Application and comparison of different ensemble learning machines combining with a novel sampling strategy for shallow landslide susceptibility mapping

The existence of shallow landslide brings huge threats to the human lives and economic development, as the Lang County, Southeastern Tibet prone to landslide. Landslide susceptibility mapping (LSM) is considered as the key for the prevention of hazard. The primary goal of the present study is to assess and compare four models: classification and regression tree, gradient boosting decision tree (GBDT), adaptive boosting-decision tree and random forest for the performance of landslide susceptibility modeling. Firstly, a landslide inventory map consisting of 229 historical shallow landslide locations was prepared and the same number of non-landslide points was determined by k-means clustering. Secondly, 12 conditioning factors were considered in the landslide susceptibility modeling. The prediction performance of the four models were estimated by fivefold cross validation and relative operating characteristic curve (ROC), area under the ROC curve (AUC) and statistical measures. The results showed that the GBDT performed best in the training and validation dataset, with the highest prediction capability (AUC = 0.986 and 0.940), highest accuracy value (95.3% and 88.1%) and highest kappa index (0.904 and 0.772), respectively. Therefore, the GBDT was considered to be the most suitable model and applied to the whole study area for LSM. The results of this study also demonstrate that the performance can be enhanced with the use of ensemble learning. The sampling strategy of non-landslide points can be improved by combining with clustering analysis which are more reasonable.

Automated summary

This study aims to evaluate and compare the performance of four models for landslide susceptibility modeling, with gradient boosting decision tree (GBDT) performing best in both training and validation datasets. The results indicate that GBDT is the most suitable model and can be further enhanced by combining with clustering analysis for improved sampling strategy of non-landslide points.