Abstract:

 Recently, the USGS Cooperative Park Studies Unit, located at The University of Arizona, conducted an accuracy assessment of the Arizona GAP vegetation map. Numerous GPS and GIS techniques were utilized to facilitate mission planning, for data gathering, and were vital in the accuracy assessment process. Integrated field-based GPS and GIS techniques were heavily depended upon to rapidly collect "ground-truthed" vegetation data from a randomly selected, stratified, set of target polygons that statistically would represented each vegetation community class. GPS-based field mapping and polygon monitoring techniques included the use of GIS coverages to monitor which polygons were sampled and targeted for survey. Vegetation survey methods included GPS-synchronized airborne videography as well as ground-based automated GPS-GIS data collection surveys. Without the use of integrated GPS and GIS technologies the 58,169 polygon, 105 community type map could not have been adequately assessed within the 12 month study period. This paper will discuss GPS and GIS field techniques, vegetation classification systems, and preliminary thematic accuracy results of vegetation-based land cover data.

Introduction:

The worldwide loss of biodiversity is a major concern for natural resource managers and conservation biologist. Currently there is a large academic debate between conservation professionals about the number of species being lost or at risk of extinction. Since the Endangered Species Act was passed in 1973, preliminary evidence in North America indicates that the number of species at risk has noticeably and "dramatically" increased (National Gap Analysis Program, 1994). With awareness of increased species at risk, ecologist in the U.S. Department of Interior began to develop scientifically based methods to examine broad-spectrum habitat loss and subsequently initiated the National Gap Analysis Program or GAP (Scott et. al. 1987, Scott et al. 1993). The GAP program assumes that the cost of maintaining species in their natural habitat, while they are relatively common and part of self-sustaining ecosystems, is less than the cost of intensive management programs to save species near extinction. Vegetation mapping is of critical importance to the GAP program because dominant vegetation types are proximate indicators for overall biological diversity (Franklin 1993). Numerous studies have shown that vegetation structure and composition significantly affects species richness and competitive interactions between individuals and species. Dominant vegetation types are also used because patterns of natural land cover are an integrated reflection of the abiotic and biotic factors that shape the environment of a given area (Whittaker 1975). As such, dominant vegetation types are the constituent parts of landscapes that are mappable and can be used to model habitat types frequently used in conservation evaluations.

The Arizona GAP program began early in 1991 and was housed within the Advanced Resources Technology Program (ART) in the School of Renewable Natural Resource at the University of Arizona. The initial program was directed by Dr. Lee Graham from April 2, 1991 to June 30, 1995. Dr. Graham utilized airborne video techniques to assist with satellite imagery classification and in the identification and description of dominate vegetation types found in polygons. In April, 1996, the USGS Cooperative Park Studies Unit at the University of Arizona received funding by the National Gap Program to: (1) assess the overall Arizona GAP mapping effort; (2) to review the adequacy of ancillary Arizona GAP information including, but not limited to metadata; (3) assess the accuracy of the vegetation/land cover, vertebrate, and ownership Geographical Information System (GIS) coverages; (4) to update and convert the Graham land-cover types to the latest National Vegetation Classification Standard (NVCS) as envisioned by the GAP program. Because the assessment process for the whole state has not been entirely complete, the thematic accuracy results reported in this paper are accuracy estimates. However, because of the amount of data, the estimates reported should be a good proximate indicator of overall thematic accuracy and trends.

The original Dr. Graham vegetation map (figure 1) was created using common imagery classification procedures and a wide variety of GIS raster and vector processing techniques. Essentially, landsat Thematic Mapper imagery was "digitally classified using a hybrid unsupervised and supervised classification methodology. First, 3 input bands (NDVI, a 5/4 band ratio indicating moisture content of vegetation, and a local texture band built from the NDVI) were used in an unsupervised maximum likelihood classification procedure. The result of the unsupervised classification was then used with digital elevation model (DEM) data as input for a supervised classification procedure, in which buffered GPS-referenced airborne video sample points (GRAHAM_AIRVIDEO.GIF) indicating vegetation association were used as training sets (Graham, 1993). The resulting image was edited manually to correct classification errors then converted to ArcInfo vector format. A series of ARC eliminate and dissolve operations were used to get the map to the GAP Program-mandated minimum mapping unit of 100 HA for upland vegetation types, and 40 HA for riparian vegetation types" (Graham, 1995).

