Development of a statistics-based nowcasting model for earthquake-triggered landslides in Taiwan

Abstract

Earthquake-triggered landslides are one of the major hazards in tectonically active mountain belts, and a nowcasting (near-real-time assessment) model for rapid assessment of coseismic landsliding is crucial for hazard mitigation and management. The purpose of this study is to develop a nowcasting model by using a logistic regression model. We use an inventory of earthquake-triggered landslides from the 1999 Mw7.6 Chi-Chi earthquake to train the model at a spatial resolution of 40 m. We use a multi-stage framework to choose dominant variables for developing an optimal model. Our result shows that the model contains three key variables, peak ground acceleration (PGA), topographic roughness, and lithology, and two essential combined variables, PGA*slope, and PGA*roughness, to enhance the interaction between ground shaking and topography. By using the landslide inventory of the 1998 Mw5.7 Jueili earthquake, the probability threshold of 0.1 is chosen to define predicted landslides, and the balanced accuracy is above 0.95 in the prediction. The development of this model can effectively estimate possible earthquake-triggered landslides with the input of real-time PGA values, and facilitate a rapid response in emergencies based on the potential landslide distributions in Taiwan.