Improving Classification Tree Analysis for Remotely Sensed Data: Boosting and Bagging Algorithms

摘要

Classification tree analysis (CTA, also known as decision trees, classification and regression trees, CART, and recursive binary partitioning) has received increasing use for classifying remotely sensed data because of its high accuracy results, ease of incorporating ancillary data, and interpretability of results. Relatively recent statistical enhancements to CTA using boosting and bagging have the potential to increase accuracies and overcome certain limitations of CTA. We examined three ensemble methods, traditional boosting using SeeS software, stochastic gradient boosting using TreeNet, and Breiman Cutler Classification using RandomForest. Classification accuracies were improved in all cases over traditional classification methods except for classification of Landsat imagery with stochastic gradient boosting. Traditional boosting using See5 was the easiest to implement because of an interface between See5 and ERDAS Imagine, while Breiman Cutler Classification had the advantage of providing a reliable internal estimate of classification accuracy without the need for independent accuracy assessment data.