关键词: 滑坡易发性评估, 地震滑坡, 机器学习, 支持向量机, Newmark位移模型

Abstract:

The Newmark displacement model is a common physically based model for earthquake induced landslide susceptibility mapping. The machine learning model is one of the statistical methods and is increasingly applied to landslide susceptibility mapping in recent years. The main purpose of this paper is to combine the Newmark displacement method with machine learning model for developing a mechanism-based earthquake-induced landslide susceptibility model and improving the predictive accuracy. The support vector machine (SVM) method was selected and eight thematic data layers, including landslide inventory, topographic relief, slope, curvature, Newmark displacement value, distance from faults, distance from drainages and distance from roads, were prepared in GIS. A total of 1,900 landslides were subsequently randomly divided into two subsets: a training subset comprising 70% of the landslides and a validation subset containing the remaining 30%. The model is then applied to Wenchuan County, which was one of the most severely affected areas during the May 12, 2008 (Mw 7.9) Wenchuan earthquake in China. The model performance was evaluated using the receiver operation characteristic (ROC) curve and the model estimation uncertainty in comparison with the other two statistical methods: frequency ratio (FR) bivariate statistical model and the logistic regression (LR) multivariate statistical model. The results show that five variables including topographic relief, Newmark displacement value, distance from faults, distance from drainages and distance from roads are finally selected based on a significance level of 0.05 and the multi-collinearity detection. The values of the area under the ROC curve (AUC) demonstrate that the SVM model exhibited the highest accuracy for the training and validation data sets with AUC values of 0.876 and 0.851, respectively, followed by the LR model (AUC values of 0.836 and 0.842 for training and validation, respectively) and FR model (AUC values of 0.844 and 0.808 for training and validation, respectively). For the evaluation of the model prediction uncertainties, the pixels classified as high susceptibility (probability ≥ 0.75) and low susceptibility (probability < 0.25) are more valuable and practical, both of which have high reliability to determine whether the location is stable. The prediction variation is low for pixels classified as high susceptibility and low susceptibility (with an average of the standard deviation less than 0.05), indicating that the SVM and LR models consistently identified these pixels as stable or unstable. Furthermore, these high and low susceptibility pixels account for about 70% of all pixels for the SVM model, which is about 25% higher than that of LR model. It means that the SVM model performs better than the LR model in these high reliability pixels. For the pixels classified as intermediate susceptibility (probability 0.25 ~ 0.75), the standard deviation of the predictive probability of the SVM model is about 0.09, which is larger than that of the LR model. It indicates that the SVM model exhibited larger uncertainty than the LR model in these intermediate susceptibility pixels. However, it is difficult to determine whether these pixels are stable or not. Also, the pixels with intermediate susceptibility only account for about 30% of the total samples for the SVM model. In general, the SVM model combined with the Newmark displacement method outperform the LR model in accuracy and uncertainty evaluation. The proposed method was applied to produce the earthquake-triggered landslide susceptibility map of Wenchuan County. The comparison of landslide susceptibility map and actual landslide distribution showed that the high susceptibility areas can account for about 80.4% of the actual landslides. This indicates that the combination of support vector machine and the Newmark displacement method has a higher predictive value. The proposed method can potentially help risk assessment and effective management of landslides caused by earthquakes.

Key words: landslide susceptibility assessment, earthquake-triggered landslides, machine learning, support vector machine, Newmark displacement model