Performance Analysis of Logistic Model Tree-Based Ensemble Learning Algorithms for Landslide Susceptibility Mapping

Abstract

Landslide susceptibility prediction (LSP) is the key technology in landslide monitoring, warning, and evaluation. In recent years, a lot of research on LSP has focused on machine learning algorithms, and the ensemble learning algorithm is a new direction to build the optimal prediction. Logistic model tree (LMT) combines the advantages of decision tree and logistic regression, which is smaller and more robust than ordinary algorithms. The main aim of this study is to construct and test LMT-based random forest (RF) and selected ensemble learning algorithms including bagging and boosting algorithms to compare their performance. Firstly, taking the county of Ziyang, China, as the study area, through historical reports, aerial-photo interpretations, and field investigations, 690 inventory maps of landslide locations were constructed and randomly divided into the 70/30 ratio for a training and validation dataset. Secondly, considering geological conditions, and landslide-induced disease and its characteristics, 14 landslide-conditioning factors was selected. Thirdly, the variance-inflation factor (VIF) and tolerance (TOL) were used to analyze the 14 factors, and the prediction ability was calculated with information-gain technology. Ultimately, the receiver-operating-characteristic (ROC) curve was applied to verify and compare model performance. Results showed that the LMT-RF model (0.897) was superior to other models, and the performance of LMT single model (0.791) was the worst. Therefore, it can be inferred that the LMT-RF model is a promising model, and the outcome of this study will be useful to planners and scientists in landslide sensitivity studies in similar situations.