Predictive modeling of forest cover type and tree canopy height in the central Rocky Mountains of Utah

Maps of forest cover type and canopy height are needed for LANDFIRE, a multi-scale fire risk assessment project designed to generate intermediate-resolution data of vegetation and fire fuel characteristics for the U.S. Here we describe an evaluation study in the central Rockies of Utah, comparing tree-based methods, multivariate adaptive regression splines (MARS), and a hybrid method for mapping forest cover and canopy height on the basis of more than 2,000 forest inventory ground plots in the seven million ha mapping zone. The two forest attributes were modeled as functions of a variety of predictor variables, including: Landsat 7 Enhanced Thematic Mapper Plus (ETM+) images acquired at three different seasons; Tasseled-cap brightness, greenness, and wetness; a forest type group map; and topographic variables derived from Digital Elevation Models (DEMs); and other ancillary variables. The hybrid modeling approach showed a marked increase in overall and within forest cover type accuracies, outperforming the tree-based and MARS approaches. Little difference was seen in global performance measures of forest canopy height models, but patterns in residual plots resulting from different modeling approaches raise questions about utility of height predictions in different applications.