Sinkhole risk assessment by using machine learning model: the case study of Guidonia-Bagni di Tivoli plain (Rome), Italy

Abstract

Sinkholes that occur in settled carbonate lands can be a critical source of risk for human properties and activities since they can abruptly produce serious damage to property and people in densely populated flat areas. This work presents a sinkhole susceptibility and risk assessment mapping in Guidonia-Bagni di Tivoli plain (Central Italy), which is a carbonate sinkhole-prone study area where sudden occurrences of sinkholes have happened in past and recent times. We consider a point-like sinkhole inventory and a series of environmental sinkhole-controlling factors on the study area, related to its geo-litho-hydrological asset, i.e. travertine thickness, and to its terrain deformational scenario, i.e. ground motion rates derived from InSAR COSMO-SkyMed imagery. A sinkhole susceptibility map was generated by means of maximum entropy algorithm  - MaxEnt model – and it was then combined with data on vulnerability and elements-at-risk economic exposure derived from cadastral inventories and market and income values, in order to provide a final sinkhole risk map of the Guidonia-Bagni di Tivoli area. The results show that areas at higher risk covers about 2% of the total study area and primarily relies on the zoning of the main urban fabric. In particular, it is worth to highlight that 5% of the whole road-network pavement and 27% of all the residential buildings fall into higher risk classes. Outcomes of this work reveal the potential of MaxEnt model to assess sinkhole susceptibility for predicting sinkhole areas, either provide a sinkhole risk map as a useful tool for geohazard risk and urban planning management strategies.