A physics-based machine learning-informed model for predicting regional earthquake-induced landslides
Citation
Gong, W., Zekkos, D., Clark, M. (2025). A physics-based machine learning-informed model for predicting regional earthquake-induced landslides. Engineering Geology, 349: 107948. Link to paper
Abstract
Reduced complexity physical models are commonly applied to regional earthquake-triggered landslide events because they are computationally efficient and can reproduce observed landsliding patterns with relatively few input parameters. However, a major shortcoming of existing physical models is the overprediction of slope failure arising from the implementation of seismic displacement models that depend on a single critical slope angle. Such models predict that all slopes steeper than the critical slope angle will fail, in contrast to observational data that indicate that only a certain portion do. High-resolution input of spatially variable ground conditions and strength parameters and consideration of topographic amplification would likely reduce overprediction, but these approaches are considered prohibitively expensive at present. In this study, we explore the opportunity to improve the prediction of regional earthquake-induced landslides obtained from physics-based models by applying a computationally efficient machine learning (ML) algorithm to classify the model results. A pseudo-three-dimensional (pseudo-3D) physics-based methodology is used to predict the landslides caused by the 2015 Mw 6.5 Lefkada earthquake and the prediction has broad geospatial overlap with observed landslides and high landslide density areas, however the model also overpredicts landslide occurrence. Classifying the results of the physics-based model using the XGBoost ML algorithm is found to enhance regional earthquake-induced landslide prediction by establishing objective criteria for eliminating false predictions from the physics-based model. The ML model considers five training features, i.e., landslide volume, calculated ratio of three-dimensional (3D) factor of safety (FS) to one-dimensional (1D) FS, slope aspect, elevation, and topographic roughness. Incorporation of these features results in a 45% reduction in the total number of predicted landslides originally predicted by the physics-based method. However, application of the ML algorithm also reduces true positive rate, underscoring the need to fairly assess the tradeoffs between true- and false-predictions in landslide hazard forecasting.