Landslide risk prediction in near-fault areas based on 3D ground motion simulation

Seismic landslide risk assessment in near-fault areas requires highly reliable ground motion input data. This study takes the 2016 Kumamoto, Japan, M w 7.1 earthquake as a case study and proposes an integrated assessment framework that combines three-dimensional (3D) ground motion simulation with a machine learning-based Random Forest (RF) model. By utilizing physics-based simulation via a 3D finite-difference method, the framework can reproduce key near-fault features, including pulse-like ground motions (PLGMs) induced by fault rupture directivity effect, thereby providing a reliable data foundation for seismic landslide risk assessment. The landslide susceptibility map generated using this integrated framework shows strong spatial consistency with the actual landslide distribution. The RF model achieves an overall prediction accuracy of 88.6%, validating the effectiveness and robustness. Furthermore, the extracted pulse PGV from the PLGM is identified as the most influential variable in the predictive model accounting for more than 40% of the total importance, indicating that PLGM characteristics are key controlling factors for co-seismic landslide triggering in the study area. This study indicates that integrating 3D ground motion simulation to account for the pulse characteristics of near-fault ground motions provides an effective physics-enhanced pathway for improving co-seismic landslide risk assessment.