Evaluation of coseismic landslide susceptibility by combining Newmark model and XGBoost algorithm

Abstract

Coseismic landslides are among the most perilous geological disasters in hilly places after earthquakes. Precise assessment of coseismic landslide susceptibility is crucial for forecasting the effects of landslides and alleviating subsequent tragedies. This research formulates a comprehensive landslide hazard assessment model by integrating the Newmark physical model with machine learning techniques. The Jiuzhaigou region serves as the study area, utilizing the comprehensive indicator Fs from the Newmark model as a principal feature parameter, which integrates several influencing aspects, including rock strata, moisture content, and slope gradient. Monte Carlo simulations (MCS) mitigate uncertainty in geotechnical parameters, augmenting Newmark model findings’ reliability. The sampling approach utilizes Newmark model outcomes to identify nonlandslide locations for developing six landslide susceptibility models, namely N_XGB, N_RF, Dn_XGB, Dn_RF, and independent XGBoost and RF models. Comparative investigation reveals that the hybrid models N_XGB and N_RF, which amalgamate Newmark model outputs with machine learning methodologies, have enhanced performance. The N_XGB model attains superior accuracy, exhibiting an AUC of 0.96 in the Jiuzhaigou region, markedly surpassing alternative hybrid machine learning models and physical models, moreover, Fs as a feature exhibits superior predictive accuracy relative to utilizing Newmark displacement. Validation in the Ludian and Luding region results in model accuracies of 0.88 and 0.86, respectively, indicating robust generalizability across various seismic situations. These findings highlight the benefits of integrating physical principles with data-driven methodologies, providing a comprehensive regional landslide risk assessment and management framework.