 5 singleton groups 9 0.732 5 singleton groups; 1 two- station group 10 0.744 7 singleton groups 25 4.----------.-----~------, r-, 1 "·~ '-N- -- :}_ ' ~ ~ 0 0 fu.... -2 <µ.. CLASS ·"__ _, SAND X CORAL -5 SEAGRASS -5 -2 1 4 FACTOR(l) Figure 7. Canonical scores plot showing the grouping of scores for each of the three habitat types. Each point represents one station included in the training set. The ellipse is a 95% confidence ellipse. 26 Table 3. Classification error matrix of the discriminant function based on training data with the number of stations in each classification. Expected habitat based on DF A factor 1. Overall accuracy= 88.5%; Tau (t) coefficient= 81.25%. Observed habitat in underwater video Habitat type User Sand Seagrass Coral reef Row accuracy total (%) Sand 23 3 4 30 77 Seagrass 1 8 1 10 80 Coral Reef 0 0 38 38 100 Column total 24 11 43 Classification Accuracy 96 73 88 (%) 27 LANDSAT 7 Jmage Recoding The LANOSAT 7 image was recoded to create a habitat map of the region surrounding Tumeffe atoll, showing coral reef, seagrass, and sand habitats as well as the terrestrial and open ocean areas of the atoll. The Model Maker function of ERDAS™ was used along with the discriminant function canonical scores computed for every pixel. Conditional values for the Model Maker Function were derived from the canonical scores of the training set (Dillon and Goldstein 1984) that placed each Landsat 7 pixel value into one of five classes ( open ocean, coral, seagrass, sand, and terrestrial). There are two additional habitat classes in this model ( open ocean and terrestrial classes) because these represented areas that were beyond my ground-truthing sample areas (deeper than 11 m or on land). Separate Model Maker processes were carried out for each factor of the discriminant function, which produced two recoded maps representing factor 1 (Figure 8) and factor 2 (Figure 9). The factor 1 map is most useful upon examination of the features within this map. One can see that blue areas are deep ocean, cyan-coded areas are coral reef, green areas are seagrass beds, yellow areas are sand flats , and red is land. Factor 1 appears to be greatly influenced by ETM+ band 4, which is important in identifying the island, and by ETM+ band 2, which would have the best ability to detect shallow water. Factor 2, however, does very little in delineating and identifying habitats. The apparent disturbance from wave and wind action in the aquatic areas can be readily discerned in this image. Figure 10 is a close-up of the recoded map shown in Figure 8 (from the DFA factor 1) , with transects that were ground-truthed at Calabash Caye overplotted with the 28 individual GPS waypoints shown. In this figure it appears that there is a small sand flat area adjacent to the land that leads into a seagrass bed, then into the coral reef, and finally out to open ocean. It appears to be most useful to examine individual waypoints and their habitat classifications with the recoded maps based on factor (Figure 8 and 10) compared to ground truth data. For example, compare Figure 10 to Figure 4 (lower left image of Calabash Caye) and one can see the transition described above in that false color image. Thus, the habitat recoding based on DF A factor 1 gives an intuitively satisfying map of a well-known section of Belize's coastal zone. However, this map needs verification in areas outside of the training set sites. 29 Figure 8. Discriminant Function Factor 1 used to recode Landsat image (Scene ID: L71019048_04820000523; WRS Path 019, Row 048) identifying 5 habitats. Color code: blue=deep water; cyan=coral reef; green=seagrass; yellow=sand bottom; red=land (o r clouds). 30 Figure 9. Discriminant Function Factor 2 used to recode Landsat image (Scene ID: L71019048_04820000523 ; WRS Path 019, Row 048). Color code: blue=deep water; cyan=coral reef; green=seagrass; yellow=sand bottom; red=land (or clouds). 31 Figure 10. Close-up of Calabash Caye Transects in the recoded LANDSAT image. Color code: blue==deep water; cyan==coral reef; green==seagrass; yellow==sand bottom; red==land (o r clouds). 32 Assessment of Classification Accuracy oft he Discriminant Function at New Sites The true accuracy of the discriminant function analysis (DF A) factor 1 can only be assessed using new data not in the training set data. New sites were visited in July 2002 to gather bottom habitat data at five different areas, each containing 10 points (Fig. 11, 70 sites). Accuracy of the DF A factor 1 predictions were found to be lower for new sites (overall accuracy 60 % and Tau (-r) coefficient 52.5 %; Table 4), because of an apparent misclassification of predicted coral reef areas that turned out to be seagrass. In addition, a more severe misclassification of predicted ocean areas that turned out to be seagrass areas occurred, and these misclassifications caused accuracies to decline overall. Indeed, all predicted habitat classes were observed to have seagrass habitats at least one site, resulting in a very low classification accuracy of 25 % for seagrass. It must be pointed out that two new habitat categories exist (land and deep ocean) which were not present in