Probabilistic Models Significantly Reduce Uncertainty in Hurricane Harvey Pluvial Flood Loss Estimates

Abstract

Pluvial flood risk is mostly excluded in urban flood risk assessment. However, the risk of pluvial flooding is a growing challenge with a projected increase of extreme rainstorms compounding with an ongoing global urbanization. Considered as a flood type with minimal impacts when rainfall rates exceed the capacity of urban drainage systems, the aftermath of rainfall-triggered flooding during Hurricane Harvey and other events show the urgent need to assess the risk of pluvial flooding. Due to the local extent and small-scale variations, the quantification of pluvial flood risk requires risk assessments on high spatial resolutions. While flood hazard and exposure information is becoming increasingly accurate, the estimation of losses is still a poorly understood component of pluvial flood risk quantification. We use a new probabilistic multivariable modeling approach to estimate pluvial flood losses of individual buildings, explicitly accounting for the associated uncertainties. Except for the water depth as the common most important predictor, we identified the drivers for having loss or not and for the degree of loss to be different. Applying this approach to estimate and validate building structure losses during Hurricane Harvey using a property level data set, we find that the reliability and dispersion of predictive loss distributions vary widely depending on the model and aggregation level of property level loss estimates. Our results show that the use of multivariable zero-inflated beta models reduce the 90% prediction intervalsfor Hurricane Harvey building structure loss estimates on average by 78% (totalling U.S.$3.8 billion) compared to commonly used models.

Hierarchical Bayesian Approach for Modeling Spatiotemporal Variability in Flood Damage Processes

“…A possible explanation for this is that more than 80% of these households had high flood experience and 99% of the households had implemented one or more private precautionary measures. ), since this functional form has been proven to be suitable (Merz et al, 2013;Rözer et al, 2019;Schröter et al, 2014;Wagenaar et al, 2017). Values of rloss lie between 0 and 1.…”

“…We test and prove the hypothesis that, in transfer scenarios, HBMs are superior compared to generalized and localized regression models. In order to improve loss predictions for regions and events for which no empirical damage data are available, we use variables pertaining to specific region-and event-characteristics representing commonly available expert knowledge as group-level predictors within the HBM.Flood intensity is influenced by duration of inundation, along with inundation depth (Rözer et al, 2019). Households experiencing longer inundation duration experience higher building damage .…”

Probabilistic Flood Loss Models for Companies

“…Still, stage‐damage functions often omit other damage influencing factors such as inundation duration or preparedness (Kelman & Spence, 2004; Middelmann‐Fernandes, 2010; Thieken et al, 2005) and, more importantly, cannot account for interactions among the variables. As a result, stage‐damage functions can only partially describe the damage processes (Gissing & Blong, 2004; Merz et al, 2004; Rözer et al, 2019; Schröter et al, 2014; Sieg, Vogel, et al, 2019). The advance of machine learning and data mining promoted the development of multivariable flood loss models, which jointly consider a variety of damage influencing factors and their interdependency.…”

The object-specific flood damage database HOWAS 21

“…direct monetary damage) as well as injured people and fatalities in Switzerland since 1972 using press articles as the main source of information. Broader overviews of eventspecific databases are provided by Tschoegl et al (2006), Gall et al (2009), De Groeve et al (2014 or Rudari et al (2017), for example.…”

“…Most recent innovations are probabilistic damage models, e.g. based on Bayesian networks (Lüdtke et al, 2019;Wagenaar et al, 2018) or Bayesian regression (Rözer et al, 2019). Their main advantage is that by returning predictive distributions instead of deterministic point estimates, they inherently provide uncertainty information together with their damage predictions.…”