Status of GAP Components This section reviews the status of the constituent parts, or components needed, to conduct Gap Analysis. Gap Analysis was begun with a focus on the terrestrial environment, however, the development of information and analyses must logically be extended to the aquatic environment; the aquatic component of Gap Analysis is also treated in this section. The status of analyses of the GAP data layers is discussed in the "Products" section of this report. Land Cover Maps Mapping natural land cover requires a higher level of effort than the development of data for vertebrate species, agency ownership, or land management, yet this does not make it any more important than the other data layers. Prior to GAP, only one or two states had contemporary maps of their natural land cover, digital or otherwise. The science and technology of large area land cover mapping have not developed commensurately with computing power and other technologies. Because of this, a variety of different methods have been tried by different state projects. Generally, the mapping of land cover is done by adopting or developing a land cover classification system, delineating areas of relative homogeneity (basic cartographic "objects"), then labeling these areas using categories defined by the classification system. More detailed attributes of the individual areas are added as more information becomes available, and a process of validating both polygon pattern and labels is applied for editing and revising the map. This is done in an iterative fashion, with the results from one step causing re-evaluation of results from another step. For example, the discovery of attributes for a given polygon may result in adjustment of its boundary. Finally, an assessment of the overall accuracy of the data is conducted. Where the database is appropriately maintained, the final assessment of accuracy will show where improvements should be made in the next update (Stoms et al. 1994). Some of the major barriers to efficient mapping of large areas at the required spatial and thematic resolutions that have been overcome are: (1) classification of land cover, (2) data acquisition, (3) delineation of land cover pattern, (4) object interpretation, and (5) assessment of final map accuracy. Each of these is discussed below.
In order to provide meaningful comparisons across large areas, a consistent land cover classification system is prerequisite. Land cover classifications must rely on specified attributes such as the structural features of plants, their floristic composition, or environmental conditions to differentiate categories evenly (Khchler and Zonneveld 1988). Although there has been much effort devoted to the classification of vegetation, there has been no previous attempt to apply a detailed classification of natural land cover across the contiguous 48 U.S. states at a 1:100,000 scale. In choosing a classification system for GAP, the following basic criteria are required:
For Gap Analysis, the system that fits best is provisionally referred to as the Natural Land Cover Classification System (NLC). In recent times, this system has also been referred to as the UNESCO/TNC system (Lins and Kleckner in press) because it is based on the structural characteristics of vegetation derived by Mueller-Dombois and Ellenberg (1974), adopted by the United Nations Educational, Scientific, and Cultural Organization (UNESCO 1973), and later modified for application to the United States by Driscoll et al. (1983, 1984). The Nature Conservancy and the Natural Heritage Network (Grossman et al. 1994) have been improving upon this system in recent years with partial funding supplied by GAP. As the GAP prototype state projects of Idaho and Oregon were completed and projects were started in most of the other Western states, the need for a consistent set of vegetation cover types was recognized. Funding provided by GAP to The Nature Conservancy was used to compile and standardize the names and descriptions for vegetation types of 11 Western states (Bourgeron and Engelking 1994) and 13 Northeastern states (Sneddon et al. 1994) from data developed by each state's Natural Heritage Program. Underlying assumptions and definitions for the classification system were described by Jennings (1993b). The Gap Analysis Program, with additional support from the Department of Defense in 1994, continues to fund the standardization of regional classifications of Natural Community Alliances in the Southeastern and Midwestern U.S. The development of this classification system is also supported by funds from the NBS/NPS Vegetation Mapping Program, an effort to map the land cover of U.S. national parks. Eventually all regional