the initial analysis of the training set (Table 3). Land and ocean image classifications were created out of a necessity to classify those pixels, which had values either less than or greater than the canonical scores for the three identified habitat classes. When these two classes are removed from the analysis, the overall accuracy improves to 67% and the Tau (-r) coefficient becomes 55.4% (Table 5). However, there is still a problem with misclassification of predicted seagrass areas into coral reef areas (71 % user accuracy for seagrass), and observed seagrass areas being classified as coral reef ( 42% classification accuracy for seagrass ). Also, coral reef classified pixels turned out to be seagrass 40% of the time; thus, user accuracy was low for coral reef (60 %). It is apparent that seagrass and coral reef habitats were difficult to 33 discriminate using Landsat 7 ETM+ data alone and this discriminant function analysis at these new sites. 34 Figure 11. Factor 1 recoded map showing the location of the July 2002 ground-tn,thing sites. Color code: blue=deep water; cyan=coral reef; green=seagrass; yellow=sand bottom; red=land (or clouds). 35 Table 4. Error Matrix for new sites visited in July 2002 with the number of stations in each class. Expected habitat image class from DF A factor 1 versus observed habitat. Overall Accuracy= 60%; Tau Coefficient= 52.4 7%. Observed habitat in underwater video User Sand Seagrass Coral Row Land Ocean accuracy reef total (%) Sand 4 1 0 0 0 5 80 Seagrass 0 5 2 0 0 7 71 Coral Reef 0 6 9 0 0 15 60 Land 0 1 0 2 0 3 67 Ocean 0 7 3 0 10 20 50 Column total 4 20 14 2 10 Classification 100 25 64 100 100 accuracy (%) 36 Table 5. Error Matrix for sites visited in July 2002 with the land and ocean classifications removed from the analysis with the number of stations in each class. Expected habitat image class from DFA factor 1 versus observed habitat class. Overall Accuracy= 67%; Tau Coefficient= 55 .35%. Observed habitat while snorkeling or using SCUBA and in underwater video Habitat type User Sand Seagrass Coral Row accuracy reef total (%) Sand 4 1 0 4 80 Seagrass 0 5 2 7 71 Coral reef 0 6 9 15 60 Column total 4 12 11 Classification accuracy 100 42 82 (%) 37 Concluding Remarks I classified the coastal habitats of a small portion of the coastline of Belize based on Landsat 7 satellite imagery, videotaped ground-truth data, cluster analysis, and a discriminant function analysis (DF A). I used a discriminant function analysis with two factors to classify a Landsat image using pixel values alone and a training set. When the image was recoded using DFA factor 1, a map that identified 5 major habitat types was produced with an overall accuracy of 81 % for the training set, but fell to 67 % for new sites away from the training set sites. This technique appears to be a useful and inexpensive means to detect and monitor tropical coastal ecosystem habitats. In contrast, the second D FA factor appears to be affected by the characteristics of ETM+ band 4 and the near infrared wavelengths in water, and not habitat. These wavelengths only penetrate water to a depth of a few centimeters and tend to detect surface disturbances, such as wave action, causing the fanning pattern seen ih the factor 2 map (Figure l 0). When the factor 2 map is examined closely with respect to habitat distribution in known areas, there is no intuitive relationship observed, in spite of the apparent discrimination observed for training set seagrass sites with this factor in the canonical scores plot (Figure 7). Therefore, for the purposes of this study, the map produced by factor 2 is of no use for habitat classification. My method, using a discriminant function multivariate statistical technique, achieved relatively accurate habitat classifications: the lowest user accuracy is 77 % correct classification (sand) in the training set and 60 % (coral) in the new area data set. However, the method suffers from low classification accuracy for seagrass (42 %). 38 Practically, this method will under-estimate the distribution of seagrass habitats, which is an essential fish habitat that must be studied in other ways. However, coral reef areas will be mapped with relatively high accuracy (100 % classification accuracy at the training set sites, and 82 % at new sites), making it an excellent method for delimiting coral-reef fish habitat. The verification process carried out in July 2002 in areas outside of the training set was essential in rigorously testing the accuracy of the discriminant function factor 1 to predict habitat type. The accuracy at the new sites visited was considerably lower than when using the training set data. Although other researchers have published map accuracy estimates of 77% for Landsat data (Mumby et al. 1997), this is the first time that a model created on a training set was tested against sites outside of the training set. My map accuracy results exceed that of Mumby et al. (1997) for predicting correctly the training set, but are slightly lower when used on sites outside the training set. As more researchers attempt to use their models on areas outside of their training sets, a better