Nowcasting the movement of the deep-seated Reissenschuh landslide based on soil-vegetation-atmosphere transfer modelling and machine learning

Monitoring deep-seated landslides is an important task to prevent impacts on society. Their movement can be highly variable in space and time, ranging from millimetres to metres per year. In the Alps, their driving factors are typically prolonged rainfall events and snow melt, causing a rise of pore water pressure and reduced shear strength and subsequently higher movement rates. Nowcasting the movement of deep-seated landslides is an essential task in disaster risk management.

In the present study we employ a combination of a soil-vegetation-atmosphere transfer (SVAT) model and machine learning for nowcasting the movement of the Reissenschuh landslide in the Schmirn valley (Tyrol, Austria). The landslide’s movement has been monitored periodically since 2016 and continuously since 2020, with annual displacements up to more than 3 m and movement rates between 0.1 to 0.6 cm/day.

In a first step, we use the SVAT model LWF-Brook90 for reproducing subsurface runoff as a proxy of pore water pressure. The model includes vegetation dynamics and their interaction with incoming precipitation, snow accumulation and melt, as well as infiltration processes into porous media. We calibrated and validated the model using time series of snow water equivalent from two monitoring locations in Tyrol (Austria). For considering lag times between the infiltration and the onset of acceleration we computed running sums of subsurface runoff, considering time windows of up to 120 days.

Based on the continuous displacement time series and the outputs of the SVAT model, we trained machine learning models (support vector machines, SVM; random forest, RF) for reproducing the temporal displacement dynamics on a daily resolution. We validated the machine learning models then using the periodical displacement measurements. Based on the combined SVAT/machine learning models, we nowcast the movement of the Reissenschuh landslide using available meteorological products. With the growing displacement time series we will further refine the machine learning models and validate their predictive performance with periodical measurement campaigns. In a next step, we will employ the combined models for predicting the landslide’s movement under selected climate change scenarios.