Co-Seismic Landslide Mobility Assessment Using Machine Learning Models

Landslide mobility is essential for landslide risk assessment because impact scales with the distance traveled by the landslide mass in many cases. This paper aims to quantify the mobility of landslides triggered by the M-w 6.5 Lefkada earthquake on November 17, 2015, and develop regression models for landslide travel distance. The study leverages a high-quality landslide inventory that includes the location, area, and volume of 716 landslides, with the source and entire area for each landslide mapped separately. Multivariate linear and machine learning models are used to predict landslide travel distance. The dependent variable, the 3D landslide travel distance, is calculated from the inventory with the digital elevation model. Independent variables include terrain characteristics, material strength, and permanent seismic displacements estimated from seismic displacement models. The results show that terrain characteristics correlate most strongly with landslide travel distances. Furthermore, the multivariate linear regression, random forest regression, and stochastic gradient descent regression have better prediction capacity than k-nearest neighbors' and XGBoost regression.