classifications will be standardized into a single national classification. In its present form, the classification continues to be improved upon by The Nature Conservancy and the Natural Heritage Network (Grossman et al. 1994). In addition to improving the physiognomic part greatly, TNC and the Natural Heritage Programs are responsible for creating a consistent set of labels and attributes for vegetation types of the United States based on the science of vegetation and community ecology. An additional development for the classification system is that during the Ecological Society of America's (ESA, a non-profit professional society) 1994 annual conference, a subcommittee on vegetation classification was formed. A standing ESA Panel on Vegetation Classification has since been established by the ESA Executive Council to:
Further, during the spring of 1995, the Federal Geographic Data Committee's Vegetation Subcommittee moved to adopt the NLC as a federal standard, though a final decision had not been issued as of this writing. The development of a broadly accepted classification system that is maintained within a scientific peer-reviewed arena and recognized by government agencies has historic significance in the natural resources fields. As discussed in the overview on state GAP projects, changes in the natural resources fields are resulting in demand for an "ecosystems" approach in research, planning, and management. Yet nowhere else is there a consistent set of defined categories for naturally occurring assemblages of species (see Orians 1993) that can be reliably used as the building blocks for characterizing ecosystems at alpha, beta, delta, and gamma scales of diversity (sensu Whittaker 1960, 1977). Ecosystems are any unit that includes the assemblage of organisms in a given area that interact with the physical environment so that a flow of energy leads to: a) clearly defined trophic structures, b) biotic diversity, and c) material cycles (Odum 1971). Gap Analysis is not only funding a good part of the classification system's development and contributing centrally to activities of the ESA Panel and the FGDC Vegetation Subcommittee, it is mapping these basic units across the U.S. And, it is developing reliable models for relating these units to the landscape-level occurrences of each native vertebrate species in the U.S., all in a digital Geographic Information System format. The information, its structure, and the institutional capabilities now forming are having a profound effect within and among state natural resource agencies, federal agencies, and private organizations. This development is requisite and fundamental for the rational planning and management of ecological systems as we now understand them (Jennings 1989, Jennings and Reganold 1992).
Although Gap Analysis projects use a wide variety of data sources, the two major types of land cover data required are Landsat Thematic Mapper (TM) satellite images and ground point data. These are also the most expensive. The TM images are used to render a digital base map and to delineate land cover patterns from spectral reflection. The ground point data are used to interpret and label the objects delineated from land cover patterns. In 1994 and 1995, substantial progress was made in both access and cost-savings for acquisition of these data. Sections a and b below discuss these achievements.
Obtaining the TM images had been a major inhibiting factor in launching new GAP projects. Problems with the acquisition of the TM data have been: (1) cost, (2) timeliness of delivery, (3) preprocessing and processing of the images, and (4) copyright restrictions on sharing the images with cooperating agencies. In the early stages of GAP, state projects were responsible for acquiring their own TM imagery. These images are privately owned and cost from $4,400 for a basic "raw" image (no corrections or "preprocessing" steps, such as corrections for geometric distortion, atmospheric haze, or system errors) to $5,500 for a fully preprocessed image. Costs therefore varied according to the size of each state, with imagery for Arizona costing about $121,000, while imagery for Montana cost about $187,000, and imagery for Arkansas cost about $55,000. Actual costs have also varied because of constraints and opportunities particular to each state project. Because TM costs constituted such a large portion of each state project's budget, many of the GAP state projects embarked on organizing a cooperative purchase among different state interests. This required a large investment of time and effort in gaining agreement and in negotiating cost sharing and details among agencies. There were frequently long delays in delivery after the state