A comparative evaluation of state-of-the-art ensemble learning algorithms for land cover classification using WorldView-2, Sentinel-2 and ROSIS imagery

Recent advances in airborne and space-based remote sensing technologies and a rapid increase in the use of machine learning (ML) techniques in digital image processing applications have led to a renewed interest in the classification of satellite imagery. Decision-tree based ensemble learning (EL) algorithms, one of the popular ML techniques, have received considerable attention from researchers due to their simplicity, computational effectiveness and interpretability compared to black-box algorithms. The main goal of this study is to evaluate the supervised classification performance of the advanced decision-tree based EL algorithms, including rotation forest (RotFor), random ferns (RFerns), canonical correlation forest (CCF), extreme gradient boosting (XGBoost), light gradient boosting machines (LightGBM) and categorical boosting (CatBoost) using satellite imagery with different spatial and spectral resolutions. Two well-known EL algorithms, namely, adaptive boosting (AdaBoost) and random forest (RF), were also considered to compare their classification performances. In order to achieve the desired goal, three satellite imagery, namely WorldView-2, Sentinel-2 and hyperspectral ROSIS, were utilized as the fundamental datasets. Results of the study showed that the highest overall accuracy values (i.e. 92.65% for WorldView-2, 92.80% for Sentinel-2 and 95.70% for Pavia datasets) were estimated using CCF, LightGBM and RotFor algorithms, respectively. According to McNemar’s test result, the difference between the predictions of RotFor and CCF algorithms on test samples of the hyperspectral image was found to be statistically insignificant. On the other hand, the lowest classification accuracy values were obtained by the RFerns algorithm in all cases. In addition, while the CCF showed superior classification performance for high spatial resolution WorldView-2 and Pavia images, the highest classification performance was acquired with the LightGBM algorithm for medium spatial resolution Sentinel-2 image. As a result, the performances of advanced EL algorithms were found to be more robust than the well-known RF and AdaBoost ensemble classifiers.