Spatio-temporal cross-validation to predict pluvial flood events in the Metropolitan City of Venice

Abstract

• A novel ML-based methodology for pluvial flood risk prediction is proposed. • Forward features selection and spatio-temporal CV permit to reduce model overfitting. • ML model reproduces biggest pluvial flood historical events with high accuracy. • A set of pluvial flood risk maps are developed to support local adaptation processes. • Distance to rivers and to main road strongly contribute to pluvial flood risk. Due to a combination of climate change and urbanization, the instances of pluvial flooding are expected to increase in the next decades posing raising threats to properties, people and productive assets. Predicting and mapping pluvial flood-prone areas is becoming a crucial step in flood mitigation and early warnings, as well as climate change adaptation strategies, to be incorporate in urban planning. Most commonly applied machine learning (ML) procedures for pluvial flood risk assessment, neglect to account for spatio-temporal constraints, leading to overoptimistic models that underestimate the prediction error. In this paper, we propose a novel ML-based methodology for pluvial flood risk prediction in the Metropolitan City of Venice which, introducing a features selection process and spatio-temporal cross-validation, permits to reduce overfitting of the resulting ML models. Spatio-temporal characteristics of floods are derived from a dataset of 60 historical events occurred in the area between 1995 and 2020. Logistic Regression (LR), Neural Networks (NN) and Random Forest (RF) models are applied to identify and prioritize sub-areas that are more likely to be affected by pluvial flood risk, considering the daily precipitation amount and 12 different triggering factors. The models were validated using Random Cross-Validation (R-CV) and Leave Location and Time Out cross-validation (LLTO-CV), that split data in training and validation set considering both time and space. In addition, a forward features selection procedure was applied to identify the features, among the triggering factors, that better face spatio-temporal overfitting in pluvial flood prediction based on the Area Under the Curve (AUC) score. Results suggest that Logistic Regression and LLTO-CV represent the most reliable model to predict pluvial flood events in new spatio-temporal conditions, while, among the triggering factors, distance to river and distance to road resulted the prominent ones.