Global database and prediction model of earthquake-triggered landslides

Earthquake-induced landslide (EQIL) models seek to map where landslide is likely to occur during earthquakes from ground-motion measures and environmental controls. Yet most models are trained almost exclusively on landslide-triggering earthquakes, encouraging overfitting to event-specific signatures, weakening transferability, and blurring how ground motion and predisposition jointly govern failure. Here we address both limitations by compiling a new global EQIL database that explicitly includes strong non-triggering earthquakes, and by developing a neural-network framework designed to learn transferable, mechanism-consistent controls on landslide occurrence. Our database extends existing public inventories by harmonizing 44 previously published landslide-triggering earthquakes and adding 24 newly mapped triggering events, alongside 44 strong earthquakes for which no widespread landslides is mapped. These non-triggering earthquakes provide event-level negative constraints that are rarely available in EQIL modelling. For each non-triggering event, we conducted a multi-temporal audit using 3-m PlanetScope imagery; any missed failures are expected to be sporadic and very small, and do not alter the event-level classification. Using the combined catalogue, we train pixel-level probabilistic models conditioned on ground motion and environmental covariates. Transferability is evaluated via leave-one-event-out cross-validation and an independent multi-continent test set spanning diverse climates and faulting styles. Incorporating non-triggering earthquakes markedly improves cross-event performance (mean ROC–AUC increases from 0.873 to 0.914) and reduces event-specific errors, yielding more robust probabilistic maps of landslide spatial patterns. To interpret learned controls, we apply SHAP-based explain ability supported by complementary statistical summaries. Terrain and material properties (for example slope/relief and lithology) exert strong inhibitory influences that keep predicted probabilities low even under high peak ground acceleration (PGA), whereas PGA acts primarily as a conditional amplifier where predisposition is high. Overall, explicitly modelling counterfactual non-triggering earthquakes offers a practical route to more accurate, transferable EQIL mapping and clearer insight into why strong earthquakes do—or do not—produce widespread landslides.