Prediction of Rainfall-Induced Shallow Landslides at a Geographic Scale Using Deep Learning: an Application to the Liguria Region (NW Italy)

The present work addresses the implementation of a probabilistic framework for the spatio-temporal prediction of rainfall-induced shallow landslides at geographic scales. The methodological approach consists of two main phases. In the first phase, based on statistical analysis and machine learning, landslide occurrence is modelled as a Bernoulli experiment. Using a historical catalogue of rainfall-induced shallow landslides in Italy and a 20-year pluviometric dataset collected from the rain gauges of the national civil protection network, the probability of rainfall-induced landslide initiation is estimated at locations where rainfall data are available. In the second phase, the rainfall-based probability is spatialized and coupled with landslide susceptibility, a spatial variable that accounts for the influence of geological and geomorphological characteristics of the territory. Two techniques were used: a deterministic approach based on the Voronoi tessellation and a stochastic model implemented using the MUSE (Modeling Uncertainty as a Support for Environments) software. The coupled spatio-temporal forecasting model was tested over the entire Liguria region during a significant rainfall event that occurred in early March 2018. The outcomes showed that integrating both temporal and spatial variables is effective in forecasting rainfall-induced shallow landslides.