Development of AI Algorithms for landslides prediction (Emilia-Romagna Region, Italy)

Landslide risk is one of the most relevant hazard that affects the Emilia-Romagna Region. Almost 80,000 landslides were mapped in the mountainous part, and the percentage of land covered by landslides exceeds in some areas 25%. Although most of the regional landslides are relatively slow, the economic impact is critical: in 2019, 1 million euros was allocated for urgent safety interventions, and it is estimated that at least another 80 would be needed to complete the plan. These numbers place the Emilia-Romagna Region among the areas with the highest landslide risk in the world. The geological characteristics of the Region, combined with the growing exploitation of the territory and the climatic changes underway, are making this problem more and more dramatic. It is now clear that emergency responses are no longer sufficient and that they must be accompanied by prevention actions devoted to mitigate the risk. 

The main objective of this work is to develop Artificial Intelligence models for the prediction of landslides in the Emilia-Romagna Region. The idea is to exploit the data collected by the University of Bologna in the last 15 years, as part of the research activities carried out in collaboration with the Regional Agency for Civil Protection and the Geological Survey of the Emilia-Romagna Region.

Available data consist of time series of rainfall, soil moisture, snow cover and displacement of some active landslides that have occurred in the region in recent years. The displacement data comes from permanent GPS stations, wire strain gauges, and robotic total stations installed in several landslides for emergency purposes. These data show clear relationships between precipitation and rate of movement. However, such relationships are difficult to reproduce using physically-based approaches.

The proposed machine learning approach was applied to the Emilia-Romagna Region of Italy taking advantage of the historical landslide archive, which includes more than 2210 rainfall events  that triggered 2363 landslide, and of the genetic classification algorithm TPOT (Tree-based Pipeline Optimization Tool) with more than 1million combinations of hyperparameters. The results show that landsliding in the study area is strongly related to rainfall event parameters (Precipitation during the event, The day of the event and in which location happened) while antecedent rainfall seems to be less important (Precipitation 30 and 60 days before the rainfall event). The distribution of landslides in the rainfall precipitation - day of the year chart shows that after the dry summer season a rain event of at least 90-100 mm is necessary to trigger a landslide. However, this number decreases as the day of the year increases, and then arrives in spring where many landslides are shown have been triggered with modest rain events (15-30 mm). The algorithm also provided an F1 test result score of 0.825, which means that it can predict a true positive (rainfall event triggers landslide) with a 70% of precision and with 95.5% about true negative (rainfall event do not triggers landslide).