Rapidly assessing coseismic landslide occurrence using logistic regression model and initial P-wave amplitude

On April 3, 2024, coseismic landslides (CL) triggered by the Hualien earthquake with local magnitude of 7.2 caused significant economic losses and casualties. To better understand the occurrence patterns and factors influencing CL, this study built a multivariate logistic regression model using the CL consists of a total number of 3,191 samples, which can provide the probability of landslide occurrence in a given area. To ensure accurate sampling of non-coseismic landslides (NCL), all polygon areas where CL occurrence existed in slope units were removed, and 3,191 random slope areas were mapped as NCL samples. Causative factors used in analysis include gradient, aspect, elevation and curvature of slope, distances to the earthquake and a fault, the angle between slope aspect and earthquake-to-slope azimuth, lithology. Seismic shacking factors including peak ground acceleration (PGA) and peak ground velocity (PGV) are used as the triggering factors. The CL and NCL samples are assigned the class label values of 1 and 0, respectively. The dataset was split into training (70%) and testing (30%) subsets, with each sample containing 29 features and 1 target class label. To balance complexity and accuracy, stepwise regression based on Akaike Information Criterion (AIC) and multicollinearity control (VIF < 5) were used to select key variables. The model was then developed to predict landslide probabilities in the test set. To determine the optimal classification threshold, the Youden index was calculated from the Receiver Operating Characteristic (ROC) curve. Model performance was evaluated using confusion matrices, with metrics such as accuracy, recall, and F1 score to assess overall effectiveness. Additionally, SHapley's Additive Interpretation (SHAP) was applied to quantify the contributions of individual variables. Model 1 (landslide threshold: 0.4569) demonstrated strong performance, achieving 96.26% accuracy, 97.30% precision, 95.16% recall, and a 96.22% F1-score on the training set, and 97.18% accuracy, 97.38% precision, 96.97% recall, and a 97.18% F1-score on the test set. To enhance interpretability, Model 2 (threshold: 0.4939) excluded variables like area, minimum slope, and slope range. By focusing on key variables, Model 2 reduced overfitting risks and improved prediction reliability, offering more consistent results in new regions or emergency scenarios. Despite a slight drop in performance, Model 2 maintained high accuracy (95.28% training, 95.82% test) and reliable metrics across precision, recall, and F1-scores. Partial correlation plots (PDP) and boxplots confirmed its enhanced reliability in predicted probabilities, showing improved consistency compared to Model 1. To further enhance disaster response, the study incorporated an early warning system using the peak displacement of initial P-wave (Pd). When the on-site vertical displacement exceeds 0.12 cm, a CL alarm is issued, providing a lead time of 6 to 14 seconds before peak ground shaking. This integrated approach bridges early warning and post-disaster assessment, enhancing resilience and preparedness in seismically active regions.

This study proposes a framework that integrates early warning and post-disaster assessment for coseismic landslides (CL). Combining logistic regression with P-wave information, the system enables timely alerts and effective damage evaluation, bridging hazard detection and recovery planning to enhance disaster resilience.