project staff managed to: (a) assemble a cooperative purchase agreement among state partners (who pays and how?), (b) sort through the available imagery and select individual images agreed to by all involved (no small task), then, (c) actually place an order for the imagery. Additionally, delivery of the wrong image, or delivery of images having unacceptable system errors, or other problems were common experiences. Images frequently had to be returned and further delays endured. For example, the New York State Gap Analysis Project experienced a delay of 18 months because of these problems. With each state Gap Analysis project pursuing its own TM imagery, there were natural variations among the many choices for preprocessing and processing of these data. While these variations may not be critical, a greater degree of standardization would facilitate and enhance the ultimate edge-matching of map products across state boundaries. Finally, the variations in the arrangement of image acquisition between state Gap Analysis projects resulted in substantially different interpretations about data access under copyright law among project staff, cooperators, and non-affiliated organizations wishing to access the imagery. This carried the potential of some legal risk to the GAP projects and possibly the Department of the Interior. The national GAP staff solved each of these problems by forming a consortium with other federal programs. Known as the Multi-Resolution Land Characteristics Consortium (MRLC), the group selected, assembled, and acquired the first full set of Landsat TM images for the 48 contiguous states. In FY 1994, GAP invested just over one-quarter of its budget ($1M out of $3.61M) in this consortium. This resulted in:
The MRLC endeavor also solved the problem for GAP of how best to archive and distribute the TM images as well as archive all final GAP data. Through the MRLC, the USGS EROS Data Center will provide these services:
The most important achievement of the MRLC is the establishment of a common set of digital land cover data across multiple levels of spatial and thematic resolution. In September 1995, the MRLC developed a specific strategy for integrating the land cover data from all member programs into a single national land cover database. In May 1995, the MRLC was recognized by the White House Closing the Circle Award for "Environmental Excellence in Government." The overall MRLC model is shown in Figure 2.
Collection of ground point data has also been problematic for Gap Analysis projects because of cost and access. The development of airborne video equipment integrated with Global Positioning System (GPS) data has had a major yield for the state projects in solving this problem. The application of airborne video for GAP land cover mapping was begun by Graham (1993), principal investigator of the Arizona Gap Analysis Project. A total of 11,223 ground point samples covering about 3 percent of the state's surface area was collected inexpensively. Each of the known vegetation types was visited on the ground, and a simple key was developed to relate ground surveys to the airborne video images. Approximately two-thirds of the airborne video frames (7,482) were used to interpret the TM and map 142 different vegetation types. One-third of the airborne video frames (3,741) were used to assess the accuracy of the draft map. Since then the application of airborne video to object interpretation has been improved even further (Slaymaker et al. in press), and national GAP continues to facilitate acquisition of these data by the state projects. Four airborne video units (cameras, recorders, GPS receivers, mounting brackets, etc.) have been acquired and are shared among the state projects, and a number of training workshops have been held. Usually the state project cooperators furnish the airplane, so for many GAP projects there is little direct cash outlay. High resolution airborne video transect samples have been acquired for Arizona, Connecticut, Maine, Massachusetts, part of Montana, New Hampshire, Oklahoma, Oregon, Rhode Island, Tennessee, Texas, and Vermont. Plans are under way to acquire airborne video imagery in Delaware, Florida, Indiana, Maryland, Michigan, Minnesota, New Jersey, Pennsylvania, West Virginia, Wisconsin, and Wyoming. While the airborne video technology has solved a major problem and is a cost-effective intermediate platform from which to obtain high-resolution images, substantial on-the-ground reconnaissance has been, and continues to be, the primary source for thematic interpretation and assessment of map accuracy (i.e., Edwards et al. 1995). Interpreted imagery is only as reliable as the field reconnaissance upon which it is based.
