In the complex realm of geohazards, earthquake-triggered landslides stand out as a formidable secondary risk, frequently exacerbating the devastation wrought by seismic events. Responsible for the loss of tens of thousands of lives and causing billions of dollars in damage annually, these landslides pose a significant challenge to disaster management and response efforts worldwide. Traditionally, the rapid identification and mapping of landslide occurrences following earthquakes have been hindered by the limitations of remote sensing technologies, which depend heavily on clear atmospheric conditions and cloud-free satellite imagery—resources often unavailable during the crucial hours immediately after seismic shocks.
Addressing these formidable challenges, a pioneering team led by Professor Xuanmei Fan at the State Key Laboratory of Geohazard Prevention and Geoenvironment Protection, Chengdu University of Technology, has developed the most extensive and rigorously validated global database of coseismic landslides to date. This comprehensive catalog encompasses nearly 400,000 accurately mapped landslide events across 38 major earthquakes around the globe, spanning from the 1970s to the present. The meticulous validation process integrates advanced remote sensing change detection techniques with expert manual verification to ensure precise delineation and temporal accuracy of landslide extents, thus providing an unprecedented repository of high-fidelity geospatial data for the scientific community.
Building upon this vast inventory, the research team constructed an intricate suite of seventeen diverse indicators encompassing topographical, geo-ecological, hydrological, and seismological parameters. Their sophisticated spatial and statistical analyses underscore the dominant roles of peak ground acceleration, slope gradient, and underlying lithology in governing global landslide susceptibility. Secondary influences emerge from terrain relief and roughness, highlighting the complex interplay of earth surface factors controlling landslide initiation. By segmenting the inventory into tectonically significant regions—the Circum-Pacific and Alpine-Himalayan belts—and further stratifying each into simplified climate zones (cold, temperate, and equatorial), the study elucidates distinctive regional controls, offering a scientifically grounded framework for tailoring predictive models to localized environmental conditions.
At the forefront of predictive geohazard modeling, the team harnessed the power of deep learning, developing a novel architecture based on a multi-scale fully convolutional regression network augmented with channel-spatial attention modules. This innovative design empowers the model to extract and emphasize the most discriminative features from multi-dimensional geospatial inputs, facilitating rapid inference of landslide occurrence probability across heterogeneous terrains. Two distinct model configurations were trained and evaluated: regional models fine-tuned to specific climatic and tectonic zones, yielding heightened accuracy in data-rich regions; and a global model leveraging the full spectrum of worldwide events to maintain robustness in data-sparse cold climates.
The model’s performance was rigorously validated using a leave-one-out cross-validation approach encompassing all 38 earthquake datasets, achieving consistent spatial predictive accuracy surpassing 82%. This efficiency is underscored by the model’s ability to process each scenario in under one minute on high-performance Tesla V100 GPUs, illustrating its suitability for near-real-time applications. Such rapid predictive capability marks a transformative advance over traditional susceptibility mapping approaches, offering critical temporal advantages for post-earthquake response.
Professor Fan highlights the operational potential of this deep learning framework, emphasizing its capacity to generate near-real-time landslide probability maps immediately after seismic events without relying on pre-existing labels or ground-truth data. This immediacy is invaluable for first responders and hazard managers, enabling them to pinpoint and prioritize at-risk regions during the agonizing early hours following an earthquake, where timely intervention can significantly reduce casualties and damages.
Complementing this perspective, co-author Professor John Jansen from the Czech Academy of Sciences underscores the practical implications for disaster risk management. By integrating model outputs with overlays of population density and critical infrastructure, emergency planners can swiftly identify vulnerable communities and assets exposed to landslide hazards, facilitating resource allocation and mitigation strategies well ahead of the availability of high-resolution imagery or ground surveys.
Looking toward the future, the research ensemble aims to expand the model’s predictive scope by incorporating additional environmental drivers such as rainfall forecasts and aftershock sequences. This evolution aspires to culminate in a comprehensive multi-hazard early warning system capable of dynamically assessing cascading risks originating from seismic events and their hydrometeorological aftermath. The team is also exploring deployment strategies leveraging cloud computing platforms, and integrating data streams from unmanned aerial vehicles and ground-based sensors, to further compress the latency from earthquake detection to actionable landslide hazard prediction.
Co-author Professor Hakan Tanyas from the University of Twente characterizes this breakthrough as a paradigm shift, pivoting away from retrospective susceptibility or hazard mapping toward proactive, real-time predictive analytics for earthquake-induced landslides. This approach promises to substantially enhance decision-support capabilities at a global scale, transforming seismic hazard management into a more anticipatory and responsive discipline.
In summation, this research represents a landmark convergence of a uniquely comprehensive global landslide database, rigorous mechanistic understanding of landslide triggers, and cutting-edge deep learning methodologies. The synergistic integration of these components lays the technological and scientific foundation for next-generation geohazard risk reduction tools. These advancements not only deepen our understanding of complex earth system processes but also empower societies worldwide to prepare for and mitigate the destructive cascading effects of major earthquakes effectively.
Subject of Research: Earthquake-triggered landslides and their rapid prediction using deep learning methodologies.
Article Title: Deep learning can predict global earthquake-triggered landslides.
Web References: http://dx.doi.org/10.1093/nsr/nwaf179
Image Credits: ©Science China Press
Keywords: Earthquake-triggered landslides, deep learning, coseismic landslide inventory, fully convolutional regression network, channel-spatial attention, geohazard prediction, seismic hazard zones, global landslide database, terrain analysis, multi-hazard early warning system
