Statistical Learning Methods for Tree Classification using Remote Sensing Imagery

Sarah Maebius

Advised By: Tom Allen

Over the past decade, Western Redcedars have reportedly been in decline. This tree species is native to the Pacific Northwest and necessary to the ecosystem. The decline of Western Redcedars lacks published literature, so this research seeks to answer the question of where Western Redcedars are located to better understand the decline of this species. The approach to this question applies statistical learning methods to remote-sensing RGB imagery at a spatial resolution of 3 meters combined with field-level data to predict six tree species: Bigleaf Maple, Douglas-Fir, English Oak, Giant Sequoia, Norway Maple, Western Redcedar. The models trained in the research were random forest models and support vector machines. The best performing model was a random forest model with 7 predictors on 5 tree species (by grouping maple and oak trees into a "Broadleaf" category) with an overall testing accuracy of 0.65 and 0.22 for Western Redcedars. This model was applied to a masked image spanning Portland's city boundary to predict the locations of Western Redcedars. The model allows researchers to locate clusters of Western Redcedars for tracking the progress of the species overtime in locations where field crews do not have the resources to visit and record tree information. This research also improves methods of classification using freely accessed satellite imagery with low spatial and spectral resolution.