Geophysical and Geo-AI based evaluation and modelling of earthquake induced landslides

Earthquake-induced landslides (EQIL) are among the most severe cascading hazards during seismic events, often resulting in significant loss of life and extensive damage to infrastructure. To minimize these impacts, it is essential to accurately understand and model EQIL. However, the influence of earthquake ground motion on EQIL is not fully understood due to limited spatially dense full waveform ground motion data, geotechnical information, and detailed landslide inventories. This data scarcity complicates the modelling of EQIL hazards, hampering the development of robust frameworks for both deterministic and probabilistic hazard assessments.
A key aspect in understanding the role of ground motion on landslide occurrence involves having spatially varying ground motion waveforms for each slope to assess seismic perturbation during an earthquake. However, this information remains unavailable due to the scarcity of dense seismic networks. Ground motion simulations, on the other hand, solve wave equations in both space and time, offering detailed shaking estimates that could explain landslide genesis. Despite this, geomorphologists have not extensively explored this approach, as limited knowledge exists regarding which ground motion parameters, whether synthetic or real, best explain coseismic landslide distributions. Furthermore, these simulations consider ground motion amplification due to surface and subsurface geometry, known as topographic amplification. This phenomenon increases shaking on hilltops, potentially destabilizing otherwise stable slopes. However, its contribution to coseismic landslides remains unquantified in strong earthquake-affected landscapes, despite its possible significance.
Historically, physics-based models were the only tools capable of comprehensively describing the spatio-temporal dynamics of landslides. Data-driven models, by contrast, have primarily focused on predicting landslide susceptibility and identifying areas at risk, as well as estimating landslide timing through early-warning systems. While some studies have explored joint spatio-temporal models, attempts to model landslide size (in terms of area or volume) over space and time remain limited.
In the field of data-driven modelling, a crucial distinction exists between statistical models, which prioritize interpretability, and machine learning models, which focus on maximizing predictive performance. This dichotomy is common across various scientific domains. In natural hazard prediction, researchers frequently face the challenge of balancing interpretability with performance. While high-performance models are beneficial for disaster risk reduction, understanding how these models function is critical for establishing trust in them. This underscores the importance of developing explainable machine and deep learning models in landslide hazard assessments.
One limitation of current deep learning models is their tendency to simplify ground motion time series into peak scalar values, disregarding the contributions of phase, amplitude, and their spatial and temporal variations. This simplification hinders the models’ ability to fully capture the ground motion waveforms and their relationship to landslide occurrences.
Another challenge in data-driven models is the typically deterministic nature of landslide hazard estimates, which often focus only on susceptibility and intensity. When both elements are modeled, they are usually treated separately, with frequency rarely examined. Since frequency and intensity are interdependent—larger events occur less frequently—a probabilistic framework is more appropriate for their estimation. However, due to a lack of multi-temporal inventories and joint statistical models, developing a unified probabilistic hazard model has been difficult.
Despite these challenges, regional-scale landslide prediction continues to rely on data-driven models, which often use proxy variables in place of direct geotechnical data. These proxy variables have been effective, as obtaining explicit geotechnical data over large areas is difficult.
To address these issues, we developed several analytical frameworks and methods, focusing on the Gorkha earthquake (M7.8, April 25, 2015). We combined Synthetic Aperture Radar Interferometry (InSAR) with strong motion data to select the optimal ground motion simulation from fault process forward modelling. With this data, we extracted a comprehensive set of intensity parameters to analyze coseismic landslide occurrences and examined how topographic amplification influences landslide spatial distribution.
Recent advances in data-driven models now allow for the modelling of all three key components: the location, timing, and size of landslides. Our research introduced a method for assessing landslide hazards by jointly modelling landslide occurrences and area density in both space and time. Utilizing a spatio-temporal landslide database for Nepal, impacted by the Gorkha earthquake, our deep-learning model—an Ensemble Neural Network—aggregated landslide occurrences and densities over 1×1 km units, classified and regressed against a 30-meter lattice. We incorporated both predisposing and triggering factors with a temporal resolution of approximately six months.
We also demonstrated the utility of explainable artificial intelligence (XAI) in landslide susceptibility modelling. Our XAI model provided valuable insights into model design and querying potential, while a deep-learning model incorporating full seismic waveforms was compared with one using scalar intensity parameters to better understand the waveform’s role in landslide genesis.
For probabilistic hazard assessment, we presented a unified model that estimates landslide hazards at the slope unit level. By integrating deep learning with extreme-value theory, we analyzed 30 years of rainfall-triggered landslide data in Nepal to assess hazards for various return periods. We also investigated landslide hazards under different climate change scenarios, equivalent to analyzing hazards under different ground motion scenarios.
Moreover, we applied a Physics Informed Neural Network (PINN) approach, incorporating an intermediate constraint to solve for permanent deformation characteristics using Newmark slope stability methods. This approach allowed the neural network to estimate geotechnical parameters from common proxy variables, minimizing a loss function based on coseismic landslide inventory, and estimating permanent deformation and landslide susceptibility.
Our research highlights that, contrary to the common focus on peak ground acceleration, velocity, or displacement, other shaking parameters—such as total displacement, frequency content, and shaking duration—play a crucial role in landslide occurrences. Topographic amplification was found to contribute to 6-17% of coseismic landslides, with its influence becoming more pronounced at greater distances from the rupture zone. This suggests that the impact of topographic amplification on landslide occurrence may be overestimated in strong earthquakes.
In joint modelling for susceptibility and area density, our model demonstrated strong performance (AUC = 0.93 for susceptibility and Pearson r = 0.93 for density prediction), providing an integrated hazard modelling framework. Explainable modelling also showed high performance (AUC = 0.89) and interpretability regarding the roles of various controlling factors. Incorporating full waveform ground motion data improved predictive capacity by 16%, particularly on gentle hillslopes with low ground shaking, highlighting the limitations of using single-intensity measures.
Our probabilistic framework achieved an accuracy of 0.78 and an AUC of 0.86, modelling landslide hazards comprehensively under different scenarios. Climate change scenarios indicated a potential doubling of landslide hazards in lower Himalayan regions, with minimal change in the middle Himalayas and a slight decrease in the upper Himalayas.
The integration of physics via the PINN approach showed excellent predictive performance for susceptibility and produced regional-scale geotechnical property maps. This architecture could significantly enhance coseismic landslide prediction and, with further validation, enable near-real-time predictions using PINN-based methods.
In conclusion, this research advances the understanding of ground motion and topographic amplification in landslide hazards and introduces several modelling frameworks using deep learning in both deterministic and probabilistic contexts. Future research should focus on integrating probabilistic hazard modelling with physicsbased approaches and full waveform simulations, as well as extending these models to other earthquake-affected regions for improved hazard assessment.