| Abstrakt: |
In recent years, numerous landslide susceptibility assessment studies using conventional machine learning models have been carried out, and a series of achievements have been made. To acquire better landslide susceptibility mapping results, various ensemble techniques have been adopted to construct strong classifiers for landslide susceptibility prediction. Generally, for the same base classifier, it is necessary to compare the effects of multiple ensemble techniques and determine the best one. In this paper, a naïve Bayes tree (NBTree) was employed as the base classifier, and three popular ensemble techniques, namely, Bagging (Bag), RandomSubSpace (RS), and MultiBoostAB (MB), were applied to build ensemble landslide susceptibility models for Jian'ge County, China. Herein, a total of 262 landslides were included in the landslide inventory map. Then, the training and validation datasets were randomly divided at a ratio of 70/30. Moreover, fifteen conditioning factors related to topography, geology, vegetation, and human activities were selected to train landslide susceptibility models. The correlations between conditioning factors and landslide occurrence were also measured by weights of evidence (WoE). Ultimately, the performance of each landslide susceptibility model was quantitatively evaluated by receiver operating characteristic (ROC) curves and areas under the curves (AUCs). The results show that all the ensemble models outperform the NBTree model with the validation datasets, and the Bag-NBTree model exhibits the best performance on the processing validation dataset (AUC = 0.852). Additionally, as landslide susceptibility levels are escalated, the corresponding frequency of landslide occurrence significantly increases, indicating that the landslide susceptibility maps (LSMs) produced by the four models are rational and effective. Overall, this study is of great significance to landslide prevention and mitigation in Jian'ge County. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink