Meta-prediction of Bromus tectorum invasion in central utah, United States

Abstract

Cheatgrass (Bromus tectorumj is an invasive, exotic grass infesting the Western US. Multi-temporal Landsat tm imagery and ancillary topographic data were used for mapping this invasion over portions of Utah. Tobit, logit, probit, and Projection Adjustment by Contribution Estimation (pace) regression, neural networks, and additive regression of regression trees were tested individually, and in an ensemble, Tobit regression had the best performance as an individual predictor. Tobit was most frequently the best predictor of zero cheatgrass coverage. A meta-predictor (classifier) to choose the best predictive model was implemented on a pixel-by-pixel basis. A J48 classification tree as a meta- predictor resulted in an increase in accuracy over the best performer in the ensemble. This study illustrated the potential for meta-prediction as a general technique for increasing accuracy from a collection of base predictors. © 2009 American Society for Photogrammetry and Remote Sensing.