Testing empirical and synthetic flood damage models: the case of Italy

Abstract. Flood risk management generally relies on economic assessments performed by using flood loss models of different complexity, ranging from simple univariable models to more complex multivariable models. The latter account for a large number of hazard, exposure and vulnerability factors, being potentially more robust when extensive input information is available. We collected a comprehensive data set related to three recent major flood events in northern Italy (Adda 2002, Bacchiglione 2010 and Secchia 2014), including flood hazard features (depth, velocity and duration), building characteristics (size, type, quality, economic value) and reported losses. The objective of this study is to compare the performances of expert-based and empirical (both uni- and multivariable) damage models for estimating the potential economic costs of flood events to residential buildings. The performances of four literature flood damage models of different natures and complexities are compared with those of univariable, bivariable and multivariable models trained and tested by using empirical records from Italy. The uni- and bivariable models are developed by using linear, logarithmic and square root regression, whereas multivariable models are based on two machine-learning techniques: random forest and artificial neural networks. Results provide important insights about the choice of the damage modelling approach for operational disaster risk management. Our findings suggest that multivariable models have better potential for producing reliable damage estimates when extensive ancillary data for flood event characterisation are available, while univariable models can be adequate if data are scarce. The analysis also highlights that expert-based synthetic models are likely better suited for transferability to other areas compared to empirically based flood damage models.