The final classification of Landsat Thematic Mapper data (TM) eventually provided the framework for a basic data set containing 58,169 polygons representing 105 vegetation and land cover types. Approximately 9,000 miles of airborne video transects (figure 2) were taped within the state for use by investigators. Review of airborne video transect tapes by video interpreters facilitated in the attribution and labeling of polygons with descriptive vegetation identification information. Polygon vegetation types that could not be visually interpreted or clearly attributed from the airborne video transects were selected for subsequent field investigations. Field data and experience was required to better delineate TM training sites and to assist personnel in obtaining the additional skills and knowledge necessary for subsequent video interpretation of problematic vegetation types. Field specialists, carrying GPS receivers and downloaded photographic video images of selected sites, were sent into the field to locate imagery training sites and to identify questionable species and vegetation community types. Through a combination of vegetation community identification techniques, species identifications in the field, extrapolations from imagery training sites, digital elevation models, and the use of historic vegetation data, the first Arizona GAP vegetation/land cover coverage was constructed with 105 vegetation land cover types.

The classification system developed by Graham to depict vegetation cover types, although roughly modeled on Brown, Lowe, and Pase (1979), was asystematic because it was built of similar vegetation units as determined by indirect satellite imagery classification procedures as opposed to other field methods that measure vegetation directly utilizing traditional floristically-based techniques. It should be pointed out that GAP vegetation maps are primarily based on algorithmic clustering of TM pixels with the resultant land cover classes being spectrally similar. Ideally, spectrally similar polygons would highly correlate with their respective vegetation type. Unfortunately, for a variety of reasons, including, but not limited to, vegetation community heterogeneity (the degree of patchiness and size of inclusion patches) and the ability of remote sensing sensors to discriminate between vegetative and non-vegetative reflectance patterns, spectral classes may not represent all vegetation types equally well. TM perceived map units or vegetation polygons and concomitant shapes and sizes do not produce a direct vegetation surface in the same way that low altitude, human interpreted, air photos have in the past. However, for the purposes of GAP the sampling unit to be tested is the TM derived polygon.

Maps depicting biotic communities are based primarily on natural vegetation. Although animal constituents are an important factor in the determination and classification of biomes, it is the natural vegetation structure and floristic composition that provides the taxonomic basis of separating one evolutionary-derived entity from others of the same rank. A comparative evolutionary approach to plant communities is consistent with an open-ended hierarchical system of classification that adapts to the accumulation of more information over time and the evolving nature of ecosystems. Although there are numerous vegetation classification systems in use world-wide, no system has been universally accepted. For example, GAP mapping standards require that vegetation communities be classified according to the NVCS. Unfortunately, this classification system is incomplete at this time. The Nature Conservancy and U.S. National Park Service (TNC-NPS) system is more highly developed but is also incomplete and not well described. The Brown, Lowe and Pase system (Brown et al. 1980, Brown 1982) is better described and has the additional advantage of being widely used in Arizona by resource managing agencies. Because more BLP vegetation communities have been defined and documented, the BLP system served as the classification and thematic assessment standard. The BLP classification is not parallel to NVCS nor TNC�NPS although they appear to converge somewhat at the lower classification levels. The BLP system (BLP) is a framework for organizing vegetation community data. The system may be thought of as an arrangement of expanding hierarchical units waiting to store additional evolving community-based information. BLP was never intended to be a non-adaptive, rigid classification system with a fixed number of "types" and formal descriptions (Brown, et al. 1979, 1980)

