Hydrometeorological features ans surface velocity measurements for the Séchilienne landslide

The Séchilienne datasets are based on a multi-sensor monitoring network installed and maintained by the French Landslide Observatory OMIV (Observatoire Multidisciplinaire des Instabilités de Versants; ano-omiv.cnrs.fr). They include daily surface-displacement measurements from five monitoring plots (tacheometric targets and extensometers; cm d⁻¹) and local precipitation records (mm) acquired at the Mont Sec weather station, complemented by the Luitel station to ensure temporal completeness. Effective rainfall (mm) is estimated as precipitation minus potential evapotranspiration, with evapotranspiration computed using the Oudin formulation (Oudin et al., 2005). Daily temperature (°C) required for this calculation is derived from the two weather stations, and the maximum soil available water capacity (SAWCMAX) is set to 100 mm. Because groundwater-level observations are not available at this site, groundwater dynamics are reconstructed using rainfall–discharge models that conceptualise hillslope hydrology, yielding two synthetic groundwater proxies: a shallow, rapidly responding reservoir (GWLI) and a deeper, more inertial reservoir (GWLM) (Béjean-Maillard et al., 2025).

Raw displacement records are processed using a sequential data-preparation workflow to mitigate instrumental errors that can affect model performance and interpretation. This workflow includes position-bias correction, reconstruction of cumulative displacement and velocity by differencing, outlier detection, and adaptive smoothing. In addition, to ensure stationarity for XAI modelling, processed velocity series are decomposed using a multiplicative decomposition to separate trend and transient components; the model is trained on the dimensionless transient component provided in the file. Technical details of the pre-processing steps are provided in the associated publication.

Hydrometeorological predictors computed from the forcing series are designed to be physics-informed and non-site-specific. They represent three complementary aspects of water-driven forcing: hydrological state (e.g. saturation versus dryness), hydrological memory (e.g. cumulative rainfall and groundwater trends/extrema), and short-term hydrological transients (e.g. intense rainfall events and rapid groundwater rises or drawdown). Predictors are computed over multiple time windows (1–90 days), enabling the model to resolve short-, intermediate- and long-term hydromechanical responses. To account for delayed kinematic responses, groundwater descriptors are additionally computed with site-specific time lags. This results in a set of 404 predictors derived from four forcing series (R, ER, GWLI and GWLM). No-data cells occur when values from the underlying forcing series are missing; the resulting gap length increases with the predictor time window (i.e. longer windows produce longer no-data segments). Predictor names follow a consistent convention that encodes the data source, descriptor type and time window (and lag, where applicable). For example, cumulative rainfall over 60 days is denoted R_60, and a 10-day groundwater-level difference computed from GWLI with a 5-day lag is denoted GWLI_l5_diff10. Full predictor definitions are provided in the associated publication. The code used to generate the predictor set is available from the certified IN2P3 GitLab repository: [to be completed]. This work contributes to a PhD thesis supported by the Université Marie et Louis Pasteur (Doctoral School ED554) – Béjean-Maillard O. (2025).