An Evaluation of Bagging, Boosting, and Random Forests for Land-Cover Classification in Cape Cod, Massachusetts, USA

Abstract

The iterative and convergent nature of ensemble learning algorithms provides potential for improving classification of complex landscapes. This study performs land-cover classification in a heterogeneous Massachusetts landscape by comparing three ensemble learning techniques (bagging, boosting, and random forests) and a non-ensemble learning algorithm (classification trees) using multiple criteria related to algorithm and training data characteristics. The ensemble learning algorithms had comparably high accuracy (Kappa range: 0.76-0.78), which was 11% higher than that of classification trees. Ensemble learning techniques were not influenced by calibration data variability, were robust to one-fifth calibration data noise, and insensitive to a 50% reduction in calibration data size.