The NOAA Hawai‘i benthic habitat mapping project began in the summer of 2000 with acquisition of suitable remotely sensed data and continued through July 2003 with the final release of benthic habitat maps. The overall mapping process was divided into four major categories (Figure 1):
The research began in the summer of 2001 with the collection of training data (Figure 1B) and continued through the final release of the mapping products in 2003 (Figure 1D). Emphasis, however, was on the accuracy assessment analyses of the various thematic maps (Figure 1C). CRAMP was actively involved in all training, ground truthing, and accuracy assessment fieldwork, using the same observer for every field station occupied. Categories A and D (Figure 1) are central to the overall fieldwork and accuracy assessment and as such are summarized. The final map production was done by one of NOAA’s private contractors, Analytical Laboratories Hawai‘i, Inc (ALH). ALH did the original photointerpretation and provided the second draft maps to CRAMP for accuracy assessment. CRAMP assisted ALH in all stages of fieldwork. ALH also revised the final map product based on the findings of the accuracy assessment analyses.
Figure 1: Flowchart of benthic habitat mapping process, including: A. Collection and processing of remotely sensed imagery, B. Draft map creation based on habitat classification scheme and field training data, C. Map accuracy assessment data collection and analyses, D. Production and release of final map product of known accuracy standards.
Four geographically distinct test areas were chosen from around the state of Hawai‘i. The size of the four test areas ranged from 16 square kilometers on Maui to 54 square kilometers on O‘ahu. They ran from the shoreline to a depth of at least 30 meters, and encompassed the various habitat types present on each island. The test areas were selected as a representative cross section of the habitats that were expected to occur throughout the remainder of the state. The first test area was located in the district of South Kohala on the west side of the island of Hawai‘i. The geographic range extends from Kawaihae Harbor south to Kiholo Bay. The second test area was located in Kāne‘ohe Bay on the island of O‘ahu. Its geographic range extends from Mokoli‘i Island south to the Sampan Channel. The third test area was located along the south shore of the island of Molokai, extending from 1 kilometer west of Pālā‘au channel to the east side of Kamiloloa. The fourth test area was in southwest Maui, extending from approximately 0.5 kilometers west of Ma‘alaea harbor southward to Mākena beach.
Remotely Sensed Data
Suitable remotely sensed imagery was acquired for roughly 60% of the coastline of the main Hawaiian Islands. Three types of remotely sensed imagery were collected or purchased for entire coverage of each of the four test areas. Although acquired for complete geographic coverage of these four test areas, some portions of the various types of imagery had cloud cover which could not be removed by adjacent coverage or repetitive acquisitions. The areas of cloud cover varied by test site and image type, but were always less than 10% of the overall test area for any image type. It should also be noted that although spatial coverage was comparable between image types for each test area, there were several large areas where some types of imagery were not suitable for mapping use due to surface glint, waves, suspended sediments and other reasons. Due to these reasons, the usable hyperspectral data were limited to a smaller region of the O‘ahu test site than the other two types of imagery. The geographic extent of the imagery at each test area always slightly exceeded the area to be mapped.
Color aerial photography and AURORO hyperspectral data were collected simultaneously by NOAA’s Aircraft Operations Center two-port aircraft, with a conventional photographic camera in one port and a hyperspectral instrument in the other. IKONOS satellite imagery was purchased for complete coverage of all four test areas. NOAA and its private contractors did all the photogrammetric processing of the data and made the resultant imagery available to CRAMP for fieldwork logistics and accuracy assessment analyses. The final imagery was also provided to Analytical Laboratories Hawai‘i for photointerpretation and map production.
Nine by nine inch color photos were taken at 1:24,000 for complete coverage of all four tests areas. Imagery collection was restricted to hours of the day when specific sun glint, weather and wave conditions permitted the best imagery for the purpose of underwater feature recognition. Along each flight line, consecutive photos were taken with 60% overlap. Where more than one flight line was required, sidelap between adjacent flight lines was approximately 30%. These methods are consistent with conventional photogrammetric requirements for processing of aerial photography (Campbell, 1998).
