Data requirements and scientific efforts for reliable large-scale assessment of landslide hazard in urban areas

Landslides in urban areas are conceived as phenomena capable of tearing the physical structure as well as the networks of socio-economic, cultural, material and immaterial relations that make up the life of cities. Landslide hazard analysis is usually mandatory for proper land use planning and management. Nevertheless, in some cases (e.g., municipality of Rome in Italy) regulatory plans lack detailed thematic mapping of geohazard-related data. In Italy, the safety of urban areas has become a very important issue in the last decade, therefore projects of national interest have been funded for the mitigation of geological risks. Shallow landslides are common mass movements in urban areas. They can be triggered by earthquakes, heavy rains or induced by proximity to specific urban assets, like road cuts or retaining walls. Reliable quantification of landslide hazardous areas is often associated with the existence of static specific predisposing factors, such as local terrain variables, land use, lithology, proximity to roads and streams as well as dynamic factors related to trigger (e.g., antecedent rainfalls). Predictive multivariate statistical analysis, among which Machine Learning (ML) models, takes as input several predisposing and conditioning factors that may reveal patterns with the spatial and temporal distribution of different types of landslides. Therefore, ancillary landslide databases are the key-data to investigate the distribution, types, pattern, recurrence, and statistics of slope failures and consequently to determine the overall landslide hazard. However, the amount and quality of available data may be inadequate to build accurate large-scale predictive models. Open-source landslide inventories are often incomplete in spatial and temporal terms, with heterogeneous geometries, thus generating a data sparse environment consisting of a variety of low-accuracy datasets that need to be integrated and cross-validated to gain reliability. In this study, the adoption of a combined approach based on GIS tools and Machine Learning techniques allowed to estimate landslide susceptibility based on both real and synthetic Landslide Initiation Points (LIPs). Open-source landslide inventories have been collected, cross-validated, and integrated in a unique database, thus creating a richer data product that contains the strengths but overcomes the weakness of each contributing dataset. As the number of LIPs was too low to train reliable ML models, we developed a methodology based on the features of occurred landslides in order to derive synthetic LIPs to boost the original database by three times. This approach has been applied to the Metropolitan area of Rome (Lazio, Central Italy), where rainfall-induced shallow landslides have been widely overlooked. The final database with LIPs and predisposing factors has been used to create and validate different ML models and the most accurate one was then deployed to estimate landslide susceptibility for the whole area of the municipality of Rome with a resolution of 5 meters. The obtained results were then compared with pre-existing, regional, national, and European scale susceptibility maps to assess their reliability in case more detailed studies are not available. Eventually, rainfall probability curves were estimated to evaluate the temporal dependence of rainfall-induced shallow landslides.