No single procedure is appropriate for the delineation of land cover patterns in all environments (Stoms 1994a). One of the different approaches being used renders land cover patterns (dominant vegetation types, water, or non-vegetated areas) by visual interpretation of false color digital TM scenes. This procedure uses "on screen" delineation of different dominant vegetation types based on criteria such as color, texture, context, and cross-reference to other information sources such as air photos and other maps as interpreted by an analyst (Davis et al. 1991, Scott et al. 1993). Another procedure delineates vegetation patterns by clustering spectral data using unsupervised or supervised methods (Colwell 1983). Many of the GAP geographers begin with a digital product from one of the procedures, then continue to refine the delineated pattern by incorporating information from the other procedure. Further improvements are often made by extracting additional detail from TM using hybrid classification (Lillesand and Kiefer 1987, Lillesand in press) along with air and ground reconnaissance. Although concern has been expressed by some that a single mapping procedure is somehow a requisite for adequacy, the use of these two procedures, their hybrids, and the further development of other methods continues because:
Rather than limit information to a continent-level least common denominator, there are five ways in which differences that could arise in the derived data among state projects are dealt with. First, a standardized method for assessing the accuracy of all land cover maps, regardless of the different sources of information or different procedures used to produce them, is being developed. Second, a full accounting of the variability among state-level products is provided to users through standardized metadata records. Third, a detailed review of data quality is undertaken when edge-matching data from adjacent states to ensure spatial and thematic consistency. Fourth, since the present effort is a first generation, these data-sets will be improved upon over time, dampening the amplitude of inter-state variation. Fifth, results from applying the two pattern delineation procedures ("on-screen digitizing" and supervised/unsupervised spectral classification) to the same area were compared. Although some differences are evident below the GAP minimum mapping unit of 100 hectares, these initial results are judged to be within acceptable limits for the purposes of Gap Analysis (Zhenkui Ma and Roland Redmond, Montana Gap Analysis Project, personal communication). Research on this topic is ongoing with newly completed state project land cover data, especially in edge-matching these data with those of adjacent states.
There have been five closely related new developments in the areas of both pattern delineation and object interpretation: (1) merging of two TM images of the same area but from different seasons into a single digital image, using phenotypic distinction to enhance detection of pattern, (2) production of a spectrally "hyperclustered" TM image product, more than doubling the number of spectral categories to work with in each image, (3) use of airborne video for interpreting and labeling the hyperclustered TM image, (4) development of the SkyKing software program that relates individual airborne video frames to the same location on a TM image, and (5) application of the SPECTRUM software program to the hyperclustered images, integrating automated pattern delineation with visual interpretation. The multi-seasonal and the hyperclustered products are discussed immediately below.
The delineation of land cover patterns from TM is easiest when the objects (i.e., dominant vegetation types) contrast strongly with each other. Pattern delineation in the Eastern deciduous forest and in grasslands are particularly problematic because during a large part of the year the spectral contrast among different vegetation types is low. One way to overcome this problem is to merge more than one TM image, of the same area but from different seasons, together into a single spectrally classified image. This captures some of the natural phenology of plant species, providing for much better definition of dominant vegetation types. Simultaneous spectral classification of more than one TM image from the same area ("multi-temporal") is conceptually and practically an easy solution to many problems of large-area land cover mapping (Dobson et al. 1995). Obtaining and processing multi-temporal TM images, however, had not previously been possible because of cost and other problems listed above in the section on data acquisition. When the MRLC TM acquisition was being planned, the acquisition of multi-temporal TM images for the Eastern deciduous forest was included as an objective. Multi-temporal images were acquired for 100 of the 430 TM scene areas which cover the 48 states (Bara 1994). The EROS Data Center now produces these multi-temporal images as a product referred to as a "twelve-band" spectrally clustered (or classified) image. In this process, the digital brightness values of the pixels in six separate spectral bands, from each of the two TM images, are mathematically combined. The final product is a single spectrally classified image that provides greater discrepancy of land cover types.
In addition to the "twelve-band" product, the application of a form of unsupervised spectral clustering to the TM imagery that generates a large number (240) of spectral clusters (compared with 80-100 as a usual range) is another significant development (Kelly and White 1994). This provides for a denser pattern of spectrally discrete land cover types and works well with the multi-seasonal imagery. The product is commonly referred to as a "hyperclustered" image. When multi-seasonal imagery is hyperclustered, the product is referred to as a "twelve-band hypercluster." These developments, along with the use of airborne video, have made the comprehensive mapping (at the spatial resolution of 30 m and the thematic resolution of natural community alliance) of the Eastern deciduous forest possible for the first time.
