(TEXT EXTRACTED FROM THE TECHNICAL REPORT)
Current (2009) extent of seagrass across the BB-LEH system
The first objective of this project was to quantify the location of seagrass across the BBLEH estuary system for the 2009 growing season. To accomplish this aerial photography dataset was collected, processed into image objects (polygons), and classified to create a GIS dataset showing the location of seagrass across the BB-LEH. The following methods sections describe the steps used to create the output GIS dataset. In addition an accuracy assessment was undertaken to determine how well this GIS dataset maps seagrass across the BB-LEH.
a). Aerial Photography Collection
An aerial photography mission was undertaking during the summer of 2009. Film aerial photography was collected on June 28, July 7, and August 4, 2009, using a Navajo HS airplane equipped with a Leica RC30 camera, lens # 13234, focal length 152.720 mm, and a variable exposure time of 260-420 milli-seconds. Two types of film were used; a grey scale AGFA 80 and color film AGFA 100. The same plane and camera was used for all three imaging missions. The plane flew at an altitude ~ 3,658 m and speed of 180 km hr-1 per hour. The plane flew three survey lines, two in the southern estuary due to bay width and one in the northern estuary for both the June 28 and July 7 aerial flyover, the August 4 date was only flown to collect imagery in the northern part of the study area. Two passes were made per day, the first to collect black and white photography and the second to collect color photography. The resultant film was then processed and scanned through a high resolution scanner resulting in a digital image with 18,278 by 18,292 pixels in a scale of 1 to 2,000. These scans were ortho-rectified and projected into Universal Transverse Mercator (Zone 18 North, North American Datum 1983, GRS Spheroid of 1980) with a horizontal positional accuracy at root mean square error of ±1- 2 meters. The resulting geo tiffs were mosaicked into 15 larger blocks for later analysis.
b). In situ data collection
A number of in situ sites were visited to collect reference information to enable the interpretation of the aerial photography (Figure 1 & Appendix I). Reference sites were selected to match a subset of the in situ references sites selected during the 2003 Lathrop study (Lathrop et al. 2006). Reference sites were not selected in a random probabilistic manner, but rather targeted transects across the study area n = 167. In addition, 15 sample sites were selected for a late season review (October of 2009) for areas of uncertainty in the imagery. An additional 120 sample points were collected in June 2009 as part of an ongoing research project (Kennish 2009 unpublished data). These data points were also included in the study as field reference sites, although their collection used a different technique than the data points used in this study. A second in situ n = 124 dataset was collected to provide a validation dataset which was selected using a stratified random sampling design to focus on shallow water habitats mimicking the depth distribution of seagrass within the BB-LEH estuary. These points were distributed to match the depth distribution on the 2003 seagrass survey. To accomplish this, 2003 seagrass presence absence data from (Lathrop et al 2006) was intersected with the NOAA Nautical Charts Depth information (Charts 12324: edition 25, 1990 and 12316: edition 25, 1992 from Lathrop et al. (2001). For each 0.3048 meter depth (1 foot) category a number of field sites were randomly chosen to match the percentage of area of all seagrass habitat at that depth. This matched the random seagrass sites depth histogram to the depth histogram of the presence/absence seagrass data from 2003. These points were distributed to match the probability of finding seagrass at a specific depth. This validation dataset was not used in the image mapping and classification process but kept as an independent data set to compare with the wall-to-wall GIS map to create an error matrix, a producer's and user's accuracy assessment, and a Kappa statistic. As a secondary step after the accuracy assessment was completed the validation dataset was used to clean up the final GIS dataset.
