Due to the important role that seagrasses play in estuaries, there has been a considerable effort at developing sampling and mapping techniques to quanitfy the spatial distribution, biomass and health of seagrass communities and monitor changes over time (Caloz and Collet, 1997; Lehmann and Lachavanne, 1997). Remote sensing approaches have seen increasing application to the mapping of seagrass beds due to their synoptic perspective and cost-effective mapping over large areas. The most widely adopted approach has been the visual interpretation and mapping from aerial photography (Zieman et al., 1989; Ferguson et al. 1993; Robbins 1997; Kendrick et al., 2000; Moore et al. 2000). More recently, this has transitioned to a direct capture through digital photogrammetric techniques or heads-up digitizing of digital rectified photography (Dobson et al., 1995).
In this study, we examine the utility of object-oriented image segmentation/classification approaches (Burnett and Blaschke, 2003; Benz et al., 2004) to seagrass mapping from airborne digital camera imagery. Our goal was to develop a methodology that was comparatively objective in delineating bed boundaries and characterizing seagrass density, was cost-effective and easily repeatable for future monitoring purposes. We apply this conceptual framework to the mapping and spatial analysis of seagrass beds and the broader benthic environment in Barnegat Bay-Little Egg Harbor estuary in New Jersey, USA.
SAV is a key indicator of the health of the Barnegat Bay estuary, and is under constant stress from a number of sources. These stresses cause changes in seagrass bed characteristics which in turn make frequent monitoring of this habitat type necessary.
2. The SAV 2003 classification data are formatted as Arc/Info raster grid. The grid is a mosaic of 14 individual raster grid SAV classifications in Barnegat Bay - Little Egg Harbor. For dissemination purposes, the raster grid was converted to an Arc/Info vector coverage format, then exported to an Arc/Info interchange file (*.e00).
3. Our remotely sensed approach allowed for determination of seagrass at 4 levels of density (including shallow sand/mud flats with < 10% seagrass cover), rather than a simple presence/absence, with a comparatively high degree of consistency and accuracy (68% overall accuracy for 4 categories and 83% accuracy for a simpler presence/absence map).
4. Other SAV GIS data available on the CRSSA web are SAV for the 1968, 1979, 1985-87 and 1996-99 time periods. Please follow the cross-reference links in the metadata.
5. PLEASE READ ALL METADATA.
While efforts have been made to ensure that these data are accurate and reliable within the state of the art, Rutgers University cannot assume liability for any damages, or misrepresentations, caused by any inaccuracies in the data. Rutgers University makes no warranty, expressed or implied, nor does the fact of distribution constitute such a warranty.
Any maps, publications, reports or any other type of document produced as a result of utilizing this data will credit the original map author(s) as listed as well as the Center for Remote Sensing and Spatial Analysis (CRSSA), Rutgers University.
Accuracy Assessment
The resulting maps were compared with the 245 field reference points. All 245 reference points were used to support the interpretation and mapping in some fashion and so can not be truly considered as completely independent validation. The resulting maps were also compared with an independent set of 41 bottom sampling points collected as part of a seagrass-sediment study conducted during the summer of 2003 by the New Jersey office of the Natural resources Conservation Service (Chris Smith and David Friedman). These additional 41 bottom sample points were collected in an area along the eastern shore of central Barnegat Bay in an area deemed of high image quality. At each sampling point, a sediment grab sample was taken and the presence/absence of seagrass determined for an approximately 5m2 area. The spatial locations of the 41 sampling points were recorded using a non-differentially collected GPS receiver (Garmin Map 12) with an approximate positional error of ± 15m. The presence/absence data for the 245 and 41 sampling points were compared with the same location from the digital seagrass map and summarized in a contingency table and producer's/user's accuracy and Kappa statistic (a measure of agreement corrected for chance agreement) computed.