The BLP system is grounded on the biome concept and the biome serves as its fundamental organizational unit and provides an ecological basis (Brown, 1982). Biomes are natural communities characterized by a distinctive vegetation physiognomy (appearance and spacing of the plants independent of their species), evolutionary history within a formation (forest, grassland, swamp, etc.), developing within the same climatic zone, and persisting together through time and space to become a geoflora. Thus they tend to be centered in, but not restricted to particular biogeographic provinces which are contiguous, non-disjunctive, geographic regions. At the higher levels of the published BLP system, the published classifications are fairly well fixed based on our field observations. This stability extends down to the BLP series level which currently is at a finer classification level than the NVCS and more coarse than the TNC alliance level. Below the series level (BLP associations and subassociations) the classification is incomplete. The lower levels of BLP that we use in the Arizona Gap program are based on published field data. The Biomes may be and often are discontinuous (disjunctive) and are named for the biogeographic province where they are centered. The naming of biogeographic provinces (a geographic concept) and biomes (a biological concept) may create confusion although the two are different. All vegetation classifications are, to a degree, artificial abstractions created by individuals to promote information organization and entity understanding.

Because biomes evolved from common ancestors in a common environment, they are a naturally defined vegetation unit. Larger (more generalized) and smaller (more specific) classification units become progressively more artificial. Thus, classifications bench marked on the biome have a greater degree of reality than those based on species presence or on climate zones, for example. Whether a natural vegetation classification is possible or not is still a contested issue. The European ecological tradition favors the existence of natural, discrete, and recurring plant assemblages across the landscape, i.e., certain plant species are always together and thus define the community. The American ecological tradition favors the view that each plant and animal species is distributed according to its biological requirements, the past history of the place where they are rooted, and the chance presence of their propagule. Thus plant communities are variable in composition and not discretely distributed. The differences between these polarities mostly result from the data used by each tradition. The BLP system takes a somewhat intermediate position in that the biome is recognized as a natural plant assemblage. BLP as delineated here recognizes 8 levels of organization. They are (in order from broad to narrowly restricted): Biogeographical realm, Upland/Wetland, Formation type, Climatic zone, Biome, Series, Association, and Subassociation.

Accuracy assessment requires obtaining data of higher quality than the original map with adequate coverage of space and classes. The task is made more difficult because of the numerous types of accuracy factors that must be examined, in varying degrees, to accurately assess a map. Examples of factors include, but are not limited to: (1) the spatial accuracy of the map units chosen to represent landscape units at various scales and resolutions; (2) thematic or attribute accuracy; (3) topological accuracy or the fidelity and structure of digital spatial data to represent, in this case, vegetation polygons that should ideally exist without processing artifacts such as the presence of "sliver" polygons or duplicate lines that may lead to erroneous map errors; and (4) temporal accuracy of the spatial data to represent dynamic landscape processes using "snapshot in time" map units.

To determine the thematic adequacy and accuracy of individual polygon land-cover map labels (which are based on imagery) the CPSU/UA collected over 4,173 vegetation 2-4 hectare relevés complete with GPS-based positions and ancillary field plot information over a six month period (figure 3). The quantity of relevés collected would not have been possible without the uses of integrated GPS and GIS Geolink mapping software. During the assessment three types of relevés were collected in the field (figure 6): (1) traditional i.e. by foot using paper field forms as specified in the vegetation delineation manuals (Bennett, et al., 1997; Rolands, et al., 1994); (2) computer-based forms using Geolink GPS-based software to collect field data from moving vehicles, and; (3) relevés interpreted from randomly sampled GPS-based points along airborne (figure 4) video transect lines. All field crews used an unlabelled polygon map (paper or electronic) to help delineate polygon boundaries and direct the collection of relevé information that would best typify the vegetation community or communities within each polygonal unit. Field crews were instructed to classify their field vegetation data into specific vegetation communities or land-cover classes from a robust list of theoretical vegetation associations that have been historically found or that are thought to occur in Arizona (Bennett, 1997). In addition, primary field crews had the option to create "new" vegetation community classification categories if warranted. Therefore, field crews were not limited to or biased towards the original Graham land cover classifications.