From the original negatives, color diapositives were created and digitally scanned in tagged image format (TIF). NOAA’s contractor used a photogrammetric quality scanner at a resolution of 500 dots per inch, resulting in TIF images with pixels that represent one square meter of coverage on the ground (or in this case underwater). These digital images were orthorectified and mosaiced using a combination of different softcopy photogrammetric software that accounted for lens distortion and incorporated airborne kinematic GPS data. The process resulted in georeferenced digital mosaics of several adjacent photos with uniform scale throughout. Histogram matching or color balancing was not attempted between adjacent photos. While this may create mosaics with noticeable color differences at individual image seams, it does not change the original photographic information that can aid the photointerpreter in delineation of habitat types.
Co-registration of adjacent imagery was preformed before mosaicing, using both image-to-image tie points (features visible in adjacent imagery) and ground control points collected in the field at features visible in the digital scans. For each ground control point occupied, at least ten minutes of carrier phase GPS data were collected. Very few ground control points were collected over water, so the image-to-image tie point method of co-registration was important for underwater and offshore features. The ground control points were collected on peninsulas, headlands and small offshore islands whenever possible to allow for the best spatial solution for the georegistration process. They were also collected at sidewalk corners, street intersections, center-lines for lanes in streets, ends of stop lines at intersections, right-angle edges of cement slabs or other manmade features, and on geodetic control monuments which could be associated with single pixels in the digital scans. A haphazard subset of the ground control points were held out from the solution process and used to measure the quality of the solution model after rectification was complete. If the spatial accuracy was unacceptable to NOAA, more GCP and tie points were included and the model was ran again and checked against the “held out” GCP’s. This process was repeated until a Root Mean Square Error (RMSE) of less than five pixels (e.g. meters) was reported at the “held out” GCP’s. Since there were very few GCP’s collected offshore, the spatial accuracy for offshore areas was expected to be less than that of the onshore areas.
The digitally scanned aerial photos were mosaiced using the best sections of
each image based on both spatial (positional) accuracy and underwater
feature recognition. Ground control points (GCP’s) collected for each island
numbered 133 for O‘ahu, 44 for Moloka‘i, 45 for Maui, and 23 for the island
of Hawai‘i. A large subset of the ground control points, along with
kinematic GPS and image-to-image tie points were applied in the geometric
solution of the photomosaics. Ten-meter digital elevation models from U.S.
Geological Survey were used to correct for radial and relief displacement.
The spatial accuracy of the mosaiced aerial photography for each island is
reported in average root mean square error (RMSE) as calculated in pixel
units (12 meter). This RMSE was calculated at a subset of the ground control
points that were held out from the geometric solution. The average RMSE for
each island test area was under two pixels (meters) in the X, Y (horizontal)
and Z (vertical) directions (Table 2). Positional error was not uniform
throughout the imagery, and was greater for areas not over land, as only
kinematic GPS data and image to image tie points were used in the offshore
areas of the geometric solutions. The RMSE increased as distance from shore
The final mosaics were saved as geotiffs and projected to the North American Datum 1983 (NAD83) in Universal Transverse Mercator (UTM) zone 4 north for all areas except the island of Hawaii, where UTM zone 5 north was used. While some areas on the island of Hawai‘i are officially designated as UTM zone 4, for the purpose of this mapping project all imagery and maps for the island of Hawai‘i were in done in zone 5.
AURORO Hyperspectral Imagery
The hyperspectral imaging data were collected at the same time as the color aerial photography, using the AURORA Hyperspectral Imaging System (HIS) from Applied Power Technologies, Inc (APTI). This instrument was set to collect seventy two narrow 10 nm bands in the visible and near infrared spectral range, with a resulting pixel size of 2.8 by 2.8 meters (based on the flying height). Real time navigation data were collected and incorporated with the spectral data using an Applanix inertial navigation system. These raw spectral and navigational data were provided to ALH for georeferencing, interpretation and mapping.
