Deep learning forecast of rainfall-induced shallow landslides in Italy

Abstract

Rainfall induced landslides occur in all mountain ranges posing severe threats to people, property, and the environment. Given the projected climate changes, in many areas the risk posed by rainfall induced landslides is expected to increase. For this reason, the ability to anticipate their occurrence is key for effective landslide risk reduction. Empirical rainfall thresholds and coupled slope-stability and rainfall infiltration models are commonly adopted to anticipate the short-term (from hours to days) occurrence of rainfall induced shallow landslides. However, empirical evidence suggests that they may not be effective for operational forecasting over large and very large areas. We proposed a deep learning based modelling strategy to link hourly rainfall measurements to landslide occurrence. We constructed a large ensemble of 2400 neural network models which we informed using hourly rainfall measurements taken by more than 2000 rain gauges and information on more than 2400 landslides in the period from February 2002 to December 2020 in Italy. Our results have indicated that (a) it is possible to effectively anticipate the occurrence of the rainfall induced shallow landslides in Italy, and (b) the location and timing of the rainfall-induced shallow landslides are controlled primarily by the precipitation. Our results open to the possibility of operational landslide forecasting in Italy, and possibly elsewhere, based on rainfall measurements and quantitative meteorological forecasts aided by deep learning based modelling.