A Unified Deep Learning Framework for Rapid Global Prediction of Coseismic Landslides

Earthquake-triggered landslides are among the most destructive secondary seismic hazards, yet their rapid prediction at global scale remains elusive due to the limitations of existing physical and statistical models. Current approaches typically depend on regional inventories, simplified assumptions, or retrospective calibration, preventing timely and reliable assessments immediately after large earthquakes. To address this gap, we compiled the largest global database to date of ~400,000 coseismic landslides from 38 major earthquakes spanning diverse tectonic and climatic settings. Using this unified dataset, we developed a multi-scale fully convolutional deep-learning framework capable of predicting coseismic landslide probability worldwide with no prior local labels.

The model integrates 14 primary control indicators, representing topography, geo-ecology, hydrology and seismology, and learns nonlinear relationships governing slope failure across global environments. Independent testing shows that the global model achieves an average AUC of ~0.83 and spatial accuracy of ~0.77, while regional models trained within specific environmental domains achieve slightly higher performance. The predictions successfully reproduce both the extent and spatial pattern of landslides for events such as the 2015 Gorkha, 2016 Kaikoura, 2021 Nippes, 2022 Luding and 2002 Denali earthquakes. Sensitivity analyses further demonstrate that model performance is robust to inventory uncertainty but strongly influenced by the quality of input seismic and fault data.

Our framework predicts landslide probability for a new earthquake in less than one minute, enabling actionable early hazard intelligence well before cloud-free satellite imagery becomes available. A hypothetical Mw 7.5 earthquake scenario in Sichuan, China illustrates that rapid prediction can identify high-impact areas and populations exposed to landslide cascades within seconds. This study establishes the first globally scalable and operational deep-learning model for earthquake-triggered landslide prediction, offering transformative potential for rapid hazard response, seismic risk management, and global multi-hazard preparedness.