Our map validation goal was to sample 10-percent of the 58,170 Graham polygons. Consequently, a list of approximately 5,800 randomly selected target polygons was generated without regard for size or location. Unfortunately, many of the selected polygons lie in remote and mountainous places making sampling by foot or backroads impractical given the time and budget constraints. Remote polygons can be sampled by observation from aircraft, but this is both more expensive and less accurate than ground-based methods. We eventually adopted both methods with the bulk of the work being done via sampling from roadways while driving 16,358 km (10,162 miles). Particular attention was taken to reach the random sample polygons, but observations were also made in each polygon traversed by the surface roads. In this manner, we made 5,811 ground-based observations. Multiple observations (within individual polygons) were made within a stratified subset of large polygons to explore vegetation heterogeneity and edge related issues. In addition, 545 observations were made from aircraft to obtain remote targeted random polygons not reached by land. Thus, a total of 6,356 vegetation community-related observations were taken. Of these 1619 were file headers, file footers, and calibration observations. The initial 10% sampling objective was not achieved owing to financial, time, and topographic constraints.

Numerous field tools were created to assist primary assessment crews in the field. Examples include, but are not limited to: a vegetation-community delineation field manual specifying data collection methods, a list of theoretically possible JBK associations, an AZGAP Flora of "prominent" plants, numerous plant identification guides, and generalized paper maps indicating roads and the randomly selected stratified set of polygons to be sampled. In addition, primary assessment teams had at least two weeks of standardized field training. As a matter of efficiency, CPSU/UA crews continuously collected relevé data electronically on each polygon traversed as they drove from target site to target site. In November 1997, all GPS-based field data was converted into ArcInfo coverages for subsequent analysis and comparison to the original map cover types. The field-based relevés were also a prerequisite for recasting the Graham land-cover types into the National Vegetation Classification System (NVCS) and or the Brown, Lowe, and Pase (BLP) vegetation community classification system. Of the 105 original land cover types, only 38 percent of the vegetation classes were consistent with BLP (or NVCS); 33 percent were combinations to two or more vegetation types; 43 percent used combinations of genera that are inconsistent with BLP; and 10 percent mixed elements from two or more biomes.

Although we did not revise (change the shape and area of TM derived polygons) the Graham 58,169 polygon land-cover coverage, we did crosswalk the original descriptive attributes and cover types to meet present GAP standards. After careful examination of the cross-walked Graham data and the addition of assessment field information, we were able to create a new GAP biome/series level map (figure 5) with community descriptions that are compatible with both NVCS and BLP hierarchial vegetation classification systems. Subsequently, the thematic resolution and accuracy assessment of the vegetation cover types could be evaluated in either NVCS or BLP classification systems. In the process many more categories required lumping than splitting. To increase the utility and thematic accuracy of the map for GAP purposes, 105 vegetation cover classes were reduced to 53 types (including 3 non-vegetative classes such as urban, agriculture, and playa).

The hierarchical structure of the BLP system (and also NVCS and TNC) makes is readily applicable to vegetation studies from generalized continental scales to those of small plots. This ability to zoom in or zoom out allows generalization and presentation of data at a level consistent with data available and permits vegetation assessment at varying taxonomic ranks and thematic resolution levels. As a result, we were able to rate each Biome/Series vegetation class (legend element) at varying levels of thematic resolution each with a separate accuracy estimate. This type of accuracy assessment should be of benefit to land managing agencies and map users alike.

We scored agreement between the Graham map polygons and our field classifications according to their level of agreement at the BLP formation, biome, series, and species ranks. A system of weighted scoring was designed to best reflect the overall objectives of the Arizona GAP effort by focusing primarily on distinguishable dominate cover types that are recognizable by a "coarse-filter" Thematic Mapper (TM) imagery approach with a minimum mapping unit of 100 hectares. Correct polygons were awarded points in this manner: a value of 8 for the correct series, 4 for the correct biome, 2 for the correct formation, and 1 for the correct species. Polygon points added together gave a total score between 0 and 15. These numerical values were chosen to reflect the relative degree of importance of correctness and because they add together to give unique values that can be decomposed to determine which taxonomic rankings or classification elements were correct. For example, correct answers for formation, biome, and series and a wrong answer for species would total 14 points (2+4+8+0=14). Similarly, correct answers for biome and series only add to 12 (4+8+0+0=12). Graham's polygonal attribution for the 6,356 relevés we collected was 66 percent correct overall. However, it is more instructive to examine the individual thematic classes and resolution accuracy assessment range estimates.