The data were spectrally processed using Research Systems Inc. ENVI remote sensing software. A subset of optimum bands from the visible spectrum were selected to reveal benthic features for habitat mapping. From this subset, several different three-band combinations were converted into Red/Green/Blue (RGB) composite images and then georeferenced and mosaiced using the real time navigational data. Where possible, ground control points collected for georeferencing the aerial photography were used to check the spatial quality of the resultant hyperspectral three-band imagery
The spatial accuracy of the AURORA hyperspectral data was much lower than either IKONOS or aerial photography. Although the hyperspectral data were collected simultaneously with the color aerial photography in same fixed wing platform, the real-time positional data collected for the hyperspectral acquisition was not as accurate as that of the photography. Attempts to post process these data using various techniques by the vendor reduced this error, but never adequately georeferenced these data to meet NOAA contract standards of horizontal RMSE at or below 2 meters. This did not prohibit comparisons of the different thematic maps, as the resultant hyperspectral imagery was photointerpreted and then an image-to-image tie point solution was applied to the accuracy assessment field data to co-register them to the hyperspectral thematic maps. The overall research objective was to compare the three types of imagery in terms of photo interpretability and resulting thematic map accuracy, which can be done to thematic maps which have limited positional accuracy if these maps can be floated to the correct or close location after image interpretation. The positional error of the hyperspectral data increased as the distance from shore increased.
IKONOS Satellite Imagery
IKONOS multispectral satellite data were acquired for complete coverage of all four test areas. These data were purchased from Space Imaging Inc, specifically for use in this mapping project. The satellite was tasked by NOAA in a special acquisition contract to collect data that would facilitate underwater feature recognition. The angle of the satellite had to be tilted from its usual orientation to allow for better sea surface penetration, and the time of acquisition had to be limited to mid-day hours to minimize glint from the ocean surface.
The spatial accuracy of the IKONOS data and resultant imagery was provided by Space Imaging Inc, and was stated to have a RMSE of 1.9 meters as per technical review. This low RMSE was based on data from the satellite utilizing ground control points to give added positional accuracy. The IKONOS imagery was orthorectified by Space Imaging, Inc. and was processed for surface glint and water column corrections by NOAA before being provided to the photointerpreter. NOAA reported the imagery was georeferenced to within 10 meters per pixel in the areas furthest from control points, where RMSE and georeferencing was suspected to be the poorest in the imagery.
The IKONOS multispectral data span four relatively wide spectral bands (65- 95 nm width), with three bands in the visible spectra and a forth in the near infrared wavelengths. The near infrared band was not used for underwater feature recognition because of lack of penetration through the water column, but was useful in delineating shoreline boundaries. The three visible bands were mapped to various combinations of the RGB guns in a cathode ray screen to generate false color imagery used for habitat delineation. The true-color band combinations were visually comparable to the other two types of remotely sensed imagery collected. Pixel size (resolution) was four meters by four meters. The final imagery was projected in WGS84 UTM zone 4 for all areas other than the island of Hawai‘i, which was UTM zone 5 north following the convention used for the aerial photography and hyperspectral imagery.
Hawai‘i Benthic Habitat Classification Scheme
A benthic habitat classification scheme was used to differentiate various coral reef habitats. The thematic digital habitat maps for all test areas were created based on this classification scheme. The classification scheme was hierarchical so as to be collapsible or expandable to different levels of detail for each class. These classes were both mutually exclusive and totally exhaustive. Two import aspects considered during the creation of this classification scheme were the (pixel) resolution of the imagery from which the classes would be delineated, and the size of the minimum mapping unit. If a feature, habitat type or potential map class could not be seen in the imagery used for interpretation, it was not included in the classification scheme. This can result from a feature being too small to see or a feature not being distinguishable from its surroundings. Quite simply, this means if the photointerpreter could not see it, then it was not included in the classification scheme nor was it mapped.
The Hawai‘i benthic habitat classification scheme has four classes in the first level of the hierarchy: Unconsolidated Sediments, Submerged Aquatic Vegetation, Coral Reef and Hard Bottom, and Other Delineations (Table 3). There are eleven second-level classes and 22 third-level classes. Not all of the second-level classes have third-level classes beneath them in the hierarchy. There are a total of 28 classes at the greatest level of detail.
A second component of the classification scheme includes zones based on insular shelf and coral reef geomorphology. Zones refer to benthic community locations, while habitats refer only to substrate cover and type. While habitats can occur across all reef zones, the zones are mutually exclusive of each other. Although the zones were determined during fieldwork, they were not part of the accuracy assessment analyses and are only mentioned summarily.
