Rescuers at work in Casamicciola, on the Ischia in Campania, on 26 November 2022, after a landslide due to heavy rains that killed 12 people. Credit: Pasquale Gargano/Pacific Press via ZUMA Press Wire.
Dangerous landslides triggered by rain are likely to increase in frequency through the effects of climate change. A new study1 describes how machine learning can be applied to the analysis of rainfall events that trigger landslides, potentially leading to a nationwide warning system that could anticipate them. Current models mostly rely on geological and morphological analyses that can only explain the predisposing conditions, but not predict actual events.
Researchers from the Institute for Geo-Hydrological Protection of Italy’s National Research Council’s analysed the available data about rainfall and landslides in Italy in the last 20 years. The catalogue contains about 2,500 rainfall events that triggered landslides and about 60 million rainfall events that did not. “Using a deep learning strategy, we proved that landslides can be anticipated accurately using only rainfall data, without the need for terrain or geological data,” explains Alessandro Mondini, a researcher at National Research Council, and the study’s first author. “This opens to the possibility of predicting landslides also where such thematic data, which are difficult and costly to collect, are not available”.
The team employed neural networks to find differences between even very similar rainfall patterns. “When we have rains whose temporal patterns are similar it is difficult to discriminate the hidden connections between the conditions that have, and have not resulted in landslides,” Mondini explains. The model ensemble classified correctly more than 90% of the rain events that had triggered landslides, and over 85% of the events that did not cause slope instabilities.
For this study, researchers assumed that landslides and rains maintain the same intensity over short periods, though this is a theoretical condition, assumed only for the purposes of the study. One of the main gaps is the absence of a historical series of at least 30 years of rains with hourly data throughout the country, as well as a high-quality data catalogue of landslides covering the same period,” says Mondini
The research team plans to integrate the network with a predictive meteorological system to exploit rainfall data in real time and predict the likelihood of a landslide occurring in a certain area. “Current landslide forecasting systems are used after the rain has already fallen and leave little time to act,” Mondini says. “This tool could increase a region’s preparedness.”
