Probabilistic back-analysis of earthquake-induced 3D landslide model parameters and risk assessment for secondary slide

Abstract

Back-analysis is an effective method for rapidly estimating soil strength parameters. However, soil spatial variability and the influence of autocorrelation function (ACF) are often inadequately considered. This study presents an efficient probabilistic Bayesian back-analysis for spatially varying soil parameters in earthquake-induced 3D landslide models. A surrogate model based on the Variance Reduction Stochastic Response Surface Method (VRSRSM) is proposed, incorporating five different variance reduction functions associated with ACFs to address the spatial variability of 3D slope under seismic conditions. An improved Hamiltonian Monte Carlo sampling method facilitates Bayesian inference with minimal computational effort. The approach is validated using a 3D simple slope under seismic conditions, accounting for numerical model uncertainty. A case study of a deep-seated landslide from the 2016 Kumamoto earthquake is then used to back-analyze soil strength parameters and unit weight, which are subsequently utilized for risk assessment of secondary slide under aftershocks. Results indicate that VRSRM accurately approximates both 3D simple slope and real landslide models, while the commonly used single exponential ACF yields an unconservative factor of safety, affecting the accuracy of the back-analyzed soil parameters. This proposed approach offers an effective tool for rapidly determining spatially varying soil parameters from landslide events, enhancing risk assessment for future aftershocks.