Recently, the team led by Professor Fan Xuanmei at Chengdu University of Technology has made important progress in the intelligent prediction of globally earthquake-induced geological hazards. The related results, titled"Deep learning can predict global earthquake-triggered landslides"were published in the international top journalNational Science Review(NSR, impact factor 16.3).

01

02

Global earthquake-induced landslide database

To achieve truly global prediction, the research team selected 38 typical events from earthquakes of magnitude 6 or above worldwide since 1970, and, combining intelligent remote sensing identification with manual verification, interpreted nearly 400,000 landslide samples, constructing the largest and highest-qualityglobal earthquake-induced landslide databaseto date (Figure 1).

Furthermore, earthquake events were divided into the Circum-Pacific and Alpine-Himalayan seismic belts, and three major climate zones: cold, temperate, and equatorial, in order to enhance the model's generalization capability under different regional geological environmental conditions.

Universal controlling factors of earthquake-induced landslides at the global scale

Based on the understanding of the mechanisms of strong-earthquake-induced landslides, this study analyzed 17 influencing factors and found that peak ground acceleration (PGA), slope, and lithology are the dominant controlling factors of earthquake-induced landslides worldwide: PGA directly reflects the intensity of seismic energy release, slope determines the stability threshold of the slope itself, and lithology affects the physical and mechanical properties of the rock and soil mass.

It is noteworthy that the landslide controlling factors in different seismic belt-climate zones show significant spatial differentiation (Figure 2): in cold regions, PGA and distance to fault are the dominant factors; in equatorial regions, terrain-ground motion coupling dominates landslide triggering; in temperate regions, PGA-slope-lithology present a态势 of parallel core factors. This regional differentiationreflects the differentiated influence of the synergistic action of geological background and climatic exogenic forces on landslide dynamic processes, breaks through the "one-size-fits-all" limitation of traditional prediction models, and provides pioneering insights for zonal optimization of the model.

Minute-level inference for global earthquake-induced landslide prediction

This study innovatively designed a multi-scale fully convolutional regression network, combined with channel-spatial attention modules, to achieve automatic extraction, focusing, and fusion of key features of induced landslides. The model adopts a"global-regional dual-track"deployment strategy (Figures 3 and 4): in regions with abundant events and data such as the equatorial and temperate zones, regional models are used to fully optimize parameters through regionalized training and obtain higher prediction accuracy; while in cold-region scenarios lacking events and samples, the global model is used, which, with massive and diverse training samples, can guarantee a robust prediction lower bound and effectively avoid the risk of overfitting.

Compared with traditional mechanistic and statistical models, this method has made significant breakthroughs in both accuracy and computational efficiency, with an averageprediction accuracy of 82%, about 20% higher than existing international models,and computation time shortened from several days to less than 1 minute, achieving near-real-time prediction of earthquake-induced geological hazards.

03