Object-based classification with rotation forest ensemble learning algorithm using very-high-resolution WorldView-2 image

Abstract

Machine learning algorithms reported to be robust and superior to the conventional parametric classifiers have been recently employed in object-based classification. Within these algorithms, ensemble learning methods that construct set of individual classifiers and combining their predictions to make final decision about unlabelled data have been successfully applied. In this study, performance and effectiveness of a novel ensemble learning algorithm, rotation forest (RotFor) aiming to build diverse and accurate classifiers, was investigated for the first time in object-based classification using a WorldView-2 (WV-2) satellite image. Also, the combination of satellite imagery and ancillary data (i.e. normalized difference vegetation index and principal components) were assessed. Random forest (RF), support vector machine (SVM) and nearest neighbour (NN) algorithms were also used as benchmark classifiers to evaluate the power of RotFor. The classification results confirmed that integration of ancillary data increased the classification accuracy in comparison to using solely spectral bands of WV-2. While RotFor and SVM generally produced similar results, they outperformed the RF and NN based on McNemar's and Wilcoxon's signed-rank test of statistical significance results.