In the assessment of individual land cover types, as anticipated, desert areas with less than 25% vegetative cover had a disporportionate number of large polygons. In addition, there was a lack of taxonomic equivalence between biogeographic provinces and vegetation types. For example, the Sonoran desert has two commonly recognized subdivisions; the Arizona Upland (Paloverde-Mixed Cacti) and Lower Colorado (Creosote-Bursage). This is in contrast to the Mohave and Chihuahuan deserts which are typically delineated at a rank or series level. In Arizona, the subdivision between the Arizona Upland (Paloverde-Mixed Cacti) and Lower Colorado (Creosote-Bursage) are clearly delineated. On the-other-hand, the differences between the Mohave Desertscrub (Creosote) and the Lower Colorado (Creosote-Bursage) are fuzzy and difficult to delineate in the field, much less with TM classified imagery.

In the preliminary assessment process of the Arizona GAP land cover map, it was determined that vegetation types found in Arizona were inherently complex, intrinsically "fuzzy" at one or more taxonomic ranks, often ecotonal and are not adequately described or equally understood at varying spatial scales or classification levels. TM imagery classification methods and GIS techniques also create additional sources of errors that make it difficult to rely on a single statistical sampling protocol to determine the appropriate sampling area that best "fits" all vegetation types and polygons. For example, polygon sizes for a single vegetation type may vary by several orders of magnitude which thereby may require different sampling protocols to determine the degree and variation of vegetation homogeneity within polygons. Long slender and small polygons, such as riparian vegetation types or small vegetation associations (relics, fragments, and inclusions), may not have been classified as well as large rectangular polygons because of TM resolution, imagery classification procedures, and minimum mapping unit constraints. Also, vegetation communities are inherently fuzzy and ecotonal, especially towards polygon edges. As polygons become smaller, the effect of edges becomes larger and harder to avoid. Whatever-the-case, our understanding and recognition of vegetation communities (legend classes) at various spatial scales and classification levels is not uniform.

 Even when one accepts "defined" units of natural vegetation, it may be difficult to draw a line separating them into distinct taxonomic classes. Moreover, it soon becomes apparent that the various classifications of community types frequently form broad ecotones, intergrading over considerable areas. In addition, natural and anthropogenic disturbances, present and past, may make it difficult to thematically assess an area's natural vegetation into distinct taxonomic groups, unique spectral classes, or vegetative land cover units. However, we conclude from these data that the Arizona GAP program produced an adequate land cover map considering the difficulties entailed with "course" filtered, TM-based imagery classifications of complex Arizona landscapes. It is also much better (finer resolution) than the1980 BLP map (figure 6). It is recommended that thematic resolution be assessed at varying classification ranks and spatial scales to increase the users understanding of "problematic" vegetation cover types and to increase the utility and value of such maps.

 With increasing GPS derived field data, higher resolution imagery, and advanced GIS tools, the ability to rapidly map has improved dramatically. However, the current state of knowledge and understanding of ecologically-defined communities and evolving ecosystems is not sufficient to delineate vegetation types at all taxonomic rankings equally well. More research is required to refine classification systems and more time is required to collect vegetation community information at all spatial and thematic scales. The GAP effort in Arizona made a significant contribution to advance state conservation goals and GAP objectives. Now a more intensive, higher resolution mapping effort will be required to adequately delineate all vegetation associations and or alliances and to determine the status and trends of ecologically important cover types throughout the state.

We would like to thank Bill Halvorson (CPSU/UA Unit Leader) and Ms. Patty Guertin for their field assistance and support. We are grateful to the National Gap Analysis Program and staff for funding to do the field work necessary to conduct an assessment of the Arizona GAP land cover map. We thank Mr. Douglas Richardson of GeoResearch Inc. (Billings, Montana) for his support and the educational use of Geolink mapping software and Mr. Darren Cosandier, of Waypoint Consulting Inc. (Calgary, Alberta) for the use of Grafnav/Grafnet GPS processing software. We are also grateful to Motorola Inc. (Scottsdale, Arizona) for the donation of LGT1000 GPS receivers. Without such corporate support, it would have been difficult to achieve the field data collection efficiencies we achieved during this project.

