A Heterogeneous Sampling Strategy to Model Earthquake-Triggered Landslides

Abstract

Regional modeling of landslide hazards is an essential tool for the assessment and management of risk in mountain environments. Previous studies that have focused on modeling earthquake-triggered landslides report high prediction accuracies. However, it is common to use a validation strategy with an equal number of landslide and non-landslide samples, scattered homogeneously across the study area. Consequently, there are overestimations in the epicenter area, and the spatial pattern of modeled locations does not agree well with real events. In order to improve landslide hazard mapping, we proposed a spatially heterogeneous non-landslide sampling strategy by considering local ratios of landslide to non-landslide area. Coseismic landslides triggered by the 2008 Wenchuan Earthquake on the eastern Tibetan Plateau were used as an example. To assess the performance of the new strategy, we trained two random forest models that shared the same hyperparameters. The first was trained using samples from the new heterogeneous strategy, and the second used the traditional approach. In each case the spatial match between modeled and measured (interpreted) landslides was examined by scatterplot, with a 2 km-by-2 km fishnet. Although the traditional approach achieved higher AUCROC (0.95) accuracy than the proposed one (0.85), the coefficient of determination (R2) for the new strategy (0.88) was much higher than for the traditional strategy (0.55). Our results indicate that the proposed strategy outperforms the traditional one when comparing against landslide inventory data. Our work demonstrates that higher prediction accuracies in landslide hazard modeling may be deceptive, and validation of the modeled spatial pattern should be prioritized. The proposed method may also be used to improve the mapping of precipitation-induced landslides. Application of the proposed strategy could benefit precise assessment of landslide risks in mountain environments.