Ground-shaking-informed regime partitioning for characterizing earthquake-induced secondary hazards

Abstract

Earthquake-induced landslides and liquefaction often occur within the same seismic event but may be controlled by different combinations of ground shaking, topography, hydrology, site conditions, and tectonic setting. This study develops an interpretable remote-sensing framework to map and compare these secondary hazards after the 2025 Dingri Ms 6.8 earthquake in the southern Tibetan Plateau. Multi-source optical and SAR indicators, seismic variables, and geo-environmental factors were integrated using ensemble machine-learning models. A peak ground acceleration (PGA)-guided multivariate Gaussian mixture model was used to stratify hazard samples into PGA-associated statistical regimes, and regime-wise XGBoost models were interpreted using SHAP. The results show that landslide predictions are mainly associated with shaking intensity and topographic conditions, whereas liquefaction predictions show stronger regime-dependent associations with site conditions, hydrological proximity, geomorphic setting, and near-fault effects. The proposed framework provides a practical way to compare model-inferred feature associations across different shaking and environmental backgrounds. The results highlight that coseismic secondary hazards in high-altitude tectonic regions are not governed by a single uniform relationship, but by spatially heterogeneous combinations of seismic and environmental conditions associated with hazard occurrence.