Short to long term space-time prediction of rain-induced landslides under uncertainty

Rainfall is the primary natural trigger for landslides, and their threat is expected to rise as the climate warms. To mitigate the consequences of rain-induced landslides, predicting where and when landslides may occur is crucial. We propose a probabilistic modelling framework for synoptic-scale, short-term (hours to days) to long-term (years to decades) space-time prediction of rain-induced landslides. The framework employs a Poisson binomial distribution for the number of successes in a set of Bernoulli trials, each representing a landslide prediction with its own success probability. Our predictors are 35 deep networks that predict landslides occurrence based on rainfall data and information on past landslides. We tested the framework in Italy using hourly rainfall data from 4031 rain gauges and historical landslide records between 2002 and 2022. Results show that hourly rainfall history provides sufficient information to predict the location and timing of landslides without the need for rainfall thresholds or to define event-based rainfall metrics. Applying the forecasting system to 184,080 h between 1 January 2002 and 31 December 2022 we generated a unique, multi-decadal representation of the expected long-term occurrence probability of rain-induced landslides in Italy, which was not otherwise available from landslide catalogues, inventory maps or susceptibility zoning. We expect the modelling framework to enhance landslide early warning systems and to support long-term landslide adaptation and risk reduction strategies. The approach opens the possibility to consider landslide hazard as a combination of independent prediction models of landslide occurrence with associated uncertainty, thus changing the existing paradigm for landslide hazard assessment.