understanding of the ability of this method or other methods to correctly map habitats will be gained. A number of explanations can be suggested to explain this difference in accuracy rates between training set data and the results of the July 2002 ground-truth data collected at sites outside the training set. The first of these concerns the inclusion of sites on the leeward side of the Atoll. It has been shown that the leeward side of the Belizean Atolls tends to have higher turbidity than the windward side due to the surface circulation in the western Caribbean that is dominated by the Caribbean current, which approaches the 39 Belize continental margin from the east (James and Ginsburg 1979). This water movement washes through the islands while moving east picking up loads of sediment, which are deposited on the leeward side. If site 2, which was a heavily sedimented sea grass bed in water 3 .4 -10 m deep on the leeward side, is removed from the assessment, the accuracy of the discriminant functi?n improves greatly (overall accuracy = 75 % and Tau ('t) = 69.25 %; Table 6). Improvements can be noted in the classification accuracy of seagrass, which improved to 50%. However, this is not satisfactory, as all coastal habitats need to be classified and we must accept that sedimentation and turbidity will occur in some areas and will remain a problem. But, by dropping this site, the model still had difficulty in identifying deep water coral reef at other sites and classified them as open ocean or seagrass (64 % classification accuracy). Another explanation for the reduction in accuracy deals with the depths of the sites visited in July 2002 compared to those used in the training set. Sites in 2002 ranged from 0.30 m to extremely deep water(> 1000 m), whereas sites used for the initial training set ranged in depth from 0.6 m to 11.0 m. The sites selected in the July 2002 work that were deeper than 11 .0 m would cause sunlight to attenuate during penetration and reflection, lowering the albedo in the visible wavelengths and all of the associated Landsat ETM+ band ' s digital values. This would have the effect of causing misclassification of observed coral reef and seagrass habitat types as predicted deep- water habitats. When deep-water sites(> 11.0 m) are removed from the analysis, the accuracy of the discriminant function again increases (overall accuracy = 77 %, Tau ('r)= 73.4 %; Table 7). Coral reef and sand habitats are correctly classified > 82 % of the time, 40 but there is still a low classification accuracy for seagrass habitats (42 %), which are misclassified as coral reef habitats. Therefore, given the limitations of the Landsat satellite's ability to sense visible light wavelengths in deep water, my recommendation for the application of this method is to restrict its use to windward reefs that are in depths less than 11.0 m. A third explanation for low accuracy is the elimination of problem (mixed) pixels in the training set data. Stations that were deemed to be a mixed pixel were eliminated in establishing the training data. In the July 2002 data these pixels are present, but are not eliminated from the analysis. This would cause lower percent accuracy for the overall analysis due to the percent cover data of these stations. A fourth possible explanation for the low accuracy of the predictions is that the habitats may have changed over the time between when the Landsat image was acquired (23 May 2000) and when the ground truth surveys were done. The surveys at the new sites away from the training set areas were done more than two years later in July 2002. If the bottom habitats changed over time, for example, if seagrass grew over the bottom in a deep water area between May 2000 and July 2002, the predictions of the DFA factor 1 model would have been in error. Although I cannot assess this source of error at the sites visited in the locations away from the Calabash Caye, I can say that the habitats near Calabash Caye did not change noticeably between June 2000 and July 2002, so I suspect that this is not a major source of classification error. Other concerns in the general use of this approach are that it is limited to scenes with low cloud cover and with good water visibility. An additional concern for the 41 general use of this approach is that there is no guarantee that the canonical discriminant functions derived can be applied without correction to new images with different radiometric gain values (which are listed in LANDSAT ca!ibration parameter files of every image). Entering a correction factor for the image gain factor into the Imagine Model Maker function when recoding the images using the discriminant function can standardize differences among images. If this image gain correction were done so that two images taken at different places or times were radiometrically balanced, then LANOSAT digital values could be converted into habitat classes by my method. With these corrections, coastal managers and ecologists could reclassify LANOSA T scenes in tropical coastal areas for use as rough biological habitat maps. Such maps would be inexpensive to prepare, and would save a tremendous