Utility of Classification and Regression Tree Analyses and Vegetation in Mountain Permafrost Models, Yukon, Canada

Classification and regression tree (CART) analyses were undertaken to test the usefulness of including vegetation variables in mountain permafrost distribution models for five widely spaced study areas in the Yukon. Digital elevation model (DEM)‐derived variables, field‐derived vegetation variables and satellite imagery‐derived vegetation variables were employed individually to classify sites into permafrost probable, permafrost improbable and permafrost ‘uncertain’ categories. The vegetation variables were subsequently combined with the DEM‐derived set to see if they could improve the latter's accuracy. Overall training accuracies for the probable and improbable permafrost categories for 102 sites ranged from 81% to 92%. Remotely sensed imagery alone had the lowest overall training (81%) and testing (50%) accuracies. The CART that combined imagery and DEM‐based variables produced high overall accuracy for training (90%) and the highest for testing (77%), had few nodes classified as ‘uncertain’ and could be used to create permafrost probability maps of the study areas. CART analyses appear useful for predicting permafrost distribution because they can incorporate non‐linear relationships between independent variables and the presence of permafrost. Remotely sensed variables relating to vegetation, specifically a normalised difference vegetation index, improved the DEM‐based results, but required considerable additional effort for data collection and processing. Copyright © 2011 John Wiley & Sons, Ltd.