The seagrass density data for the 245 field reference points were categorized into 4 seagrass density classes (absent, sparse, moderate and dense), compared with the same location from the digital seagrass map and summarized in a contingency table (Table 1a). The overall accuracy was 68.2% and Kappa statistic was 56.5%, which can be considered as a moderate degree of agreement between the two data sets. Aggregating the data into a simple presence vs. absence comparison (Table 1b) shows a higher level of agreement with an overall accuracy of 82.8% and a Kappa statistic of 63.1%. Examination of Table 1b reveals that most of the disagreement was due to a high error of omission, i.e., a number of points confirmed as seagrass in the field sampling data were not mapped as seagrass (32 out of 245 points or 13.1%). 20 out of these 32 points (62.5 %) were categorized as Sparse Seagrass (i.e., 10-39%) in the field.
The presence/absence data for the 41 independent sampling points were compared with the same location from the digital seagrass map and summarized in a contingency table (Table 2). The overall accuracy was 70.7% and Kappa statistic was 43%, which can be considered as a moderate degree of agreement between the two data sets. Examination of the Table 2 reveals that most of the disagreement was due to a high error of commission, i.e., a number of points mapped as seagrass were not confirmed as seagrass in the field sampling data (9 out of 41 points or 22.0%). These 9 points were relatively equally spaced across the 3 categories of seagrass density (3 in 10-39%, 2 in 40-79%, and 4 in 80-100%).
The agreement between the mapped results and the original field reference as well as independent reference data were only moderate (i.e., 68% for the 4 category map and 83% for the presence/absence map based on the original field reference data and 71% for the simple presence/absence map as compared to independent reference data). The comparison with the original reference data suggests that most of the error is due to the omission of Sparse Seagrass beds. These results are similar to Moore et al. (2000) who found that their aerial photo-interpretation tended to underestimate percent cover at low SAV densities. It should also be noted that while the imagery was collected in early May, the field reference points were not sampled until after the imagery collection, in some cases up to several weeks later. Thus reference points that may not have had distinctly visible seagrass at the time of data collection only to have sparse seagrass densities later in the growing season. A majority of the disagreement in the independent data comparison was due to a comparatively high error of commission and may not be a true measure of the map accuracy but rather be due to: 1) the mismatch between the footprint area of the reference sample in relation to the size of the minimum mapping unit for the seagrass maps; and 2) high positional error (± 15m) of the reference samples. Due to the natural fine scale patchiness within even dense beds, the comparatively small footprint of the reference data (approximately 5m2) could sample bare patches (i.e., below the minimum mapping unit size of 1 ha) within an otherwise extent bed. Likewise, the high positional error (± 15m) of the reference samples coupled with the fine scale patchiness could also result in a disagreement between the reference data and the mapping.
Tables 1a, 1b and 2 are available in the supplemental document, 'sav2003_Accuracy_Assessment.pdf'
Field Surveys
To support the image interpretation and mapping, extensive field reference data were collected in the weeks before and after the image acquisition. Existing maps of seagrass distribution derived from boat based surveys from the mid 1990's were used to plan the reference data collection. A series of transects were established to sample full range of conditions in the Barnegat Bay-Little Egg Harbor study area. Eighteen transects were visited, and data points were recorded at intervals of approximately 250 meters. The transects, perpendicular to the eastern (barrier island) shoreline, extended from shallow inshore areas, across the seagrass beds and into deeper mid-bay water. These intervals were not strict, and often data points were recorded in between the intervals at areas where there appeared to be a noticeable change in seagrass coverage. Additional reference points were also collected to spot check areas of uncertainty. In general, field reference points were collected in areas (i.e., approximately 5 x 5 meters area) where the seagrass bed was reasonably consistent in coverage and distribution. ESRI's ArcMap and Trimble's GPS Pathfinder were used to support the field reference data collection.