Benthic Habitat Map Preparation
First Draft Benthic Habitat Maps
The mapping contractor (ALH) produced first-draft benthic habitat maps for each type of imagery at each of the four test areas. Three different maps were made at each of the four test areas; one from each of the three different types of remotely sensed imagery (e.g. Figure 3). In total, there were twelve individual maps. These maps were created by “heads-up” computer screen photointerpretation of the various types of imagery. Delineation of habitat boundaries were digitized using a mouse and a custom digitizing extension that operates within Environmental Systems Research Institute, Inc. (ESRI) ArcView software. NOAA staff developed this editable ArcView “Habitat Digitizer” extension, which allows for a custom habitat classification scheme to populate the attribute table of the polygons (classes) delineated. Traditional methods of “grease pencil” delineation of photointerpreted habitat classes have almost completely been replaced by these computerized “heads up” digitizing methods. The resulting thematic maps are digital ESRI shape files which can be used in various GIS software packages. For each set of maps, benthic features were classified using the NOAA hierarchical coral reef habitat classification scheme (Table 3).
All delineations of habitat boundaries were conducted by ALH with the ArcView image scale set at 1:6,000. This was a contractual requirement from NOAA and ensured that the level of detail produced by the photointerpreter was uniform between different types of remotely sensed imagery throughout the project. NOAA has demonstrated from similar mapping efforts in Florida, the Caribbean, and Gulf of Mexico that little additional information was gained from interpreting similar imagery at a larger scale while the labor intensity for doing so was increased significantly (pers. com.).
The minimum mapping unit (MMU), or smallest area delineated in any of the maps produced, was restricted to one acre. This was also a contractual requirement from NOAA, based in part on the large geographic range of the overall mapping project. A smaller MMU would require a substantial increase in photointerpretation man hours, and therefore would have been both time and cost prohibitive for the Hawai‘i mapping project. The custom ArcView digitizing extension allows the photointerpreter to set a pre-defined MMU area. The extension informs the photointerpreter if a delineated polygon has an area below this pre-defined MMU, and provides the option of either including or eliminating that polygon in the final map. The set working scale and pre-defined MMU standards ensure a uniform mapping product.
No field visits or ground truthing was done before creating the first draft maps. Their thematic polygons were delineated based on what different habitats or features could be seen in the imagery. The photointerpreter delineated polygons based on boundaries that were visible in the imagery without necessarily knowing to which habitat class these polygons belonged. Sometimes the interpreter could tell which classes were being delineated, but other times the classes could not be determined until the areas and features were visited in the field. Although the boundaries of these map’s polygons did not change dramatically throughout the length of the project, the benthic habitat type attributed to these polygons changed as ground validation and training data were collected and applied.
Of all three types of remotely sensed data, the aerial photography had the largest area of usable imagery at three of the four test areas. For this reason, first draft maps from color aerial photography were used as the basis for the ground validation and training data collection. Unidentified features which were visible in the hyperspectral or IKONOS imagery but not visible in the color imagery also had GPS waypoints created for ground validation and training data collection.
Second Draft Habitat Maps
The first draft maps were subsequently revised based on the training and validation field data, resulting in “second draft” habitat maps. When required, the “first draft” habitat polygons were reattributed to the correct map class. Sometimes new polygon boundaries were drawn based on information collected in the field.
draft habitat maps were created for each of the three types of remotely
sensed data collected at each of the four test areas, resulting in twelve
separate maps. Imagery types differed in aerial extent within any test area
due to differences in cloud cover, glint, waves, turbidity, missing data and
other factors preventing substrate feature recognition. The total area
mapped using each type of imagery was comparable within any test site, with
the exception of the Kāne‘ohe Bay, O‘ahu hyperspectral imagery map (Table
4). The Kāne‘ohe Bay map made from hyperspectral imagery covered
approximately 32 Km, or 60% of the area covered by either of the other two
maps for this test site. The number of polygons delineated for each imagery
type was comparable within the four test areas. This is more evident when
normalized for total area mapped (Table 4: Polygons per sq. Km). When
complete, NOAA and ALH supplied these second draft thematic maps to CRAMP in
ESRI polygon shapefile format for fieldwork in support of map accuracy
Last Update: 04/21/2008
By: Lea Hollingsworth
Hawai‘i Coral Reef Assessment & Monitoring Program
Hawai‘i Institute of Marine Biology
P.O. Box 1346
Kāne‘ohe, HI 96744