Comparison of data-driven models of loess landslide runout distance estimation

Abstract

Irrigation-induced landslides with long runout distances endanger local communities. Estimating runout distance of landslides may contribute to the mitigation of potential hazards. Conventional mechanism-related methods require a series of experiments and/or numerical simulations that are commonly time-consuming and expensive, yet data-driven models reduce the experimental workload and require less prior knowledge in the geological history as well as mechanical behavior of the material. A data-driven model is proposed to forecast landslide runout distance using geometrical characteristics of the landslide. The geometrical dataset of the shallow loess landslides and loess-bedrock landslides occurred in Heifangtai terrace, China, was employed to develop the model. All geometrical datasets were obtained from field investigation and monitoring. Seven data-mining techniques were used and compared for runout estimation, among which the most optimal technique was integrated in the estimation model for loess slope failures. The multi-layer perceptron method outperforms other algorithms, and thus it was selected for the runout distance estimation model. Parametric models are constructed to fit runout distance based on the estimation. Hazard analysis measurements, including value-at-risk (VaR) and tail-value-at-risk (TVaR), are computed for the parametric distributions, which shows the potential area of impact and number of residential clusters at risk.