amount of labor and money in performing habitat surveys for fish population estimates, habitat area measurements, and establishment of coastal marine protected areas. Such habitat maps would be based on a classification of biological (percent bottom cover and depth) data not radiometric data from Landsat imagery alone, making my approach biologically meaningful. In addition, if combined with change detection methods (Michalek et al. 1991 ), this approach can be used to monitor habitat boundary shifts over time. This is potentially an important tool for ecologists and coastal managers to monitor change in habitat areal coverage over time. 42 Table 6. Error Matrix for new sites visited in July 2002 without site 2. Expected Habitat Based on Discriminant Function versus observed habitat. Overall Accuracy = 75%; Tau Coefficient= 69.25%. Observed habitat in underwater video User Coral Row Sand Seagrass Land Ocean accuracy reef total (%) Sand 4 1 0 0 0 5 80 Seagrass 0 5 2 0 0 7 71 Coral reef 0 3 9 0 0 12 75 Land 0 1 0 2 0 3 67 Ocean 0 0 3 0 10 13 77 Column total 4 10 14 2 10 Classification 100 50 64 100 100 accuracy (%) 43 Table 7. Error Matrix for new sites visited in July 2002 removing sites with a depth > 11 .0 m. Expected habitat based on discriminant function versus observed habitat. Overall Accuracy= 77%; Tau Coefficient= 73.40%. Observed habitat in underwater video User Coral Row Sand Seagrass Land Ocean accuracy reef total (%) Sand 4 1 0 0 0 5 80 Seagrass 0 5 2 0 0 7 71 Coral reef 0 5 9 0 0 14 64 Land 0 1 0 2 0 3 67 Ocean 0 0 0 0 10 10 100 Column total 4 12 11 2 10 Classification 100 42 82 100 100 accuracy (%) REFERENCES Andrefouet, S., and Payri, C. 2001. Scaling-up carbon and carbonate metabolism of coral reefs using in-situ data and remote sensing. Coral Reefs 19 (3): 259-269. Andrefouet, S., Muller-Karger, F. E., Hochberg, E. J., Chuanmin, R , and Carder, K. L. 2001. Change detection in shallow coral reef environments using Landsat 7 ETM+ data. Remote Sensing of Environment 78: 150-162. Aronson, J., and Le Floc'h, E. L. 1996. Vital landscape attributes: missing tools for restoration ecology. Restoration Ecology 4 (4): 377-387. Baird, R. C. 1999. Introduction. Fish Habitat: Essential fish habitat and rehabilitation. Proceedings of the Sea Grant symposium of fish habitat. AFS Bethesda, MD Bonefish and Tarpon Unlimited. www.tarbone.org. March 2004. Carriquiry, J. D., Cupul-Magana, A. L., Rodriguez-Zaragoza, F. , and Medina-Rosas, P. 2001. Coral bleaching and mortality in the Mexican Pacific during the 1997-98 El Nino and prediction from a remote sensing approach. Bulletin of Marine Science 69 (1): 237-249. Cendrero, A. 1989. Mapping and evaluation of coastal areas for planning. Ocean and Shoreline Management 12: 427-462. Chauvaud, S., Bouchon, C., and Maniere, R. 1998. Remote sensing techniques adapted to high resolution mapping of tropical coastal marine ecosystems ( coral reefs, seagrass beds and mangrove). International Journal of Remote Sensing 19 (18): 3625-3639. Chuanmin, H., Muller-Karger, F.E., Andrefouet, S., and Carder, K.L. 2001. Atmospheric correction and cross-calibration of LANDSAT 7 ETM+ imagery over aquatic environments: a multiplatform approach using SeaW iFS/MODIS. Remote Sensing of Environment 78: 99-107. Dillon, W. R., and Goldstein, M. 1984. Multivariate Analysis, Methods and Applications. Wiley, New York. Foody, G. M., and Boyd, D. S. 1999. Fuzzy mapping of tropical land cover along an environmental gradient from remotely sensed data with an artificial neural network. Journal of Geographical Systems 1: 23-35. Geography network. March 8, 2004. www.geographynetwork.com. Environmental System Research Institute (ESRI). 45 Hardisky, M. A., Gross, M. F., and Klemas, V. 1986. Remote sensing of coastal wetlands. Bioscience 36 (7) : 453-459. Hochberg, E. J., and Atkinson, M. J. 2000. Spectral discrimination of coral reef benthic communities. Coral Reefs 19 (2) : 164-171. Holden, H., and LeDrew E. 1998. Spectral discrimination of healthy and non-healthy corals based on cluster analysis, principal components analysis, and derivative spectroscopy. Remote Sensing of Environment 65 (2): 217-224. Holden, H., and LeDrew, E. 1998. The scientific issues surrounding remote detection of submerged coral ecosystems. Progress in Physical Geography 22 (2): 190-221. Holden, H. and LeDrew, E. 1999. Hyperspectral identification of coral reef features. International Journal of Remote Sensing 20 (13): 2545-2563. James, N. P., and Ginsburg, R. K. The geological setting of Belize reefs. The seaward margin of Belize barrier and atoll reefs. Blackwell Scientific Publications. Oxford. Kenchington, R. A. , and Claasen, D. R. 1998. Australia' s Great Barrier Reef- management technology. Proceedings of Symposium on remote Sensing of the Coastal Zone, Gold Coast, Queensland. Department of Geographic Information, Brisbane, VA2.l-VA2.9. Lubin, D., Li, W., Dustan, P., Mazel, C. H., and Stamnes, K. 2001. Spectral signatures of coral reefs: Features from space. Remote Sensing of Environment 75 (1) : 127- 137. Luczkovich, J. J. , Wagner, T. W., and Stoffle, R. W. 1993. Discrimination of coral reefs, seagrass meadows, and sand bottom habitat types from space: A Dominican-Republic case study. Photogrammetric Engineering and Remote Sensing 59 (3) : 385-389. Lyzenga, D. R. 1981. Remote sensing of bottom reflectance and water attenuation parameters in shallow water using aircraft and Landsat data. International Journal of Remote Sensing 2 (1): 71-82. Ma, Z. and Redmond, R. L. 1995. Tau coefficients for accuracy assessment of classification of remote sensing data. Photogrammetric Engineering and Remote Sensing 61: 435-439. 