An Experimental Comparison of the Effects of Ensemble Learning Methods, Mapping Scale, and Information Hierarchy on Prediction Accuracy of Rare Events

• Activity Number: 526
• Type: Contributed
• Date/Time: Thursday, August 11, 2005 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Computing
• Author(s): Zhaofei Fan*+ and Stephen S. Lee and Stephen Shifley and Frank R. Thompson and David R. Larsen
• Companies: University of Missouri, Columbia and University of Idaho and U.S.D.A. Forest Service and U.S.D.A. Forest Service and University of Missouri, Columbia
• Keywords: spatial prediction ; cavity tree ; statistical model

Abstract

The accuracy of spatial prediction and mapping of rare events (e.g., cavity trees used by wildlife, dead trees, or unusual plant and animal species) is greatly affected by three interrelated factors: relative frequency of objects (events), mapping scale used for sampling or modeling, and the strength of ecological relationships (represented by statistical models or rules that link the object with a set of ecological and environmental factors that serve as covariates). Information about the frequency and location of rare events is typically difficult and costly to obtain, so rare event occurrence is often modeled using covariates. However, such models generally suffer from poor precision. Compared to results for single models of rare events, model ensemble learning (simultaneously employing ensembles of as many as several hundred models) can compensate in varying degrees for the loss of prediction accuracy due to the incomplete information. We resampled cavity tree (rare event) distribution data from the Missouri Ozark Forest Ecosystem Project to represent levels of object occurrence (frequency).