The single most significant advancement in object interpretation for GAP in 1994 and 1995 has been the use of airborne video for interpreting and labeling the twelve-band hyperclustered TM image (Slaymaker et al. in press). However, the development of the SkyKing and the SPECTRUM software programs is important, and their potential is enormous. Each of these is discussed below.
Gap Analysis for the New England states began in 1992 with a single effort at the University of Massachusetts to develop the land cover mapping techniques. This effort resulted in a land cover map of the region consisting only of general categories such as "mixed deciduous forest," "conifer forest," or "grassland." The land cover of New England is commonly over 70 percent forested. The forests have a complex plant species composition, spatial configuration, and spectral properties, making for poor discrimination of forest types from a single TM image. These properties are characteristic of the forest canopy that stretches across the Eastern U.S. from Maine to the Gulf of Mexico coast. It has often been characterized as a "green rug" by the remote sensing community, because the well-mixed tree species cause its texture and color to appear fine-grained and uniform, making it difficult to distinguish the dominant vegetation types. There are no current maps of natural vegetation characterized by dominant species or habitat types that cover large areas of this ecosystem. In 1994, an intensive effort was focused on a single TM scene area (path 13, row 31) in the upper Connecticut River watershed (Slaymaker et al. in press). Here, the new twelve-band hyperclustered digital image product was interpreted using aerial videography. Two Hi8-band video cameras were mounted side-by-side. One was set at wide-angle, and the other at 12X zoom. This resulted in a wide-angle image swath of one-half km with a 30-meter-wide swath down the middle of it, providing a great deal of flexibility in the selection and the density of sample points. Since each video frame was GPS-registered, selected frames could be printed as a photo, then located on the ground for field-checking. These were used to create a visual key to the forest types, and each photo-interpreter was trained to consistently identify each type. Display monitors for both video images and the hyperclustered TM image were arranged side-by-side, and the video frames were photo-interpreted on-screen. This combination produced a major breakthrough in the ability to map deciduous forest vegetation types across large regions with an acceptable degree of accuracy and discrimination (Slaymaker et al. in press).
SkyKing was developed in the Mapping Sciences Laboratory at Texas A&M University specifically for the Texas GAP project. It is designed to provide a rapid method for linking high resolution video image interpretations with the unsupervised classifications from Landsat imagery (Wunneburger et al. 1990, Wunneburger et al. 1993). By integrating aircraft attitude measurement, provided by two or three separate GPS receivers, corner points of individual video frames are registered to a Universal Transverse Mercator grid. By knowing these corner points (+/ - 30 meters), SkyKing can interpolate the UTM position of any TM pixel on the video image. Using this method, from the computer screen, the analyst identifies an Alliance land cover type, or mix of types, within a given video frame. This estimate is guided by visual image training on Alliances within the hierarchical vegetative cover classification scheme, from which the interpreter selects an entry. By generating a list of coordinates with corresponding classification values, a seed procedure for identifying and validating unsupervised classifications of Landsat scenes is available for export from SkyKing to image analysis software. Ongoing software development will allow an ArcView or VistaMap front-end, whereby an area map of video frame locations may be displayed, then a window of the video frames (or an individual scene) can be selected and opened for interpretation. For more information on SkyKing, contact Dr. Douglas Wunneburger, Research Scientist, Mapping Sciences Laboratory, Texas A&M University, College Station, TX 77843-2120, voice: (409) 862-4253, e-mail: douglas@rsgis3.tamu.edu.
