Uncertainty research of landslide susceptibility mapping based deep ensemble learning: different basic classifier and ensemble strategy

The Jurassic red-strata of the Three Gorges Reservoir Area in China is interbedded of thick siltstone and thin sandy-mudstone and contains many clay minerals, such as montmorillonite and illite, which is water sensitive, weak and expansive, and easy to decompose by water weathering. In particular, due to the seasonal rainfall, development of settlements, and large-scale reservoir impoundment, many slow-moving landslides (e.g., deep rotation and planar landslides) often occur. Notwithstanding, the reconnaissance, updating, and mapping of kinematic features of township area landslides lack the appropriate attention of the government and researchers. Landslide susceptibility mapping is necessary prerequisites for landslide hazard and risk assessment. But a certain proportion of unpredictability is always closely related to modeling. The main objective of this work is to introduce deep ensemble learning into landslide susceptibility assessment to improve the performance of maximum likelihood models. Therefore, the current model construction has focused on three basic classifiers: decision tree, support vector machine, multi-layer perceptron neural network model, and two homogeneous ensemble models: random forest and extreme gradient boosting. Two prominent ensemble techniques—homogeneous/heterogeneous model ensemble and bagging, boosting, stacking ensemble strategy—were applied to implement the deep ensemble learning. Then, thirteen influencing factors were prepared as predictors and dependent variables. The landslide susceptibility maps were validated by the area under the receiver operating characteristic curve. The results of validation showed that the ensemble model shows that the ROC/AUC value is higher than 0.9, which is improved compared with the basic classifiers. Deep ensemble learning focuses more on detecting the landslide susceptibility area with the highest probability of occurrence. The Stacking based RF-XGBoost model obtained the best verification score (AUC=0.955). The comparison between the susceptibility map and landslide inventory data is encouraging as most of the recorded landslide pixels (about 83.3%) are at a high susceptibility level. Besides, from the information gain rate, we found that the Yangtze River and human engineering activities mainly affect the results, which is consistent with the current situation in the study area. The research results in the township-level landslide susceptibility map can also be extended to other urban and rural areas affected by landslides to reduce the landslide disaster risk and formulate further development strategies.