CAS-Climate: A Novel Process-Driven Method for Flood Frequency Analysis Based on Mixed Distributions

Arizona State University, Scottsdale AZ

Investigators

Abstract

Floods are among the most common and impactful natural hazards. In the U.S., these extreme events caused a total of $144.4 billion in damages and 2550 fatalities from 1991 to 2020. A crucial task to limit the impacts of flooding, design infrastructure, and manage water resources is to increase the accuracy of flood frequency estimates. These are currently generated through statistical analyses of annual peak flows under the implicit assumption that flood events at a given site are caused by the same physical mechanism. This assumption has been challenged by observational evidence and demonstrated to lead to inaccurate estimates. This project will address key limitations of current flood frequency methods by designing a novel approach that incorporates the effect of multiple physical mechanisms leading to flood generation into a statistical model. The approach will be tested at more than 1000 stream gages covering a large range of climatic conditions in the U.S. The knowledge generated by the project will (1) contribute to improving national guidelines for flood estimation, (2) be disseminated to regional stakeholders involved in flood management through training activities, and (3) innovate curricula at Arizona State University. An undergraduate and graduate students will be directly involved in the project activities. The main research hypothesis of this project is that the accuracy of flood frequency analysis is improved by using mixed probability distributions of peak-over-threshold (POT) flood series associated with a set of dominant atmospheric and hydrologic processes. To investigate this hypothesis, the dominant large-scale meteorological patterns (LSMPs) causing flood events will be first identified from atmospheric reanalyses. Key hydrologic processes and conditions occurring in the basins under different flood-producing LSMPs will be obtained from the recent retrospective hydrologic simulations of the National Water Model. Machine learning will be applied to variables characterizing LSMPs and hydrologic processes to group flood events in each basin into a set of dominant flood-generating mechanisms. The hypothesis that the corresponding flood sub-samples are drawn from statistically heterogeneous populations will be tested with a new regional framework based on statistical tests, Monte Carlo simulations, and physical considerations. These physical and statistical insights will be incorporated into a novel method for flood frequency analysis based on mixed distributions of POT series. Performance and uncertainty of the mixed POT model will be quantified and compared with those of homogenous distributions fitted to annual peak flows, as in current approaches, and to POT series. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.