Probabilistic Simulation of Monthly Landslide Velocity Under Hydro-meteorological Variability

Hydrometeorological forcing (rainfall, snowmelt, groundwater fluctuations) acts across multiple timescales and is a primary driver of surface velocity dynamics in slow-moving landslides. Many studies use trained AI-based models to simulate daily-to-monthly velocities over validation periods defined by specific historical hydrometeoroligical contexts. Although these models achieve accurate predictive skill, they are typically deterministic and therefore provide limited insight into the range of plausibly velocity responses under alternative, yet realistic, forcing conditions.

To address this gap, we introduce a probabilistic framework built around two axes. Forcing variability is represented by generating 500 plausible meteorological time series using a modified Richardson-type weather generator (rainfall and air temperature). These series are then propagated through a transfer-function hydrological model to simulate groundwater-level variability driven by generated effective rainfall. Second, daily velocities are simulated using a trained XGBoost model based on a set of hydrometeorological predictors. The resulting ensemble is summarised as monthly velocity distributions over a one-year horizon, thereby capturing distinct dynamics across a full hydrological cycle. Distributional performance is evaluated using the Prediction Interval Coverage Probability (PICP) and the Mean Interval Score (MIS).

We evaluate the framework on three slow-moving landslides spanning contrasting lithologies, deformation mechanisms and kinematics: Viella (morainic and colluvial deposits ; France), Villerville (chalk, sand and colluvial deposits ovelying marl substrate ; France), and Vallcebre (clayey siltstone and colluvial debris overlying limestone substrate), to demonstrate the scalability of the approach.

The modified Richardson-type generator reproduces key statistical properties of historical meteorological records. Calibrated groundwater models capture the main dynamics of groundwater fluctuations, with R² values of 0.84 (Viella), 0.76 (Villerville) and 0.53 (Vallcebre). The simulated monthly velocity distributions exhibit clear seasonality, with more contrasted annual cycles at Viella and Villerville, consistent with site-specific hydrogeological behaviour. On average, prediction intervals encompass a substantial fraction of observed monthly velocities (mean PICP: 53% for Viella, 40% for Villerville and 76% for Vallcebre), with strong variability across months. Remaining discrepancies mainly reflect data availability, limitations in groundwater simulations, and constraints in the learned forcing–velocity relationships within the XGBoost model, highlighting priorities for further methodological improvements. Overall, the proposed framework provides a first practical tool to quantify the range of probable landslide-velocity responses under multiple plausible hydro-meteorological scenarios.