All transect endpoints and individual check points were first mapped on the GPS, endpoints were then loaded onto a GPS for navigation on the water. Real time data collection (approximately ± 1-3 meter accuracy) was a sufficient level of accuracy for our purposes. A total of 245 transect and individual points were collected (Figure 2). The objective was to understand how bed characteristics changed from shallow to deep water, and to be able to understand the difference in visual signal on the imagery between beds in shallow (less than or equal to 1.5 m) and deep water (greater than 1.5 m). All 245 points were used to support the interpretation and mapping, none were reserved as independent validation. For each field reference point, the following data were collected: " GPS location (UTM) " Time " Date " SAV species presence/dominance: Zostera marina or Ruppia maritima or Algae " Depth (meters) " % cover (10 % intervals) determined by visual estimation " Blade Height of 5 tallest seagrass blades " Shoot density (# of shoots per 1/9 m2 quadrat that was extracted and counted on the boat) " Distribution (patchy/uniform) " Substrate (mud/sand) " Additional Comments
Classification
From a landscape ecology perspective as well as from our object-oriented classification approach, the spatial structure of the seagrass habitats was conceptualized at 3 different levels: 1) bed, a spatially contiguous area of seagrass of varying % cover composition; 2) density class, a spatially contiguous area of overall similar % cover composition; and 3) patch, small discrete clumps of seagrass or areas of open bay bottom. This conceptual spatial framework was then broadened to develop a hierarchical classification scheme to encompass the larger bay system for mapping purposes. The bay was categorized into 4 levels of attribute detail (Figure 3). Level 1 differentiated land and emergent wetlands from open water. Level 2 differentiated deep water/channels (> 1.5-2 m depth) and shallow water (<1.5-2m depth) bottom habitats. At Level 3, the shallow bottom habitats were then differentiated into: 1) shallow sand/mud flats (<1.5m depth); macro algae beds (i.e. Ulva lactuca and assorted macro algae dominated; scattered seagrass may be present) and seagrass beds (i.e., Zostera marina and Ruppia maritima). At Level 4, the seagrass beds were partitioned into 3 categories based on the % cover: dense (80-100 % coverage), moderate (40-80% coverage), sparse (10-40% coverage). At Level 5, were the individual patches of seagrass or bare bottom. The very detailed Level 5 delineations were not included in the final output maps.
While this seagrass classification does not represent equal % cover intervals, the class breaks were based on thresholds that appeared to be consistently discernable in both the image interpretation and corresponding field data. These seagrass density class ranges are similar to the scheme used by Moore et. al., 2000. The relative dominance of Zostera vs. Ruppia was not distinguished. The shallow sand/mud flats can in some ways be considered as potential seagrass habitat as our field surveys showed that seagrass was often present at low levels (i.e., < 10% cover). In these cases, the seagrass generally did not form cohesive clumps but rather a sparse and/or discontinuous covering of individual seagrass plants. Previous experience has shown that some of these areas develop denser cover of seagrass later in the growing season. This may especially be true in the more mesohaline areas of the bay where Ruppia is the dominant seagrass.
An object-oriented classification approach was performed using eCognition software (Standard Version 3.0) to segment the image into image objects. Image objects are delineated to minimize within object variance and maximize between object variances. A multi-resolution segmentation can be used to create a hierarchical framework of decomposable image objects (Benz et al., 2004). In other words, a super-object is composed of objects which in turn can be composed of sub-objects. As sub-objects are aggregated to form an object, interior boundaries disappear but exterior boundaries remain stable. This multi-resolution approach was adopted to segment the water portion of the image into 3 general levels of spatial detail using what is termed a classification-based multi-resolution segmentation (CBMS). The first step was to segment the image at a fine level of detail, which corresponded with our conceptual Level 5, i.e., the individual patches of seagrass. The size of a minimum mapping unit for the individual seagrass beds was on the order of 1 ha in size. Next, the segmentation was coarsened to the next higher level of aggregation (determined by the scale parameter), corresponding to conceptual Level 4 where individual sub-object (patches) are combined to create image objects (macro-patches) of similar density class. Finally, the objects (macro-patches) were combined into super-objects to correspond with the conceptual model Level 3 seagrass beds. Within the eCognition software environment, segmentation parameters can be weighted to take into account object scale, color and shape factors; resulting in drastically different image objects. Optimizing these parameters for the study at hand was an iterative trial and error process. While there was no clear correct set of parameters, certain parameter combinations (affected heavily by the scale parameter) made for more useful image object arrangements than others. These parameters differed from one image mosaic to the next because each image mosaic's radiometry and geographic extent were unique. Once the objects are delineated, they can then be classified using a rules-based approach. While initially we proposed to develop a "universal" set of rules to classify the seagrass and bottom types in a comparatively automated classification approach across the entire study area, we realized that a more manual, analyst assisted approach was necessary. As in determining the segmentation parameters, due to the variability in spectral response between the individual digital photos and the image mosaics as well as the spectral variation of seagrass across varying % cover, water clarity, depth and substrate, it was difficult to determine a set of universally applicable rules.
