Scalable XAI-based forecasting of landslide surface velocities from environmental forcings

Forecasting the evolution of slow-moving landslides is a challenge because landslide motion is modulated by hydrometeorological forcing (rainfall, snowmelt, groundwater fluctuations) acting across multiple timescales, resulting in complex and strongly non-linear forcing–response relationships. By leveraging long-term multi-parameter monitoring, AI-based models can help characterise and simulate these dynamics. However, two limitations persist. First, many approaches rely on deep-learning architectures (RNNs, GRUs, LSTMs) that successfully reproduce non-linear dynamics, but do not constrain landslide physics and have limited interpretability and transferability. Second, few AI applications address landslides governed by the combined influence of multiple hydrometeorological drivers. Existing applications remain largely site-specific, relying on tailored predictor sets and local calibration. Addressing these limitations requires interpretable modelling frameworks capable of operating across multiple landslide sites including data-scarce settings.

Here, we introduce a scalable and eXplainable Artificial Intelligence (XAI) modelling framework using eXtreme Gradient Boosting (XGBoost) and based on a set of 248 and physically grounded, non site-specific hydrometeorological predictors computed from net rainfall, effective rainfall, and groundwater level time series. Predictors are designed to represent three complementary aspects of landslide water-related forcing: (i) the hydrological state of the system, (ii) hydrological memory effects, and (iii) short-term hydrological transient processes. To capture multi-timescale hydromechanical dependencies, predictors are computed over multiple time windows ranging from 1 to 90 days. The approach simulates daily landslide velocities, evaluates predictive skill using RMSE and MAE metrics, and provides interpretable and explainable constraints on the predictor influence using features importance ranking and SHAP-based attribution tools.

We evaluate the framework on three slow-moving landslides in France: Séchilienne (fractured miscaschist), Viella (morainic and colluvial deposits), and Villerville (chalk, sand and colluvial deposits ovelying marl substrate) spanning contrasting lithologies, deformation mechanisms and kinematics to demonstrate the scalability of the approach.

The XAI framework accurately reproduces landslide velocity time series across sites and testing periods with small residual errors relative to the amplitude of observed velocity variations (Séchilienne,  0.005-0.015 cm.d-¹ ; Viella, 0.01-0.035 cm.d-¹ ; Villerville, 0.02-0.06 cm.d-¹). The identified predictors per landslide align with contrasting physical processes, including delayed pore-water pressure build-up driven by slow matrix infiltration in impermeable slope material (Villerville) and rapid responses to rainfall in more permeable (Viella) or fractured (Séchilienne) slope materials. Together, these results show that XAI frameworks can recover site-specific landslide behaviour while preserving physical interpretability across diverse settings, and demonstrate one of the first applications of a common model structure and non-site-specific predictor set across multiple distinct landslide case studies.