Improved landslide assessment using support vector machine with bagging, boosting, and stacking ensemble machine learning framework in a mountainous watershed, Japan

• Published in: Landslides 2020-03, Vol.17 (3), p.641-658
• Main Authors: Dou, Jie, Yunus, Ali P., Bui, Dieu Tien, Merghadi, Abdelaziz, Sahana, Mehebub, Zhu, Zhongfan, Chen, Chi-Wen, Han, Zheng, Pham, Binh Thai
• Format: Article
• Language: English
• Subjects: Aerial photographs; Aerial photography; Agriculture; Algorithms; Avalanches; Bagging; Civil Engineering; Debris flow; Dependent variables; Disaster management; Disasters; Earth and Environmental Science; Earth Sciences; Emergency preparedness; Fieldwork; Geography; Heavy rainfall; Landslides; Landslides & mudslides; Learning algorithms; Machine learning; Model accuracy; Modelling; Mountains; Natural Hazards; Original Paper; Rain; Rainfall; Stacking; Support vector machines; Tectonics; Training; Watersheds
• ISSN: 1612-510X; 1612-5118
• DOI: 10.1007/s10346-019-01286-5

Heavy rainfall in mountainous terrain can trigger numerous landslides in hill slopes. These landslides can be deadly to the community living downslope with their fast pace, turning failures into catastrophic debris flows and avalanches. Active tectonics coupled with rugged topography in a complex geoenvironment multiplies this likelihood. The available hazard maps are usually helpful in mitigating disasters. However, fool-proof predicting landslide susceptibility identification remains a challenge in landslide discipline. Recently, ensemble machine learning (ML) techniques have proved the potential to provide a more accurate and efficient solution in spatial modeling. The main purposes of the current study are to examine and evaluate the predictive capability of support vector machine hybrid ensemble ML algorithms, i.e., the bagging, boosting, and stacking for modeling the catastrophic rainfall-induced landslide occurrences in the Northern parts of Kyushu Island, at the watershed scale in Japan. In this study, a landslide inventory map containing 265 landslide polygons was first interpreted from the aerial photographs and fieldwork after the September 2017 rainfall event. The raw data were randomly separated into two parts using a 70/30 sampling strategy for training and validating the landslide models. Then, 13 predisposing factors were prepared as predictors and dependent variable. The landslide susceptibility maps (LSM) were validated by the area under the receiver operating characteristic curve (AUC). The results of validation showed that the AUC values of the four models (SVM-Stacking, SVM, SVM-Bagging, and SVM-Boosting) varied from 0.74 to 0.91. The SVM-boosting model outperformed the other models, while SVM-stacking model has found to be the lowest performance. The outcome suggests that an ensemble ML model does not necessarily mean good performance. It is always preferable to select an appropriate model, such as the one proposed the hybrid novel ensemble SVM-boosting model, which could significantly improve the accuracies of LSM. Also, from Information Gain Ratio (IGR) we found that the rainfall factor mainly affects the results, that agrees with the analogy of present study.