In this study, we present an empirically guided predictive approach to estimate the likelihood of one or more fatality incidents to occur in a specific flash flood event. To do so, flash flood events with and without human losses from 2001 to 2011 in the U.S. Storm Data are supplemented with other extra datasets describing the storm event, the spatial distribution of the sensitive characteristics of the exposed population and built environment at the county level. The study focus on the vehicle-related incidents as the literature proposes that the majority of people perish while inside their vehicle or is attempting to escape from a vehicle being swept. Therefore, we propose indicator variables (candidate risk predictors) relevant for quantifying vulnerability and the prominent human risk in vehicle-related circumstances during flash floods. Random forest, a well-known decision-tree based ensemble machine-learning algorithm for classification is applied then to assess the probability of vehicle-related fatalities occurrence in a flash flood given the conjunction of selected risk predictors.
The dynamics of the flash flood event in this model are represented by distributed hydrologic model-based discharge forecasts generated by the Flooded Locations And Simulated Hydrographs (FLASH) system. Hydrologic simulations are inserted to the model to estimate human risk on a daily or hourly basis for every U.S. County. As a prototype case, the model is applied to the catastrophic flash floods of May 2015 in the U.S. with focus on Texas and Oklahoma. The estimated probabilities of vehicle-related human risk to flash flooding are mapped dynamically for every county in the study area and compared with impact observations from the Storm Data. The results indicate the importance of time and space-dependent human vulnerability and risk assessment for short-fuse flood events, and suggest machine learning as a promising approach in disaster research. This method, based mainly on publicly available national datasets, can support a nationwide pre-operational prediction tool for forecasters and emergency managers to target their warnings on anticipated human impacts, using the model combined with real-time hydrologic forecasts. The study strongly recommends more systematic human impact data collection to advance impact-based predictive models for flash flood casualties using machine-learning approaches in the future.