For all of the in situ data collected for this project (the reference dataset n = 167 and the validation dataset n=124), field collection was accomplished as follows. The field survey was conducted from the Rutgers University Marine Field Station (RUMFS) using a 20 foot maritime skiff. Navigation to field locations was accomplished with a Garmin 530s marine GPS/Sonar system. Upon arrival at the preselected field locations, the boat weighed anchor. Next, an L shaped 4 meter x 5 meter grid made of 1.905 cm pvc (figure. 3) was lowered over the side of the boat. A diver entered the water and affixed a GPS Magellan Mobile Mapper 6 (±2-5 meter horizontal accuracy) to the outside L of the survey grid (marked in Figure 2). A compass reading was taken along the left-hand axis of the sampling grid. The compass reading and the GPS position allowed precise placement of the sampling grid on the benthos to a higher level of accuracy than the boatbased GPS unit. The diver then visited grid 1 through 8 and recorded information on SAV presence / absence (yes no), percent cover of seagrass species (R. maritima and Z. marina) (0 to 100 in 10% increments), and percent coverage macroalgae (0 to 100 in 10% increments). This data was verbally relayed to the boat captain who recorded the data on write-in-the-rain paper. Upon completion of field data collection, the GPS unit was removed and the sampling grid returned to the boat. Field sheets were then signed, dated, and entered into a digital database. The precise location of each sampling grid was determined using MatlabT and simple geometry using the GPS location in UTM coordinates and the compass bearing. A correction for magnetic declination (difference between the North Pole and the magnetic North Pole) was calculated using NOAA website (http://www.ngdc.noaa.gov/geomagmodels/Declination.jsp) for July 15th, 2009, 39.9745 N 74.1514 W magnetic declination equals 12 degrees and 47 minutes.
c) Image pre-processing
An important step in image classification is the clumping of similar pixels into image objects for classification. To accomplish that task, each image collected in 2009 was filtered using the aggregate command available in Arc GridT for a 2x2 grid window selecting the median cell value. This was done to remove areas of local light scatter from wave tops, Langmuir circulation lines, and to reduce the size of the imagery for processing. The median was selected over the mean to avoid skewing from light scattering which can cause areas of high image reflectance (white capping) and shadows. The rectified mosaicked color photography was then imported into eCognition(TM) to support image segmentation and classification. eCognition(TM) is an image analysis software package that segments raster data in an unsupervised method minimizing the intra-polygon (image object) variance while maximizing inter-polygon (image object) variance. The user can control the weight of each imagery band by changing a coefficient between 0 and 1 (0 no input for that band 1 full input) by band and a unit less scale parameter which determines the average image object area. As the scale parameter increases greater spectral heterogeneity is allowed increasing the average size of the image objects. Multiple scale image objects can be created by running a multiple resolution segmentation procedure. Two-scale parameters were used for each image mosaic layer 1); a small scale parameter between 10-15; 2) a large-scale parameter 50-70 (Figure 3). The smaller scale parameter resulted in image objects with a mean size of .073 ha, mode of .045 ha, 25 percentile of .02 ha, and the 75 percentile at .09. This scale parameter was selected to meet the target minimum mapping unit of .05 ha (500 m^2). The minimum mapping unit defines the smallest feature delineated in the map or the amount of detail a map contains. The band coefficients used were 1 for blue, 0.7 for green, and 0.5 for red. The coefficients were selected by trial and error by the operator to maximize the difference between seagrass and other benthic habitats.
d). Image object classification
A manual classification where each image object was visually interpreted and assigned to one of four classes of seagrass density (high 100-80% percent cover, medium < 80%- 40% cover, sparse > 40% and <= 10% cover, and no seagrass <10% - 0%). The field reference data was used to inform the interpretation. The larger scale image objects (scale parameter 50-70) were first manually classified using eCognitionT. The large image object classifications were then forced down into the smaller image objects (scale parameter of 10-15) based on the nested polygon structure. Smaller image objects on edge areas and internal to the larger image objects were then manually reclassified when necessary. This method sped up the manual classification effort allowing large contiguous areas of seagrass to be classified quickly while also allowing precise classification on seagrass edge and gap areas (Figure 4 from Lathrop et al. 2006). The reference data also contained information on seagrass species and macroalgae percent cover these categories were not mapped as part of the manual classification. To create the final GIS dataset and accuracy assessment dataset the finer-scale image objects were exported to Environmental Research Institute ESRIT shapefile format. To determine how well the image objects described seagrass presence/absence and density across the BB-LEH an accuracy assessment was undertaken. To accomplish this the classified image objects were compared to the validation dataset within a GIS to create an accuracy assessment matrix, error of omission and commission, overall accuracy assessment, and a Kappa coefficient. This is similar to the methods employed by Lathrop et al. (2006). The Kappa coefficient is a measure of agreement between two categorical datasets correcting for the random chance that categories will agree. These measures of accuracy were completed to determine how accurately seagrass vs. all other habitats were mapped, and to determine how well the maps reflected the density of seagrass habitat based on the in situ data.
Richard G. Lathrop
Grant F. Walton Center for Remote Sensing and Spatial Analysis (CRSSA)
M-F 8.30am - 4.30pm EST USA