46 Mcfield, M., Wells, S., and Gibson, J. (eds.) 1996. State of the Coastal Zone Report, Belize, 1995, Coastal Zone Management Programme, Government of Belize, Belize City, Belize, pp. 255, June, 1996. McNeill, S. E. 1994. The selection and design of marine protected areas: Australia as a case study. Biodiversity Conservation 3: 586-605. Menges, C. H., Hill, G. J. E., and Ahmad, W. 1998. Landsat TM data and potential feeding grounds for threatened marine turtle species in northern Australia. International Journal of Remote Sensing 19 (6): 1207-1221. Michalek, J. L., Wagner, T.W., Luczkovich, J. J., and Stoffle, R. W. "Multispectral change vector analysis for monitoring coastal marine environments", Photographic Engineering & Remote Sensing, Vol. 59, No. 3, pp. 381-384, March 1993. Mumby, P. J., Clark, C. D., Green, E. P., and Edwards, A.J. 1998. Benefits of water column correction and contextual editing for mapping coral reefs. International Journal of Remote Sensing 19 (1): 203-210. Mumby, P. J., Green, E. P., Edwards, A. J., and Clark, C. D. 1999. The cost- effectiveness of remote sensing for tropical coastal resources assessment and management. Journal of Environmental Management 55 (3): 157-166. Mumby, P. J., and Harborne, A. R. 1999. Development of a systematic classification scheme of marine habitats to facilitate regional management and mapping of Caribbean coral reefs. Biological Conservation 88 (2): 155-163. Mumby, P. J., Green, E. P., Edwards, A. J., and Clark, C. D. 1997. Coral reefhabitat- mapping: how much detail can remote sensing provide? Marine Biology 130 (2): 193-202. Mumby, P. J., Raines, P. S., Gray, D. A., and Gibson, J. P. 1995. Geographic information-systems - A tool for integrated coastal zone management in Belize. Coastal Management 23 (2): 111-121. Mumby, P. J. 2001. Beta and habitat diversity in marine systems: a new approach to measurement, scaling and interpretation. Oecologia 128: 274-280. NASA. March 8, 2004. http://landsat.gsfc.nasa.gov/ 0 'Regan, P. R. 1995. The use of contemporary information technologies for coastal research and management-a review. Journal of Coastal Research 12 (1): 193-204. 47 Phinney, J. T., Muller-Karger, F. , Dustan, P. , and Sobel, J. 2001. Using remote sensing to reassess the mass mortality of Diadema antillarum 1983-1984.Conservation Biology 15 ( 4), 885-891. Rigol, J. P ., and Chica-Olmo, M. 1998. Merging remote-sensing images for geological environmental mapping: application to the Cabo de Gata-Nijar Natural Park, Spain. Environmental Geology 34 (2/3): 194-202. Rubec, P. J. , Bexley, J. C. W., Norris, H. , Coyne, M. S. , Monaco, M. E., Smith, S. G. , and Ault, J. S. 1999. Suitability modeling to delineate habitat essential to sustainable fisheries. . Fish Habitat: Essential fish habitat and rehabilitation. Proceedings of the Sea Grant symposium of fish habitat. AFS Bethesda, MD. Sale, P. F., Chavez, E. A., Hatcher, B. G. , Mayfield, C. , and Ciborowski, J. H. 2001. Guidelines for developing a regional monitoring and environmental information system. Conservation and sustainable use of the Mesoamerican barrier reef system (MBRS) in Mexico, Belize, Guatemala, and Honduras. Stoffle, R. W. , Halmo, D. B. , Wagner, T. W., and Luczkovich, J. J. 1994. Reefs from space: Satellite imagery, marine ecology, and ethnography in the Dominican- Republic. Human Ecology 22 (3): 355-378. Smits, P. C. 2001. Combining supervised remote sensing image classifiers based on individual class performances. Lecture Notes in Computer Science 2096: 269- 278. Appendix A. Ground-Truthing Data for the 126 Stations. % Coraline % Hard % Soft Algae "live Waypoint Latitude N Longitude W Depth (m) % Sand % Seagrass % Algae Corals Corals rock" % Rubble 110 17'19'12.0 87'47'39.4 1.83 70 30 0 0 0 0 0 111 17'19'12.8 87'47'40.1 1.83 90 10 0 0 0 0 0 112 17'19'13.6 87'47'41 .0 2.13 65 5 30 0 0 0 0 113 17'19'14.3 87'47'41.4 2.74 85 0 0 0 0 0 15 114 17'19'14.8 87'47'42.2 2.44 60 0 30 0 10 0 0 115 17'19'15.4 87'47'43.1 1.52 70 25 5 0 0 0 0 116 17'19'15.9 87'47'43.9 0.91 0 100 0 0 0 0 0 126 17'16'49.1 87'48'34.0 2.74 30 30 0 5 20 15 0 127 17'16'48.6 87'48'34.4 3.05 45 0 25 5 25 0 0 128 17'16'47.2 87'48'34.3 3.35 50 0 30 5 15 0 0 129 17'16'46.3 87'48'34.4 3.35 60 0 15 0 20 0 5 130 17'16'45.3 87'48'34.8 3.05 0 0 0 20 20 60 0 131 17'16'44.6 87'48'35.5 3.66 40 5 0 10 25 20 0 132 17'16'44.4 87'48'36.3 3.66 40 0 0 15 20 25 0 133 17'16'44.5 87'48'37.5 3.35 I 30 0 0 20 25 15 10 134 17'16'50.6 87'48'34.9 2.74 50 50 0 0 0 0 0 135 17'16'50.0 87'48'34.4 2.74 65 35 0 0 0 0 0 136 17'16'49.3 87'48'33.7 2.74 70 30 0 0 0 0 0 137 17'16'48.6 87'48'32.9 2.74 55 45 0 0 0 0 0 138 17'16'48.0 