Bennett, Peter S., R. Roy Johnson, and Michael R. Kunzmann. 1997. "Descriptions of Arizona vegetation from Brown, Lowe, and Pase system at the series level with modified coding and emendations", in press. USGS Cooperative Park Studies Unit, University of Arizona, Tucson, Arizona.

Bennett, Peter S., Michael R. Kunzmann, and William L. Halvorson. 1997. "Field delineation manual for the Arizona GAP program", in press. USGS Cooperative Park Studies Unit, University of Arizona, Tucson, Arizona.

Brown, D.E., C.H. Lowe, and C.P. Pase. 1979. "A digitized classification system for the biotic communities of North America, with community (series) and association examples for the Southwest." Journal Ariz. Nev. Acad. Sci.; 14(suppl. 1): 1-16.

Brown, D.E., C.H. Lowe, and C.P. Pase. 1980. "A digitized systematic classification for ecosystems with an illustrated summary of the natural vegetation of North America." USDA Forest Service General Technical Report RM-73. Rocky Mountain Forest & Range Experiment Station, Ft. Collins, Colorado.

Brown, D. E.. (ed.). 1982. "Biotic communities of the American southwest-United States and Mexico", Desert Plants 4:1-341.

Franklin, J.F.. 1993. "Preserving biodiversity: Species, ecosystems, or landscapes?" Ecological Applications, 3(2):202-205.

Graham, L.A.. 1993. "Airborne video for near-real-time vegetation mapping". Journal of Forestry, 91(8):28-32.

Graham, L.A., V. Ackerly, and V. Meretsky. 1995. "Metadata: Arizona GAP Analysis Program". Unpublished data. The University of Arizona, School Renewable Natural Resources, Tucson, Arizona.

National Gap Analysis Program. 1994. "A handbook for gap analysis." Moscow, Idaho.

Rolands, Peter G.. 1994. (Ed.) "Colorado Plateau Vegetation Assessment and Classification Manual", Technical Report NPS/NAUCPRS/NRTR-94/06, 73pp, USGS Colorado Plateau Research Station, Flagstaff, Arizona.

Scott, J.M., B. Csuti, J.D. Jacobi, and J.E. Estes. 1987. "Species richness". BioScience 37(11):782-787.

Scott, J.M., F. Davis, B. Csuti, R. Noss, B. Butterfield, C. Groves, H. Anderson, S. Caicco, F. D'Erchia, T. Edwards, J. Ulliman, and G. Wright. 1993. "Gap analysis: A geographic approach to protection of biological diversity". Journal of Wildlife Management 57(1) supplement, Wildlife Monographs No. 123.

Whittaker, R.H.. 1975. "Communities and ecosystems", 2nd Ed. Macmillan Publishing Co., Inc., New York.

Michael R. Kunzmann is an USGS Ecologist at the Cooperative Park Studies Unit located at The University of Arizona, Tucson Arizona. Correspondence may be sent to 125 Biological Sciences East, University of Arizona, Tucson, Arizona, 85721. Mike may also be reached by email: mrsk@npscpsu.srnr.arizona.edu. or by telephone at (520) 621-7282.

Royden J. Hobbs is a Research Associate in The University of Arizona's School of Renewable Natural Resources. Correspondence may be sent to 125 Biological Sciences East, University of Arizona, Tucson, Arizona, 85721. Roy may also be reached by telephone at (520) 621-1174 or by email: rhobbs@nexus.srnr.arizona.edu.

Peter S. Bennett is a Senior Research Scientist at the Cooperative Park Studies Unit located at The University of Arizona, Tucson Arizona. Correspondence may be sent to 125 Biological Sciences East, University of Arizona, Tucson, Arizona, 85721. Pete may also be reached by email: Peter_Bennett@USGS.gov or by telephone at (520) 670-6896.