Co-seismic landslides are common secondary hazards triggered by strong earthquakes, particularly in mountainous areas with complex geology. Traditional Landslide Susceptibility Prediction models primarily use horizontal ground motion parameters, limiting their ability to capture the three-dimensional propagation and directional energy of seismic waves. To address this, we propose an intelligent risk assessment framework that integrates inverse distance weighting with real seismic data, extracting multi-dimensional parameters like Peak-type, Energy-type, and RMS metrics. The Extremely Randomised Trees algorithm builds spatial susceptibility models, incorporating varying seismic parameters. This framework uses advanced machine learning techniques, particularly autonomous feature extraction, to process large-scale seismic data and identify key parameters without manual input. It adapts dynamically to changing seismic conditions, offering real-time analysis and predictions of landslide susceptibility. A case study of the 2022 Ms 6.8 Luding earthquake in Sichuan, China, demonstrates the superiority of models using three-directional RMS displacement, achieving higher predictive accuracy (AUC = 0.971; F1 score = 0.942). This method captures the sustained, multi-directional nature of seismic excitation, enhancing model accuracy and spatial resolution. This approach provides a robust tool for co-seismic landslide assessment, advancing data-driven early warning and emergency response systems for geohazard management.
Zhang et al. (Sun,) studied this question.