87'48'32.1 2.44 0 0 5 0 0 90 5 139 17'16'47.4 87'48'31 .3 3.35 40 0 0 55 5 0 0 140 17'16'46.5 87'48'30.9 3.35 0 0 0 20 25 55 0 141 17'16'45.7 87'48'30.2 4.27 0 0 0 25 0 75 0 142 17'16'45.1 87'48'29.5 4.88 60 0 0 25 0 0 0 143 17'16'44.5 87'48'28.8 5.18 0 0 0 35 30 45 15 151 17'16'02.0 87'49'09.6 2.44 0 100 0 0 0 0 0 152 17'16'01 .3 87'49'08.8 2.13 55 45 0 0 0 0 0 Appendix A. Ground-Truthing Data for the 126 Stations. % Coraline % Hard % Soft Algae "live Watpoint Latitude N Lon9itude W Depth (m) % Sand % Sea9rass % Al9ae Corals Corals rock" % Rubble 153 17'16'00.6 87'49'08.1 1.22 0 0 0 20 10 70 0 154 17'15'59.7 87'49'07.6 1.83 0 0 0 0 20 80 0 155 17'15'59.3 87'49'06.6 2.74 0 0 0 85 0 15 0 156 17'15'58.5 87'49'05.8 4.27 0 0 0 5 50 45 0 157 17'15'58.2 87'49'04.9 4.88 0 0 0 45 25 30 0 158 17'15'57.7 87'49'04.1 3.96 0 0 0 15 0 85 0 170 17'19'04.7 87'47'33.2 6.10 0 0 0 35 10 55 0 169 17'19'03.8 87'47'33.6 6.71 0 0 0 5 5 90 0 168 17'19'03.6 87'47'34.3 5.18 0 0 0 35 15 50 0 167 17'19'02.7 87'47'34.1 5.18 50 0 25 10 15 0 0 166 17'19'02.5 87'47'33.8 6.40 0 0 5 15 0 80 0 165 17'19'01 .6 87'47'34.6 7.01 0 0 0 0 20 80 0 164 17'19'01 .8 87'47'35.0 5.18 0 0 20 5 45 30 0 163 17'19'01 .1 87'47'35.7 4.88 0 0 0 15 20 65 0 162 17'18'00.6 87'47'35.4 4.88 l 0 0 0 25 5 70 0 161 17'18'59.6 87'47'35.4 5.79 0 0 0 30 25 45 0 160 17'18'59.8 87'47'36.3 5.18 15 0 0 15 45 25 0 159 17'18'58.8 87'47'36.4 5.49 25 0 15 10 20 30 0 173 17'19'05.1 87'47'35.7 3.66 0 0 0 20 40 40 0 174 17'19'03.9 87'47'35.5 4.27 20 0 0 10 10 60 0 175 17'19'04.2 87'47'36.6 3.66 0 0 0 20 40 40 0 176 17'19'03.4 87'47'37.0 3.35 0 0 0 10 50 40 0 177 17'19'02.9 87'47'36.4 3.35 0 0 0 70 25 5 0 178 17'19'02.1 87'47'36.8 3.66 0 0 0 10 30 60 0 179 17'19'02.3 87'47'37.6 3.96 0 0 0 20 15 65 0 180 17'19'01 .5 87'47'38.1 3.66 65 0 5 0 15 15 0 181 17'19'00.9 87'47'37.1 4.57 0 0 0 35 35 30 0 ~ \Ci Appendix A. Ground-Truthing Data for the 126 Stations. % Coraline % Hard % Soft Algae "live Wa~point Latitude N Longitude W Depth (m) % Sand % Seagrass % Algae Corals Corals rock" % Rubble 182 17'19'00.2 87'47'37.5 4.57 0 0 0 10 0 90 0 172 17'19'00.3 87'47'38.3 3.96 0 0 0 30 15 55 0 171 17'18'59.7 87'47'38.8 3.96 0 0 0 0 40 60 0 183 17'16'19.2 87'48'43.6 8.84 0 0 0 25 30 45 0 185 17'16'18.3 87'48'43.9 8.84 10 0 0 25 25 40 0 186 17'16'18.7 87'48'44.7 7.01 0 0 0 15 15 55 20 187 17'16'17.8 87'48'45.1 6.71 0 0 0 25 25 50 0 188 17'16'17.5 87'48'44.2 8.53 10 0 0 30 30 40 0 189 17'16'16.7 87'48'44.7 7.32 0 0 0 5 50 25 20 190 17'16'17.0 87'48'45.5 5.79 0 0 0 25 25 15 0 191 17'16'16.2 87'48'46.1 6.10 0 0 0 40 40 10 0 192 17'16'15.7 87'48'45.1 7.01 0 0 0 35 35 30 0 193 17'16'14.7 87'48'45.1 7.62 0 0 0 5 55 15 25 194 17'16'14.8 87'48'46.2 6.40 5 0 0 40 35 20 0 195 17'16'13.8 87'48'46.4 6.71 '10 0 0 35 35 0 40 196 17'17'28.4 87'48'32.1 1.22 5 95 0 0 0 0 0 197 17'17'29.4 87'48'33.9 1.22 45 40 15 0 0 0 0 198 17'17'30.0 87'48'34.7 0.91 20 80 0 0 0 0 0 199 17'17'30.5 87'48'35.6 0.91 5 95 0 0 0 0 0 200 17'17'30.8 87'48'36.5 0.91 0 100 0 0 0 0 0 34 17'16'55.2 87'48'39.4 1.52 44 56 0 0 0 0 0 35 17'16'54.3 87'48'38.6 1.83 20 78 0 2 0 0 0 36 17'16'53.0 87'48'37.9 2.13 26 74 0 0 0 0 0 37 17'16'51 .9 87'48'37.0 2.44 32 68 0 0 0 0 0 38 17'16'51.1 87'48'36.0 2.74 50 50 0 0 0 0 0 39 17'16'50.0 87'48'35.4 2.74 74 26 0 0 0 0 0 40 17'16'49.3 87'48'34.6 3.05 62 38 0 0 0 0 0 Vl 0 Appendix A. Ground-Truthing Data for the 126 Stations. % Coraline % Hard % Soft Algae "live Wateoint Latitude N Lon9itude W Deeth (m) % Sand % Sea9rass % Al9ae Corals Corals rock" % Rubble 41 17'16'49.1 87'48'33.5 2.13 46 54 0 0 0 0 0 42 17'16'48.1 87'48'32.7 2.74 0 0 0 14 4 82 0 43 17'16'47.1 87'48'32.0 3.35 58 0 0 10 0 32 0 45 17' 16'45.8 87'48'31.3 3.96 42 0 0 0 14 0 20 46 17'16'44.6 87'48'30.6 4.27 40 0 0 0 12 18 30 47 17'16'43.5 87'48'30.2 4.88 50 0 0 10 8 6 26 48 17'16'42.2 87'48'29.2 5.79 4 0 0 28 0 68 0 49 17'16'40.9 87'48'28.5 6.10 0 0 0 32 4 64 0 50 17'16'39.8 87'48'26.5 10.36 72 0 0 0 6 0 11 51 17'16'40.8 87'48'27.3 7.32 4 0 0 22 0 60 14 55 17'17'41.9 87'48'20.0 1.22 100 0 0 0 0 0 0 56 17'17'42.6 87'48'21.4 1.22 88 0 12 0 0 0 0 57 17'17'43.2 87'48'22.5 1.22 88 0 12 0 0 0 0 58 17'17'43.9 87'48'23.7 1.52 97 0 3 0 0 0 0 59 17'17'44.4 87'48'25.3 0.91 I 97 0 3 0 0 0 0 60 17'17'44.9 87'48'26.6 0.91 97 0 3 0 0 0 0 61 17'17'45.5 87'48'27.8 0.91 98 0 2 0 0 0 0 62 17'17'46.1 87'48'29.0 1.22 91 0 9 0 0 0 0 63 17'17'46.6 87'48'30.3 0.91 94 0 6 0 0 0 0 64 17'17'47.3 87'48'31 .5 1.52 64 36 0 0 0 0 0 65 17'17'47.6 87'48'32.9 1.83 72 24 4 0 0 0 0 66 17'17'48.1 87'48'34.1 1.52 60 40 0 0 0 0 0 67 17'17'48.5 87'48'35.3 1.52 78 22 0 0 0 0 0 68 17'17'49.0 87'48'36.4 1.52 66 34 0 0 0 0 0 69 17'17'49.4 87'48'37.7 1.52 56 44 0 0 0 0 0 70 17'17'49.7 87'48'39.0 1.52 74 26 0 0 0 0 0 71 17'17'50.0 87'48'40.2 1.22 80 20 0 0 0 0 0 V-, Appendix A. Gmund-Truthing Data ,m tne 126 'Statio .., s. % Coraline % Hard % Soft Algae "live Waypoint Latitude N Longitude W Deeth (m) % Sand % Seagrass % Algae Corals Corals rock" % Rubble 72 17'17'50.4 87'48'41.5 0.91 62 32 0 0 0 0 0 73 17'17'50.8 87'48'43.0 0.91 80 20 0 0 0 0 0 74 17'17'51 .1 87'48'44.5 1.52 20 80 0 0 0 0 0 75 17'19'07.4 87'47'29.2 8.23 99 0 1 0 0 0 0 76 17'19'08.0 87'47'30.4 7.01 80 0 20 0 0 0 