Evaluating MM3 Multimodal Model Against High-Resolution Mapping for Earthquake-Induced Landslide Assessment in Kaikoura, New Zealand

Abstract

On November 14, 2016, New Zealand experienced an Mw 7.8 earthquake that triggered over 30,000 landslides. A high-resolution landslide inventory was developed within the study area, mapping 1,082 landslides. Approximately 48% of the inventory consisted of first-time failures, while approximately 52% were classified as reactivation failures. Landslides were categorized into two failure mechanisms: shallow and deep. Geospatial data analysis and frequency ratio methods applied to the inventory revealed that topographic amplification factor (TAF), slope, geology, and normalized difference vegetation index (NDVI) are variables that strongly correlate with and significantly influence the occurrence of earthquake-induced landslides. A sensitivity analysis of the Multimodal Physics-Based Model (MM3) showed that peak ground acceleration (PGA), friction angle, and cohesion are the most influential variables for determining failure probability. The shallow model also demonstrated high sensitivity to soil depth. MM3 was then validated against high-resolution mapping to evaluate its predictive performance through more than 25,000 simulations. This evaluation resulted in a global model accuracy (area under the curve, AUC) of 0.82 and 0.79 for shallow and deep landslides, respectively.