The following approach was adopted to map bottom types: 1) the entire image was segmented at a fine (i.e, Level 5); 2) using the clear distinction between land and water in the near infrared waveband image, a simple NIR membership rule was established to mask out land; 3) the image segmentation was then coarsened to merge areas of like classes (i.e., Level 4); 4) the Level 4 image objects were visually interpreted and manual encoded as to the appropriate bottom type (Figure 2) with the help of field reference data; 5) the class coding was "forced down" to the level 5 sub-objects; 6) the Level 5 sub-objects were then visually evaluated atop the original imagery to ensure that a proper identification was made and the classification revised where necessary; and 7) the revised Level 5 sub-objects were then transmitted back up the hierarchy and the Level 4 image objects revised accordingly (this was done by specifying "existence based on sub-objects" as a rule for each class). This approach expedited the process by undertaking the manual classification at a coarser scale with fewer objects to code but without losing the boundary detail afforded by the more detailed segmentation. Using the above approach, each of the 14 image mosaics were classified independently and merged to create a complete bay-wide classification. In addition, to the difficulty in developing consistent classification rules across mosaics, a super-mosaic of all 14 sub-areas would have required enormous computational power and time for a multi-resolution segmentation.
Quality Assurance
The resulting maps were compared with the 245 field reference points. All 245 reference points were used to support the interpretation and mapping in some fashion and so can not be truly considered as completely independent validation. The resulting maps were also compared with an independent set of 41 bottom sampling points collected as part of a seagrass-sediment study conducted during the summer of 2003 (Smith and Friedman, 2004). These additional 41 bottom sample points were collected in an area along the eastern shore of central Barnegat Bay in an area deemed of high image quality. At each sampling point, a sediment grab sample was taken and the presence/absence of seagrass determined for an approximately 5m2 area. The spatial locations of the 41 sampling points were recorded using a non-differentially collected GPS receiver (Garmin Map 12) with an approximate positional error of ± 15m. The presence/absence data for the 245 and 41 sampling points were compared with the same location from the digital seagrass map and summarized in a contingency table and producer's/user's accuracy and Kappa statistic (a measure of agreement corrected for chance agreement) computed.
Spatial Pattern Analysis
Using the resulting study area-wide classified GIS map, we examined the spatial structure of the seagrass beds by analyzing the spatial pattern of seagrass density classes and their shared edge lengths. The health and productivity of seagrass is highly dependent on an adequate amount of solar illumination which in turn is heavily influenced by the water clarity (Dennison et al., 1993). Seagrass beds in deeper water or seagrass at the deep water edge of the bed are therefore more vulnerable to turbid water conditions and a limited light environment. The classified map was used to analyze seagrass adjacency to deep water to highlight areas of greatest vulnerability as well as to examine within bed spatial structure. The amount of border to each contiguous seagrass patch was calculated, and expressed as a percentage of the total border. The Level 3 seagrass beds (i.e., 3 classes of seagrass density grouped together) were coarsened to 5m grid cell resolution and analyzed using the ArcINFO Version 8.3.
While efforts have been made to ensure that these data are accurate and reliable within the state of the art, Rutgers University cannot assume liability for any damages, or misrepresentations, caused by any inaccuracies in the data. Rutgers University makes no warranty, expressed or implied, nor does the fact of distribution constitute such a warranty.
Any maps, publications, reports or any other type of document produced as a result of utilizing this data will credit the original map author(s) as listed as well as the Center for Remote Sensing and Spatial Analysis (CRSSA), Rutgers University.