0 77 17'19'08.6 87'47'31.9 4.27 12 0 0 12 28 48 0 78 17'19'09.1 87'47'33.2 3.35 8 0 0 12 0 80 0 79 17'19'10.0 87'47'34.3 1.83 4 0 0 26 16 54 0 80 17'19'10.8 87'47'35.0 1.52 84 0 0 0 8 0 8 81 17'19'11 .8 87'47'35.5 1.52 62 0 0 2 0 22 14 82 17'19'12.9 87'47'35.8 2.74 38 0 0 0 8 0 54 83 17'19'13.6 87'47'36.7 3.66 100 0 0 0 0 0 0 84 17'19'14.5 87'47'37.7 0.91 77 10 0 0 5 0 8 85 17'19'14.5 87'47'38.7 1.83 76 8 0 6 2 8 0 86 17'19'14.6 87'47'40.0 2.44 l 92 0 0 0 8 0 0 87 17'19'14.9 87'47'41.2 2.44 74 0 0 0 8 18 0 88 17'19'15.3 87'47'42.5 1.83 40 0 20 0 0 0 40 89 17'19'16.1 87'47'43.4 0.61 0 100 0 0 0 0 0 Vl N ppend1x A. Ground:t ruthrnq Data to1 the 126 Stations. Optimization Clustering Classification wax:eoint Latitude N Lon9itude W (3- group) Band 1 Band 2 . Band 3 Band 4 Band 5 Band 7 Band 8 110 17'19'12.0 87'47'39.4 1 186 148 101 36 49 40 63 111 17'19'12.8 87'47'40.1 1 178 141 91 36 47 40 61 112 17'19'13.6 87'47'41 .0 1 185 144 92 37 47 41 59 113 17'19'14.3 87'47'41.4 1 170 131 90 36 47 40 56 114 17'19'14.8 87'47'42.2 1 174 132 91 38 48 40 56 115 17'19'15.4 87'47'43.1 1 169 128 89 38 48 40 58 116 17'19'15.9 87'47'43.9 2 148 118 93 39 50 42 55 126 17'16'49.1 87'48'34.0 1 160 117 84 36 47 39 52 127 17'16'48.6 87'48'34.4 1 157 116 85 35 47 40 53 128 17'16'47.2 87'48'34.3 1 154 111 83 35 48 40 53 129 17'16'46.3 87'48'34.4 1 154 111 83 35 48 40 51 130 17'16'45.3 87'48'34.8 3 151 111 82 35 48 41 53 131 17'16'44.6 87'48'35.5 1 147 108 82 35 47 39 52 132 17'16'44.4 87'48'36.3 1 148 110 83 35 48 39 52 133 17'16'44.5 87'48'37.5 ~ 149 110 83 35 47 41 51 134 17'16'50.6 87'48'34.9 1 166 125 87 35 47 41 55 135 17'16'50.0 87'48'34.4 1 161 120 87 35 47 41 53 136 17'16'49.3 87'48'33.7 1 161 119 84 36 47 41 54 137 17'16'48.6 87'48'32.9 1 158 116 83 35 50 39 50 138 17'16'48.0 87'48'32.1 3 148 109 82 35 46 41 54 139 17'16'47.4 87'48'31 .3 1 146 108 82 35 46 39 47 140 17'16'46.5 87'48'30.9 3 148 105 81 34 46 39 52 141 17'16'45.7 87'48'30.2 3 144 102 78 35 46 39 49 142 17'16'45.1 87'48'29.5 1 144 103 80 35 46 40 48 143 17'16'44.5 87'48'28.8 3 143 100 79 35 46 40 47 151 17'16'02.0 87'49'09.6 2 147 106 82 34 47 39 51 152 17'16'01 .3 87'49'08.8 1 147 106 82 34 46 39 49 Vl vJ Aopend•x A. Gmund-TruthmQ Oat;,1 for the 126 Statio11s. Optimization Clustering Classification Watpoint Latitude N Longitude W (3- group) Band 1 Band 2 Band 3 Band 4 Band 5 Band 7 Band 8 153 17'16'00.6 87'49'08.1 3 134 97 82 34 46 38 50 154 17'15'59.7 87'49'07.6 3 130 95 81 35 45 39 49 155 17'15'59.3 87'49'06.6 3 127 93 79 35 46 40 50 156 17'15'58.5 87'49'05.8 3 128 93 80 35 46 39 46 157 17'15'58.2 87'49'04.9 3 127 92 79 34 45 41 48 158 17'15'57.7 87'49'04.1 3 131 91 78 35 46 38 44 170 17'19'04.7 87'47'33.2 3 161 99 76 34 45 38 50 169 17'19'03.8 87'47'33.6 3 154 99 77 34 46 40 48 168 17'19'03.6 87'47'34.3 3 154 99 77 34 46 39 49 167 17'19'02.7 87'47'34.1 1 160 101 78 34 46 38 49 166 17'19'02.5 87'47'33.8 3 160 101 78 34 46 39 44 165 17'19'01 .6 87'47'34.6 3 155 99 77 34 46 39 47 164 17'19'01 .8 87'47'35.0 3 155 99 77 34 46 39 47 163 17'19'01.1 87'47'35.7 3 143 96 77 33 45 38 45 162 17'18'00.6 87'47'35.4 3 143 96 77 33 45 37 46 161 17'18'59.6 87'47'35.4 3 149 98 78 34 46 39 48 160 17'18'59.8 87'47'36.3 3 149 98 78 35 45 39 46 159 17'18'58.8 87'47'36.4 3 138 95 76 35 47 39 46 173 17'19'05.1 87'47'35.7 3 142 101 79 34 45 38 50 174 17'19'03.9 87'47'35.5 3 142 100 78 34 46 38 48 175 17'19'04.2 87'47'36.6 3 138 100 78 34 45 39 49 176 17'19'03.4 87'47'37.0 3 138 100 78 33 42 37 48 177 17'19'02.9 87'47'36.4 3 133 96 75 · 33 42 37 46 178 17'19'02.1 87'47'36.8 3 136 95 77 33 42 37 44 179 17'19'02.3 87'47'37.6 3 138 99 78 33 43 37 49 180 17'19'01 .5 87'47'38.1 1 138 99 78 34 47 41 47 181 17'19'00.9 87'47'37.1 3 135 96 77 35 46 39 48 Vl .+>, Appendix A. Gro•md-Truthmg Data tor the 126 t,tat1on Optimization Clustering Classification Wateoint Latitude N Lon9itude W (3- group) Band 1 Band 2 Band 3 Band 4 Band 5 Band 7 Band 8 182 17'19'00.2 87'47'37.5 3 142 99 80 35 46 39 48 172 17'19'00.3 87'47'38.3 3 142 99 80 35 48 39 50 171 17'18'59.7 87'47'38.8 3 141 101 81 34 46 37 50 183 17'16'19.2 87'48'43.6 3 140 95 79 35 46 38 46 185 17'16'18.3 87'48'43.9 3 139 92 79 34 46 39 47 186 17'16'18.7 87'48'44.7 3 138 96 76 34 47 40 46 187 17'16'17.8 87'48'45.1 3 140 95 80 34 47 40 48 188 17'16'17.5 87'48'44.2 3 139 92 78 34 46 39 48 189 17'16'16.7 87'48'44.7 3 142 95 76 34 47 40 46 190 17'16'17.0 87'48'45.5 3 140 95 79 34 47 40 48 191 17'16'16.2 87'48'46.1 3 139 93 79 34 47 40 48 192 17'16'15.7 87'48'45.1 3 136 92 78 34 48 40 47 193 17'16'14.7 87'48'45.1 3 136 91 78 34 47 38 43 194 17'16'14.8 87'48'46.2 3 135 94 78 34 46 38 47 195 17'16'13.8 87'48'46.4 2 ' 136 93 78 34 46 39 48 196 17'17'28.4 87'48'32.1 2 162 134 80 36 48 40 48 197 17'17'29.4 87'48'33.9 1 148 121 109 36 49 40 55 198 17'17'30.0 87'48'34.7 2 148 120 102 36 50 42 58 199 17'17'30.5 87'48'35.6 2 151 119 102 37 49 38 60 200 17'17'30.8 87'48'36.5 2 144 114 100 37 48 41 55 34 17'16'55.2 87'48'39.4 1 141 108 100 37 48 40 54 35 17'16'54.3 87'48'38.6 2 138 109 89 37 49 40 53 36 17'16'53.0 87'48'37.9 2 145 111 87 36 48 40 53 37 17'16'51.9 87'48'37.0 1 162 122 85 36 48 40 54 38 17'16'51 .1 87'48'36.0 1 164 122 85 36 47 40 56 