Landslide Displacement Prediction with Machine Learning Techniques

Landslides are natural hazards with a large socio-economic impact in mountainous regions, causing fatalities and damaging properties and infrastructure worldwide. Predicting slope movements and their temporal occurrence is a key component of early warning systems, which reduce landslide risk by delivering timely and meaningful warnings. Moreover, monitoring data are fundamental to understanding the relationship between triggering factors and slope movements. In practice, alert thresholds are set by experts based on their judgment rather than on an automatic procedure. In this study, we apply machine learning techniques to forecast the dynamic evolution of landslides based on monitoring data. Two past case studies with different characteristics are considered, namely the Huangtupo landslide in the Three Gorges Reservoir area (China) and a shallow slope failure in Laakirchen (Austria). Support Vector Regression, Extreme Gradient Boosting, and Long-Short Term Memory neural networks are employed in the framework of a time series forecasting problem. A Bayesian approach is applied to optimize the hyperparameters of the machine learning algorithms. Available meteorological data such as rainfall, air temperature, and the reservoir water level for the Huangtupo landslide are the main features for the multivariate forecasting. With feature engineering, additional features are created based on the monitoring data, and among these, only the relevant variables are selected through grey relational analysis. The performance of these machine learning techniques is evaluated and compared in terms of the root mean square error. Preliminary results show that predictions are highly accurate, especially if the deformation pattern is similar to the landslide movements experienced in the past.