PARENT SESSION
Poster Session #44: Remote Sensing and GIS.
Wednesday, August 7. Presentation from 5:00 PM to 6:30 PM. Exhibit Hall B & C, TCC

Modern CART-based classifiers: higher prediction accuracy for niche-based vegetation mapping.

Sexton, Joseph^*,1,2, Cutler, Adele³, ¹ Landscape Ecology: Modeling and Analysis Center, Logan, Utah² Department of Forest Resources, Logan, Utah³ Department of Mathematics and Statistics, Logan, Utah

ABSTRACT- The static (i.e., Hutchinsonian) view of realized niche can be represented as a subspace of an environmental hypervolume, wherein the entity of interest (e.g., species, community, landcover type) is logically true. This concept is analogous to a geometric description of binary statistical classification. Traditionally, statistical and other classification techniques have been employed in simultaneously predictive and interpretive roles (usually to the detriment of both), and CART has been demonstrated as a reasonable compromise between the two purposes. However, in many applications the value of accurate prediction outweighs that of human interpretation; among these cases is the production of geographic maps of species distributions, e.g., Gap Analysis. To improve prediction and model stability, aggregate tree-based modeling techniques have been developed in statistical and computer science research since the 1984 debut of CART. The current state of the art includes random forests and Adaboost. Each of these techniques can be used to aggregate the predictions of individual CART models, using--rather than avoiding--the inherent instability of single models in order to provide better model generalization accuracy as defined by independent dataset misclassification rates. En route to mapping the occupied and unoccupied habitat of quaking aspen (Populus tremuloides, Michx.) in the Book Cliffs of Utah, we modeled the static niche of aspen over microclimatic variables (e.g., monthly estimates of potential direct solar radiation and evapotranspiration) in a raster GIS. Favoring predictive accuracy over interpretability, we compared several statistical and algorithmic classification techniques, including random forests and Adaboost. Results thus far indicate that, although computationally intensive, modern agglomerative tree-based classifiers can outperform linear and single-tree classifiers when prediction is the ultimate goal and data are not limiting.

KEY WORDS: GIS, biogeography, niche, landcover mapping