39 17'16'50.0 87'48'35.4 1 160 119 87 36 49 40 54 40 17'16'49.3 87'48'34.6 1 161 120 85 35 47 41 53 Vi Vi Appendix A. Ground-Truth1n!:I Data tor the 126 '3tat,ons. Optimization Clustering Classification Watpoint Latitude N Longitude W (3- group) Band 1 Band 2 Band 3 Band 4 Band 5 Band 7 Band 8 41 17'16'49.1 87'48'33.5 1 , 160 117 87 36 47 39 52 42 17'16'48.1 87'48'32.7 3 153 111 84 35 50 41 53 43 17'16'47.1 87'48'32.0 1 149 106 81 35 46 39 48 45 17' 16'45.8 87'48'31 .3 1 • 147 105 81 35 47 41 50 46 17'16'44.6 87'48'30.6 1 145 101 81 35 48 40 49 47 17'16'43.5 87'48'30.2 1 143 102 83 36 47 42 50 48 17'16'42.2 87'48'29.2 3 151 103 83 35 50 40 47 49 17'16'40.9 87'48'28.5 3 143 96 79 34 47 38 47 50 17'16'39.8 87'48'26.5 1 152 95 79 35 46 41 47 51 17'16'40.8 87'48'27.3 3 143 94 78 35 45 40 48 55 17'17'41.9 87'48'20.0 1 199 176 78 35 47 41 71 56 17'17'42.6 87'48'21.4 1 197 171 134 35 49 40 71 57 17'17'43.2 87'48'22.5 1 191 167 127 35 47 42 65 58 17'17'43.9 87'48'23.7 1 191 167 120 35 46 39 67 59 17'17'44.4 87'48'25.3 1 l 195 169 121 35 46 38 67 60 17'17'44.9 87'48'26.6 1 203 176 123 35 46 40 69 61 17'17'45.5 87'48'27.8 1 201 172 131 35 47 38 66 62 17'17'46.1 87'48'29.0 1 197 172 127 35 46 39 68 63 17'17'46.6 87'48'30.3 1 201 176 128 34 47 39 72 64 17'17'47.3 87'48'31 .5 1 206 180 133 34 47 40 66 65 17'17'47.6 87'48'32.9 1 168 138 142 35 46 37 58 66 17'17'48.1 87'48'34.1 1 167 139 104 34 45 39 59 67 17'17'48.5 87'48'35.3 1 162 133 103 34 46 40 57 68 17'17'49.0 87'48'36.4 1 174 137 102 34 46 40 62 69 17'17'49.4 87'48'37.7 1 180 152 104 34 46 39 62 70 17'17'49.7 87'48'39.0 1 179 148 110 34 47 39 60 71 17'17'50.0 87'48'40.2 1 176 153 113 35 47 41 60 Vl 0\ Appendix A. Ground-Truthing Data for the "1 26 Stat,ons. Optimization Clustering Classification Watpoint Latitude N Longitude W (3- group) Band 1 Band 2 Band 3 Band 4 Band 5 Band 7 Band 8 72 17'17'50.4 87'48'41 .5 1 189 163 116 35 48 41 66 73 17'17'50.8 87'48'43.0 1 192 161 123 35 48 40 67 74 17'17'51 .1 87'48'44.5 2 177 146 127 35 46 40 63 75 17'19'07.4 87'47'29.2 1 168 100 112 34 46 39 48 76 17'19'08.0 87'47'30.4 1 · 173 109 78 34 48 40 50 77 17'19'08.6 87'47'31 .9 3 168 108 80 35 47 40 49 78 17'19'09.1 87'47'33.2 3 141 102 79 34 · 45 38 49 79 17'19'10.0 87'47'34.3 3 151 113 81 34 46 39 48 80 17'19'10.8 87'47'35.0 1 152 118 85 34 46 38 48 81 17'19'11 .8 87'47'35.5 1 153 121 91 38 48 39 64 82 17'19'12.9 87'47'35.8 1 170 138 103 37 49 42 57 83 17'19'13.6 87'47'36.7 . 1 186 144 106 36 50 42 62 84 17'19'14.5 87'47'37.7 1 196 155 99 37 47 40 61 85 17'19'14.5 87'47'38. 7 1 196 152 103 36 49 41 61 86 17'19'14.6 87'47'40.0 1 l 187 146 99 36 49 41 60 87 17'19'14.9 87'47'41 .2 1 174 132 96 36 48 40 61 88 17'19'15.3 87'47'42.5 1 169 128 91 38 48 40 58 89 17'19'16.1 87'47'43.4 2 148 118 89 39 50 42 55 Vl --..J Appendix B. Canonical Scores of the 78 Stations Used in the Analysis. Canonical score Canonical score Habitat Case Number factor 1 factor 2 1 (Sand) 1 2.235 0.334 2 1.61 0.114 3 1.075 -0.257 4 1.892 -1.904 7 -0.048 0.203 8 0.33 -0.527 47 0.385 -0.689 48 -0.094 0.255 52 2.513 2.089 53 3.586 0.602 54 3.179 0.389 55 3.016 0.571 56 3.176 0.827 57 3.615 1.255 58 3.613 0.989 59 3.463 0.756 60 3.216 1.792 61 3.562 1.973 62 2.045 -1.136 63 0.701 0.577 64 0.368 0.348 65 0.675 1.274 66 1.37 1.182 67 2.083 -0.123 68 2.736 0.618 69 2.861 0.743 71 -0 .652 1.532 75 -0.789 0.389 76 2.192 0.315 77 2.123 0.526 2 (Coral Reef) 6 -0.665 -0.26 9 -0.787 -0.457 10 -1.457 0.516 11 -1.215 -0.496 13 -1 .951 -0.468 14 -1.624 -1.56 15 -1.755 -1.647 16 -2.319 -0.841 17 -1.838 -1 .287 18 -1.627 1.907 19 -1.703 1.291 59 Appendix B. Canonical Scores of the 78 Stations Used in the Analysis. 20 -1.703 1.291 21 -1.515 1.715 22 -1.689 1.375 23 -2.466 1.284 24 -2.466 1.284 25 -1.791 0.872 26 -1.304 0.036 27 -1.748 0.173 28 -1 .868 -0.11 29 -2.355 0.726 30 -2.606 0.722 31 -2.395 0.753 32 -1.606 -1.062 33 -1.317 -0.634 34 -1.317 -0 .634 35 -1.715 0.037 36 -1 .528 -0.666 37 -2.149 0.169 38 -1.992 0.144 39 -2.172 0.195 40 -2.056 0.417 49 -0.59 -0.1 43 50 -0.96 -0.064 51 -1.933 0.396 72 -0.589 1.307 73 -1.722 0.061 74 -1.095 0.547 3 (Seagrass) 5 1.776 -4.337 12 -1.408 0.378 41 0.848 -0 .754 42 0.603 -2.118 43 1.093 -2.674 44 0.748 -3.073 45 0.208 -3.154 46 -0.147 -1 .731 70 2.144 -0.209 78 1.683 -4.233 Appendix C. CD of Digital Images of the Ground-Truthing Stations. wp_ 11 O.bmp wp_111.bmp wp_ 112.bmp wp_ 112a.bmp wp_ 113.bmp wp_ 114.bmp wp_ 115.bmp wp_ 116.bmp wp_ 126.bmp wp_ 127.bmp wp_ 128.bmp wp_ 129.bmp wp_ 130.bmp wp_ 131.bmp wp_132.bmp wp_133.bmp wp_134.bmp wp_135.bmp wp_ 136.bmp wp_ 137.bmp wp_138.bmp wp_139.bmp wp_140.bmp wp_141.bmp wp_142.bmp wp_143.bmp wp_151.bmp wp_152.bmp wp_153.bmp wp_154.bmp wp_155.bmp wp_156.bmp wp_ 157.bmp wp_ 158.bmp wp_ 159.bmp wp_ 160.bmp wp_ 161.bmp wp_ 162.bmp wp_ 163.bmp wp_ 164.bmp wp_ 165.bmp wp_ 166.bmp wp_ 167.bmp wp_ 168.bmp wp_ 169.bmp wp_ 170.bmp wp_171.bmp wp_172.bmp wp_ 173.bmp wp_ 174.bmp wp_ 175.bmp wp_ 176.bmp wp_ 177.bmp wp_ 178.bmp wp_179.bmp wp_180.bmp wp_181.bmp wp_182.bmp wp_ 186.bmp wp_ 187.bmp wp_188.bmp wp_189.bmp wp_190.bmp wp_191.bmp wp_ 192.bmp wp_ 193.bmp wp_ 194.bmp wp_ 195.bmp wp_ 196.bmp wp_ 197.bmp wp_198.bmp wp_199.bmp wp_200.bmp