image: (A) Coseismic landslide inventories within the Alpine-Himalayan and Circum-Pacific belts. (B) The enlarged view of highlighted 38 earthquakes. The five Köppen-Geiger climate zones are here simplified into three for region division: cold (polar plus snow); temperate (warm plus arid); and equatorial. Beachballs signify primary seismic rupture dynamics.
Credit: ©Science China Press
Earthquake-triggered landslides are a major secondary hazard, often responsible for tens of thousands of fatalities and billions of dollars in economic losses each year. Yet rapid identification of where landslides will occur remains a challenge: remote sensing can map past failures but relies on cloud-free imagery that may not be available during the critical hours after a quake.
To overcome these limitations, Prof. Xuanmei Fan’s team at State Key Laboratory of Geohazard Prevention and Geoenvironment Protection (Chengdu University of Technology) compiled the most comprehensive global database of coseismic landslides to date, 398,698 mapped events across 38 major earthquakes since the 1970s. Each landslide polygon was rigorously validated integrating remote sensing-based change detection and expert manual review to ensure boundary accuracy and temporal consistency.
Building on this inventory, the researchers engineered a suite of 17 topographical, geo-ecological, hydrological, and seismological indicators. Spatial and statistical analysis revealed that peak ground acceleration, slope, and lithology are the key drivers of coseismic landsliding in global scale, with secondary influence from relief and terrain roughness. Moreover, they partitioned the global inventory into the Circum-Pacific and Alpine-Himalayan earthquake belts, each further divided into cold, temperate, and equatorial climate zones, revealing distinct dominant controls in each region and providing a basis for optimizing regional model parameters.
Another core innovation is the prediction model based on multi-scale fully convolutional regression network enhanced by channel-spatial attention modules. This architecture enables to generate and focus on the most discriminative features via multi-dimensional inputs, achieving rapid inference of landslide occurrence. In this work, two model variants were trained and tested.
(a) Regional models, tuned to local climatic and tectonic zones (e.g., equatorial and. temperate regions), deliver the highest accuracy where sufficient training samples supplied.
(b) The global model ensures robust performance in data-sparse, cold regions by leveraging the full diversity of global events.
Using a leave-one-out cross-validation on all 38 events, the framework consistently achieved spatial accuracy exceeding 82% and processed each scenario in less than one minute on two Tesla V100 GPUs.
“Our deep learning model enable to provide near-real-time probability maps of landslide occurrence immediately following an earthquake without any prior labels” says Prof. Fan. “This capability can guide first responders and hazard managers to the most at-risk areas within the crucial early hours.”
Prof. John Jansen, co-author from Czech Academy of Sciences, emphasizes the operational potential: “By integrating our model outputs with population and infrastructure overlays, we can estimate exposed at-risk communities in seconds, far before high-resolution imagery becomes available.”
Looking ahead, the team plans to incorporate rainfall forecasts and aftershock sequences into the model, moving toward a unified “multi-hazard” early warning system. They are also exploring deployment via cloud-based platforms and integration with unmanned aerial and ground sensors to further shrink the time between earthquake occurrence and hazard prediction.
“This work represents a paradigm shift from retrospective susceptibility mapping to proactive, predictive prediction of earthquake-triggered landslides,” says Prof. Hakan Tanyas, co-author from University of Twente. “It opens the door to real-time, global-scale decision support in seismic hazard zones.”
With its combination of a unique global landslide database, rigorous mechanistic analysis, and cutting-edge deep learning, this research lays the foundation for next-generation geohazard risk reduction tools, transforming how societies prepare for and respond to the destructive aftermath of major earthquakes.
###
See the article:
Deep learning can predict global earthquake-triggered landslides
https://doi.org/10.1093/nsr/nwaf179
Journal
National Science Review