Novel Ensemble-Based Machine Learning Models Based on The Bagging, Boosting and Random Subspace Methods for Landslide Susceptibility Mapping
Abstract Indivisual machine learning models show different limitations such as low generalization power for modeling nonlinear phenomena with complex behavior. In recent years, one of the best approaches to this issue is to use ensemble models. The purpose of this paper is to investigate the predictive power and modeling of three novel ensemble models constructed with four machine learning models: Decision Tree (DT), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Naive Bayes (NB) models based on three approaches of Bagging, boosting and Random Subspace (RS) in landslide susceptibility mapping (LSM) in the Province of Ajloun in Jordan. A total number of 91 landslide locations along with 16 conditioning factors in LSM were identified and used. Also, before modeling, the selection of effective conditioning factors in LSM was done using genetic algorithm and four single models including DT, KNN, NB and SVM. The selected factors were used in modeling with individual and ensemble models. The results show that the area under the receiver operating characteristic curve (AUROC) for ensemble models is significantly higher than the individual models and the AUC for ensemble models was on average 14% higher than individual models. Based on the results, the most accurate models were RS ensemble model (AUROC = 0.850), Boosting (AUROC = 0.848) and Bagging (AUROC = 0.814), respectively. This study showed that by combining the results of simple machine learning models and making ensemble models, models with the desired accuracy can be achieved.