From Flash Flood Vulnerability and Risk Assessment to Property Damage Prediction: the Value of Machine Learning

Atieh Alipour,Ali Ahmadalipour,Peyman Abbaszadeh,Hamid Moradkhani

doi:10.5194/egusphere-egu2020-19057

Abstract

&lt;p&gt;Flash floods, as a result of frequent torrential rainfalls caused by tropical storms, thunderstorms,&lt;br&gt;and hurricanes, are a prevalent natural disaster in the southeast U.S. (SEUS), which frequently&lt;br&gt;threaten human lives and properties in the region. According to the U.S. National Weather&lt;br&gt;Service (NWS), flash floods generally initiate within less than six hours of an intense rainfall&lt;br&gt;onset. Therefore, there is a limited chance for effective and timely decision-making. Due to the&lt;br&gt;rapid onset of flash floods, they are costly events, such that only during 1996 to 2017 flash&lt;br&gt;floods imposed 7.5 billion dollars property damage to the SEUS. Therefore, estimating the&lt;br&gt;potential economic damages as a result of flash floods are crucial for flood risk management and&lt;br&gt;financial appraisals for decision makers. A multitude of studies have focused on flood damage&lt;br&gt;modeling, few of which investigated the issue on a large domain. Here, we propose a systematic&lt;br&gt;framework that considers a variety of factors that explain different risk components (i.e., hazard,&lt;br&gt;vulnerability, and exposure) and leverages Machine Learning (ML) for flood damage prediction.&lt;br&gt;Over 14,000 flash flood events during 1996 to 2017 were assessed to analyze their characteristics&lt;br&gt;including frequency, duration, and intensity. Also, different data sources were utilized to derive&lt;br&gt;information related to each event. The most influential features are then selected using a multi&lt;br&gt;criteria variable selection approach. Then, the ML model is implemented for not only binary&lt;br&gt;classification of damage (i.e., whether a flash flood event caused any damage or not), but also for&lt;br&gt;developing a model to predict the financial consequences associated with flash flood events. The&lt;br&gt;results indicate a high accuracy for the classifier, significant correlation and relatively low bias&lt;br&gt;between the predicted and observed property damages showing the effectiveness of proposed&lt;br&gt;methodology for flash flood damage modeling applicable to variety of flood prone regions.&lt;/p&gt;

Full Text