You are here:
TESTING TREE-CLASSIFIER VARIANTS AND ALTERNATE MODELING METHODOLOGIES IN THE EAST GREAT BASIN MAPPING UNIT OF THE SOUTHWEST REGIONAL GAP ANALYSIS PROJECT (SW REGAP)
Citation:
Sajwaj, T. D., W G. Kepner, AND D F. Bradford. TESTING TREE-CLASSIFIER VARIANTS AND ALTERNATE MODELING METHODOLOGIES IN THE EAST GREAT BASIN MAPPING UNIT OF THE SOUTHWEST REGIONAL GAP ANALYSIS PROJECT (SW REGAP). Presented at 13th Annual National Gap Analysis Program Meeting, Fort Collins, CO, October 7-10, 2003.
Impact/Purpose:
The primary objectives of this research are to:
Develop methodologies so that landscape indicator values generated from different sensors on different dates (but in the same areas) are comparable; differences in metric values result from landscape changes and not differences in the sensors;
Quantify relationships between landscape metrics generated from wall-to-wall spatial data and (1) specific parameters related to water resource conditions in different environmental settings across the US, including but not limited to nutrients, sediment, and benthic communities, and (2) multi-species habitat suitability;
Develop and validate multivariate models based on quantification studies;
Develop GIS/model assessment protocols and tools to characterize risk of nutrient and sediment TMDL exceedence;
Complete an initial draft (potentially web based) of a national landscape condition assessment.
This research directly supports long-term goals established in ORDs multiyear plans related to GPRA Goal 2 (Water) and GPRA Goal 4 (Healthy Communities and Ecosystems), although funding for this task comes from Goal 4. Relative to the GRPA Goal 2 multiyear plan, this research is intended to "provide tools to assess and diagnose impairment in aquatic systems and the sources of associated stressors." Relative to the Goal 4 Multiyear Plan this research is intended to (1) provide states and tribes with an ability to assess the condition of waterbodies in a scientifically defensible and representative way, while allowing for aggregation and assessment of trends at multiple scales, (2) assist Federal, State and Local managers in diagnosing the probable cause and forecasting future conditions in a scientifically defensible manner to protect and restore ecosystems, and (3) provide Federal, State and Local managers with a scientifically defensible way to assess current and future ecological conditions, and probable causes of impairments, and a way to evaluate alternative future management scenarios.
Description:
We tested two methods for dataset generation and model construction, and three tree-classifier variants to identify the most parsimonious and thematically accurate mapping methodology for the SW ReGAP project. Competing methodologies were tested in the East Great Basin mapping unit comprising four mapzones in Nevada. Competing approaches to data set generation included the use of averaged digital data values within a training site polygon or use of randomly selected individual pixel values to create modeling data sets. Use of averaged values was faster but created smaller data sets. Use of individual pixels created larger data sets but was slower. Competing approaches to model construction included use of a single model for all vegetation types versus use of multiple presence/absence models for individual vegetation types. Use of the single model required minimal time but did not map all sampled communities, particularly rare types. Use of multiple models mapped all community types, typically with greater map accuracies, but was slower. Competing tree-classifier variants included the use of a simple CART algorithm and two iterative tree algorithms (See5 and Random Forests software). The simple CART algorithm used binary splits of dependent variables to classify data points into "pure" groups. The See5 algorithm used a subset of the data pool to construct a set of decision rules, and then iteratively reconstructed the decision rules based on inaccuracies. The third algorithm used subsets of the data pool and dependent variables to construct numerous decision trees. Each set of decision rules got a "vote" in the final outcome of each pixel in the classified vegetation map. Alternate methodologies were compared based on required time, internal model validation, and accuracy assessments. The alternate methodologies are discussed in terms of the competing interests of time required for completion, final thematic accuracies, and complexity .