Initially developed at the Los Alamos National Laboratory for military use, SPECTRUM allows a user to quickly analyze and categorize the TM-derived spectral clusters in a menu-driven computing environment. A map legend showing the number of categories, their labels and assigned colors can be easily designed. A SPECTRUM user can manipulate the six- or twelve-band hyperclustered image (or other transformations), and a window can be opened to display the positions of multiple clusters from within a selected area. These cluster positions are displayed on a two-dimensional scatter plot of spectral values along axes specified by the user. Or, the spectral values associated with any given pixel can likewise be displayed. The SPECTRUM software is a part of Khoros, an overall data visualization system. It is owned by KHORAL Research, Inc. and is available to users under a free access license. Khoros runs in an X Windows environment. An analyst can use these features in a variety of ways. Image data can be categorized by choosing an area of a known vegetation type and assigning all associated pixels in the image the same color code. For example, a user could locate a natural community with the upper vegetation canopy dominated by a beachgrass (e.g., Ammophila arenaria). The pixels within the beachgrass area can then be highlighted, and all other beachgrass areas in the entire image (about 31,000 sq. km) would be simultaneously highlighted. Using the scatter plot window, clusters can be deselected and reselected directly. Different two-dimensional axes for scatter diagrams can be displayed, allowing the analyst to interpret and categorize the data by looking at different mathematical transformations of the cluster positions, while results of the process are immediately displayed in the image. Once a group of clusters is identified as one or more vegetation types, they can be quickly labeled and merged across the entire image (Kelly and White 1994). The potential of SPECTRUM for the various land cover mapping efforts being done by the MRLC group was first identified by Gail Thelin of the USGS NAWQA program. Prototype testing, application, and refinement were carried out by Susan Benjamin of the USGS EROS Data Center. Thelin and Benjamin developed materials and have conducted training workshops on its use for the USGS NAWQA program. For GAP, there are still some problems to be overcome with the access to, use of, and support for SPECTRUM. These problems exist primarily because the development of SPECTRUM for widespread use is not a significant priority of any one organization. SPECTRUM was acquired by the Texas GAP project and made usable only with difficulty. In addition to the Texas project, it has been applied in the Maryland/Delaware/New Jersey, New York, and Pennsylvania GAP projects. It is expected, though, that SPECTRUM will become a major tool for GAP projects, especially when integrated with airborne videography. The degree to which GAP can utilize the SPECTRUM-related tools and products paid for by NAWQA remains unknown.
The need for a standard methodology for assessing the accuracy of GAP land cover data has long been recognized. One of the recommendations made by the GAP peer review panel was that a protocol needed to be developed to provide estimates for the accuracy of GAP data. A variety of methodologies have been described (e.g., Congalton 1991). The central problem is that when working with large areas, say the size of a single TM scene (~31,000 sq. km) or larger (many states encompass 10-20 TM scene areas), the size of the sample needed to statistically represent the accuracy of a land cover map can cost as much or more in time and resources as the primary map data (Dobson et al. 1995). Nonetheless, a method for assessing the accuracy of the GAP land cover maps is critical, and work on this issue began in earnest in February 1994 with a workshop to outline the major issues of accuracy assessment and to develop a recommended protocol. Twenty-three participants attended the workshop, representing a wide variety of organizations and having expertise in the fields of land cover mapping, spatial statistics, and remote sensing. The workshop resulted in a report (Stoms et al. 1994) that frames the technical issues and discusses methods that are suited for GAP (Stoms et al. 1994). It presents a set of guidelines intended to establish the minimum acceptable level of accuracy assessment for the state projects and describes a procedure that, with testing and refinement, could provide the basis for a standard protocol (Stoms et al. 1994). The procedure is a first approximation, and the report's authors point out that, "There are several reasons that it is not practical to present a national protocol at this time:
The basic criteria for assessment of map accuracy that were identified are that it must be scientifically sound, economically feasible, nationally applicable, and coordinated and consistent with other federal efforts. The assessment data will be used to improve GAP land cover maps iteratively by identifying where the maps are inaccurate and by how much, thus indicating where more effort may be invested to improve them. Also, the field sampling may identify new land cover types which were not previously identified or described. The assessment data will be valuable to other agencies and organizations for use in their own land cover characterization activities. Accuracy assessment for land cover maps covering large areas is an emerging science, and GAP will continue its development efforts within fiscal constraints. |
