Ensemble learning framework for landslide susceptibility mapping: Different basic classifier and ensemble strategy

The application of ensemble learning models has been continuously improved in recent landslide suscep-tibility research, but most studies have no unified ensemble framework. Moreover, few papers have dis-cussed the applicability of the ensemble learning model in landslide susceptibility mapping at the township level. This study aims at defining a robust ensemble framework that can become the bench-mark method for future research dealing with the comparison of different ensemble models. For this pur-pose, the present work focuses on three different basic classifiers: decision tree (DT), support vector machine (SVM), and multi-layer perceptron neural network model (MLPNN) and two homogeneous ensemble models such as random forest (RF) and extreme gradient boosting (XGBoost). The hierarchical construction of deep ensemble relied on two leading ensemble technologies (i.e., homogeneous/hetero-geneous model ensemble and bagging, boosting, stacking ensemble strategy) to provide a more accurate and effective spatial probability of landslide occurrence. The selected study area is Dazhou town, located in the Jurassic red-strata area in the Three Gorges Reservoir Area of China, which is a strategic economic area currently characterized by widespread landslide risk. Based on a long-term field investigation, the inventory counting thirty-three slow-moving landslide polygons was drawn. The results show that the ensemble models do not necessarily perform better; for instance, the Bagging based DT-SVM-MLPNN-XGBoost model performed worse than the single XGBoost model. Amongst the eleven tested models, the Stacking based RF-XGBoost model, which is a homogeneous model based on bagging, boosting, and stacking ensemble, showed the highest capability of predicting the landslide-affected areas. Besides, the factor behaviors of DT, SVM, MLPNN, RF and XGBoost models reflected the characteristics of slow-moving landslides in the Three Gorges reservoir area, wherein unfavorable lithological conditions and intense human engineering activities (i.e., reservoir water level fluctuation, residential area construc-tion, and farmland development) are proven to be the key triggers. The presented approach could be used for landslide spatial occurrence prediction in similar regions and other fields.& COPY; 2023 China University of Geosciences (Beijing) and Peking University. Published by Elsevier B.V. on behalf of China University of Geosciences (Beijing). This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Automated summary

This study aims to define a robust ensemble framework that can be used as a benchmark method for future research comparing different ensemble models. By using decision tree, support vector machine, multi-layer perceptron neural network, random forest, and extreme gradient boosting models, it provides accurate and effective spatial probability predictions of landslide occurrence. Results from the study conducted in Dazhou town, China, show that the stacking based random forest and extreme gradient boosting model has the highest capability in predicting landslide-affected areas.