Assessing the Impact of Strong Ground Motion Intensity Measures on Earthquake-induced Landslide Susceptibility in the Indian Himalayan Region Using Machine Learning

Abstract Earthquake-induced landslides (EQILs) pose a major hazard in seismically active mountain regions; however, the relative influence of different strong ground-motion intensity measures (IMs) on landslide susceptibility remains poorly constrained. This study evaluates the role of key IMs in controlling earthquake-induced landslide susceptibility (EQ-LS) in the Sikkim Himalaya, India, using the 2011 Sikkim earthquake (Mw 6.9) as a case study. For this, a comprehensive inventory of 1,120 coseismic landslides was first compiled within the epicentral region. Seismic, topographic, and hydrological conditioning factors were integrated to develop EQ-LS models using six machine-learning algorithms. Strong ground-motion intensity measures (IMs), Peak Ground Acceleration (PGA), Peak Ground Velocity (PGV), and Arias Intensity (Ia) were then tested individually and in combination to assess their influence on model performance. Among the evaluated models, the Extremely Randomized Trees (ET) achieved the highest predictive accuracy (0.84) and area under the receiver operating characteristic curve (AUC = 0.92). Models incorporating combined IMs consistently outperformed those based on single parameters, with the PGV–Ia combination yielding the best performance (AUC = 0.93). Feature importance analysis using SHapley Additive exPlanations (SHAP) indicates that velocity-based ground-motion metrics exert a dominant control on landslide initiation. These findings underscore the critical role of PGV in governing EQ-LS. The proposed machine-learning framework offers a robust and transferable approach for EQ-LS mapping in seismically active mountainous regions, with direct implications for improved landslide hazard assessment and disaster-risk reduction planning. Graphical Abstract This graphical abstract illustrates a machine-learning (ML)–based framework for assessing earthquake-induced landslide susceptibility (EQ-LS) in the Sikkim Himalaya, using the 2011 Mw 6.9 Sikkim earthquake as a case study. A comprehensive inventory of 1,120 coseismic landslides was analysed together with key landslide-conditioning factors, including topographic, geological, hydrological, and seismic variables. Six ML algorithms—Random Forest (RF), XGBoost (XGB), Extremely Randomized Trees (ET), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Linear Regression (LR) were initially trained using Peak Ground Acceleration (PGA) as the sole seismic input. The dataset was partitioned into 70% for training and 30% for testing. Model performance was evaluated using multiple metrics, including accuracy, precision, recall, F1-score, Cohen’s kappa, Matthews correlation coefficient (MCC), and ROC-AUC, which identified the ET model as the best-performing algorithm. A key novelty of this study is the systematic evaluation of different combinations of seismic intensity measures (IMs) for predicting earthquake-induced landslides (EQILs). After selecting the ET model based on PGA alone, various IM combinations—Peak Ground Acceleration (PGA), Peak Ground Velocity (PGV), and Arias Intensity (Ia) were incorporated into the model. The results demonstrate that multi-parameter IM combinations significantly improve predictive performance, with the PGV–Ia combination achieving the highest skill (AUC = 0.93). This selection was supported by consistent improvements across multiple evaluation metrics and validation against the observed landslide inventory. Feature importance analysis using SHapley Additive exPlanations (SHAP) indicates that elevation, slope angle, and aspect are the dominant controls on EQ-LS, followed closely by velocity- and energy-based seismic IMs, particularly PGV and Ia. Overall, the proposed framework offers a robust and transferable approach for EQ-LS mapping, highlighting the importance of multi-parametric seismic inputs for improved landslide hazard prediction in seismically active mountainous regions.