HC-DT/SVM: a tightly coupled hybrid decision tree and support vector machines algorithm with application to land cover change detections

Abstract

Change detection techniques have been widely used in satellite based environmental monitoring. Multi-date classification is an important change detection technique in remote sensing. In this study, we propose a hybrid algorithm called HC-DT/SVM, that tightly couples a Decision Tree (DT) algorithm and a Support Vector Machine (SVM) algorithm for land cover change detections. We aim at improving the interpretability of the classification results and classification accuracies simultaneously. The hybrid algorithm first constructs a DT classifier using all the training samples and then sends the samples under the ill-classified decision tree branches to a SVM classifier for further training. The ill-classified decision tree branches are linked to the SVM classifier and testing samples are classified jointly by the linked DT and SVM classifiers. Experiments using a dataset that consists of two Landsat TM scenes of southern China region show that the hybrid algorithm can significantly improve the classification accuracies of the classic DT classifier and improve its interpretability at the same time.