Pluvial flood risk assessment for 2021-2050 under climate change scenarios in the Metropolitan City of Venice

Pluvial flood is a natural hazard occurring from extreme rainfall events that affect millions of people around the world, causing damages to their properties and lives. The magnitude of projected climate risks indicates the urgency of putting in place actions to increase climate resilience. Through this study, we develop a Machine Learning (ML) model to predict pluvial flood risk under Representative Concentration Pathways (RCP) 4.5 and 8.5 for future scenarios of precipitation for the period 2021-2050, considering different triggering factors and precipitation patterns. The analysis is focused on the case study area of the Metropolitan City of Venice (MCV) and considers 212 historical pluvial flood events occurred in the timeframe 1995-2020. The methodology developed implements spatio-temporal constraints in the ML model to improve pluvial flood risk prediction under future scenarios of climate change. Accordingly, a cross-validation approach was applied to frame a model able to predict pluvial flood at any time and space. This was complemented with historical pluvial flood data and the selection of nine triggering factors representative of territorial features that contribute to pluvial flood events. Logistic Regression was the most reliable model, with the highest AUC score, providing robust result both in the validation and test set. Maximum cumulative rainfall of 14 days was the most important feature contributing to pluvial flood occurrence. The final output is represented by a suite of risk maps of the flood-prone areas in the MCV for each quarter of the year for the period 1995-2020 based on historical data, and risk maps for each quarter of the period 2021-2050 under RCP4.5 and 8.5 of future precipitation scenarios. Overall, the results underline a consistent increase in extreme events (i.e., very high and extremely high risk of pluvial flooding) under the more catastrophic scenario RCP8.5 for future decades compared to the baseline.