PARENT SESSION
Special Session - The Southwest Regional Gap Analysis Project Chair(s): Kepner, William ¹, Ramsey, Douglas², Prior-McGee, Julie³, ¹ Landscape Ecology Branch, Las Vegas, NV² Utah State University, Logan, UT³ New Mexico State University, Las Cruces, NM
Wednesday, March 31, 2004 10:00 AM - 5:00 PM Apollo Room 7

The Gap Analysis Program (GAP) is a national interagency program that maps the distribution of plant communities and selected animal species and compares these distributions with land stewardship to identify biotic elements at potential risk of endangerment. GAP uses Geographic Information System (GIS) technology to assemble and view large amounts of biological and land management data to identify areas (gaps) where conservation efforts may not be sufficient to maintain diversity of living natural resources. Historically, GAP has been conducted by individual states; however this has resulted in inconsistencies in mapped distributions of vegetation types and animal habitat across state lines because of differences in mapping and modeling protocols. This was further compounded from the lack of a national vegetation classification nomenclature. In response to these limitations, GAP embarked on a second-generation effort to conduct the program at a regional scale, using a vegetation classification scheme applicable across the US, and ecoregional units as the basis for segmenting the landscape into manageable units. The program’s first formalized multi-state regional effort includes the five states (Arizona, Colorado, Nevada, New Mexico, and Utah) comprising the Southwest Regional GAP Analysis Project (SW ReGAP).

Overcoming chance agreement in classification tree modeling: predictor variables, training data, and spatial autocorrelation considerations. ^*WALLER, ERIC K. , ¹ Colorado Division of Wildlife, 6060 N. Broadway, Denver, CO, USA

ABSTRACT- Classification trees have substantial merit for vegetation mapping, but their application over large areas requires some special considerations. Error can occur in modeling due to chance agreement arising from inadequate sampling of the diversity of map classes with respect to any combination of predictor variables. The likelihood of this chance agreement increases with the addition of any predictor variables and with spatial autocorrelation. The coincidence of spatial autocorrelation in the sampled field data and spatial autocorrelation in the predictor variables that does not relate to natural patterns is especially problematic. The pseudoreplicating approach employed by the Southwest Regional GAP ensures that its training data are spatially autocorrelated. Even different sample sites can be autocorrelated – the more so the less an area is sampled for its diversity. Additionally, many of the predictor variables used by Southwest ReGAP (especially those that are DEM derived) are heavily spatially autocorrelated as well. It was determined that these predictor variables did not always help in mapping the high plains of Colorado, and often had negative effects. Rather than sacrificing many of these predictor variables, an effective strategy involved identifying areas of disagreement between maps derived with and without DEM derived variables, and increasing the sampling of class diversity in these areas through satellite image interpretation. Although remotely sensed data have lower labeling confidence than field-collected data, overall map accuracy was improved. For large area vegetation mapping, obtaining training data that represent the diversity of classes across landscape variability appears to be paramount.

KEY WORDS: vegetation mapping, classification trees, remote sensing, spatial autocorrelation, Southwest Regional Gap Analysis Project