Predicting Landslides After Earthquakes With AI

Jim Crocker
12th August, 2025

Predicting Landslides After Earthquakes With AI

To construct the N_XGBoost susceptibility model, nine key environmental variables were selected via correlation analysis, comprising the digital elevation model (a), distance to road (b), distance to parallel fault (c), slope (d), distance to river (e), peak ground acceleration (f), slope direction (g), epicentral distance (h), and curvature (i).

Image adapted from: Zhang et al. / CC BY (Source)

Key Findings

  • Researchers at Henan University developed a new hybrid model for the Jiuzhaigou region, combining a physics-based model with machine learning to better predict earthquake-triggered landslides
  • This new N_XGB model significantly improved prediction accuracy for landslides in Jiuzhaigou, outperforming other methods by integrating a key slope stability factor
  • Crucially, the N_XGB model also proved highly effective in predicting landslides in other diverse regions, making it a robust tool for wider disaster risk management
Landslides triggered by earthquakes, known as coseismic landslides, are among the most dangerous natural disasters in mountainous regions. Accurately predicting where and when these landslides might occur is essential for protecting lives and infrastructure. However, developing models that can reliably predict landslide susceptibility – the likelihood of an area experiencing a landslide – has been a significant challenge due to the complex and varied nature of the factors involved. Recent research[1] from Henan University, Chinese Government & Industry, has developed a new approach to improve the assessment of coseismic landslide susceptibility. The study focuses on integrating a physical model, called the Newmark model, with advanced machine learning techniques. The Newmark model is a widely used tool in engineering geology that estimates the displacement of a slope during an earthquake by considering factors such as the strength of the rock, its moisture content, and the steepness of the slope. The output of this model, a comprehensive stability indicator called Fs, was used as a key input for the machine learning part of the new framework. To make the Newmark model's findings more reliable, especially given the inherent uncertainties in geological data, the researchers employed Monte Carlo simulations (MCS). This technique involves running the model many times with randomly varied inputs within a plausible range, allowing for an assessment of how these uncertainties might affect the outcome. This helps to provide a more robust and realistic picture of slope stability. The study then used the outputs from the Newmark model, specifically the Fs indicator, to train and develop six different landslide susceptibility models. These included hybrid models combining Newmark outputs with machine learning algorithms like XGBoost (eXtreme Gradient Boosting) and Random Forest (RF), as well as independent XGBoost and RF models. XGBoost and Random Forest are powerful machine learning algorithms known for their ability to identify complex patterns in large datasets. The research found that the hybrid models, particularly the N_XGB model which combined the Newmark model's Fs indicator with XGBoost, showed significantly enhanced performance. For the Jiuzhaigou region, which served as the primary study area, the N_XGB model achieved an impressive accuracy, measured by an AUC (Area Under the Receiver Operating Characteristic Curve) of 0.96. AUC is a common metric for evaluating classification models, with values closer to 1 indicating higher accuracy. This performance notably surpassed other hybrid models and purely physical models. The study also highlighted that using Fs as a feature provided superior predictive accuracy compared to using Newmark displacement directly. The choice of Jiuzhaigou as a study area is particularly relevant because it was the site of a major Mw6.5 earthquake in 2017, which triggered a vast number of coseismic landslides. Previous detailed investigations, such as one conducted through field surveys and high-resolution remote sensing, established a comprehensive inventory of over 9,400 landslides in the region[2]. This type of detailed data, identifying specific landslide locations and their characteristics (like altitude, slope, and proximity to roads or faults), is crucial for training and validating predictive models like the one developed in the current study. The factors identified in the Jiuzhaigou landslide inventory[2], such as altitude, slope, and lithology (rock type), are precisely the types of influencing factors that the Newmark model's Fs indicator integrates. A key challenge in landslide susceptibility modeling is the "spatial heterogeneity" of influencing factors, meaning that the importance of different factors can vary significantly from one region to another. This often leads to models having poor "generalizability," meaning they perform well in the area they were trained on but poorly when applied to new, different regions[3]. The new study directly addresses this by testing the N_XGB model's generalizability. When applied to other seismic areas, Ludian and Luding, the model maintained robust accuracies of 0.88 and 0.86 respectively. This demonstrates that the integrated physical-machine learning approach can overcome some of the limitations of models that struggle to generalize across different landscapes, as discussed in earlier research[3]. The ability of the N_XGB model to perform well in new regions suggests that incorporating physical principles (via the Newmark model) provides a more fundamental understanding of landslide mechanics, which makes the model more adaptable than purely data-driven approaches that might overfit to specific regional characteristics. In essence, this research builds upon the detailed understanding of specific landslide events, like the Jiuzhaigou earthquake[2], and the identified challenges in model generalizability[3]. By combining the physics-based Newmark model with the pattern-recognition capabilities of machine learning, it creates a more comprehensive and robust framework for assessing landslide risk. This integrated approach offers a significant step forward in predicting potential landslide hazard areas and provides a valuable tool for regional risk assessment and management.

EnvironmentSustainability

References

Main Study

1) Evaluation of coseismic landslide susceptibility by combining Newmark model and XGBoost algorithm

Published 11th August, 2025

https://doi.org/10.1371/journal.pone.0328705

Related Studies

2) An essential update on the inventory of landslides triggered by the Jiuzhaigou Mw6.5 earthquake in China on 8 August 2017, with their spatial distribution analyses.

https://doi.org/10.1016/j.heliyon.2024.e24787

3) Insights into geospatial heterogeneity of landslide susceptibility based on the SHAP-XGBoost model.

https://doi.org/10.1016/j.jenvman.2023.117357


Related Articles

An unhandled error has occurred. Reload 🗙