Efficient Stochastic Analysis of Landslide Post-Failure Characteristics Using Dimension Reduction Method and Machine Learning

Landslides pose a significant global threat due to their sudden occurrence and their widespread influence. The inherent spatial variability of soil properties brings great uncertainty to the post-failure behavior of landslides. Generally, the assessment of the uncertainties of slope post-failure characteristics through probabilistic large deformation analysis is determined for assessing and mitigating landslide risks. However, probabilistic landslide simulation based on brute-force Monte Carlo is complex and time-consuming. Therefore, this study proposes a two-stage dimension reduction method and machine learning model-based stochastic approach for improving the efficiency of quantifying the uncertainties of slope post-failure characteristics, which considers the spatial variability of shear strength. In the proposed method, the random field is firstly generated using the Karhunen-Loève expansion (KLE) method. Then, the slope post-failure characteristics parameters of pre analysis samples are obtained using the generalized interpolation material point method (GIMP). Finally, based on the extreme gradient boosting (XGBoost) with sliced inverse regression (SIR) technique, a binary classification surrogate model is constructed to relate random field variables to the maximum relative displacement (Rd) in the characteristic parameters. Results show that the reduced input dimension from 1080 to 100 through KLE is still too high for constructing the surrogate model, while SIR is used for the second dimensional reduction from 100 to 20 of input variables to improve the accuracy of the surrogate model up to 0.98. The proposed model is used to identify hazardous samples, and we simply utilize large deformation analysis on hazardous samples. It suggests that the proposed method in this study is capable of quantifying the uncertainties of slope post-failure characteristics with accuracy and high efficiency dropping time-consuming GIMP.