Observation-Driven Understanding and Prediction of Urban Flood Hazard

Abstract

Rapid urbanization has increased exposure to flood hazards in the United States. Flood hazard is the result of complex interactions of spatially and temporally variable extreme rainfall with heterogeneous land surface, subsurface, and drainage network properties. The time and length scales at which these interactions occur in urban watersheds are shorter than in unaltered settings. Existing flood hazard assessment procedures neglect the complexity of these interactions, with unknown implications for resulting flood risk estimates. This dissertation aims to utilize modern high-resolution hydrometeorological observations and computational capabilities to examine urban flood processes at fine temporal and spatial scales, and to use the results to propose a variety of alternative approaches to the estimation of urban flood hazards.

A variety of hydrometeorological datasets are introduced, including long-term (ten-year) records of bias-corrected high-resolution (1 km, 15-minute) radar rainfall fields developed using the Hydro-NEXRAD processing system for two urban study areas in the Southeastern United States: Charlotte, North Carolina and Atlanta, Georgia. These long-term radar rainfall records are used throughout the dissertation for a variety of hydrometeorological analyses. Their accuracy is evaluated relative to observations from dense urban rain gage networks. Simple bias correction techniques provide high-quality radar rainfall estimates, which have important advantages over coarser-resolution operational radar-based products for urban hydrology applications. Other datasets are used in the analyses, including high-resolution land cover, streamflow, and cloud-to-ground lightning observations.

Radar records show evidence of urban rainfall modification for both study sites, though more work is needed to understand the role of topography and other large-scale features on rainfall in urban areas. Increases in annual peak discharges in urban catchments since the 1950s are attributable to changes in runoff and channel processes due to urbanization, rather than changes in the properties of extreme rainfall. Median annual peak discharge varies systematically with watershed area, but there are also important differences in flood response linked to the spatial distribution of land use and other watershed properties. Precise description of flood response based on basin properties, however, remains a major challenge.

The most extreme rainfall events within the ten-year Charlotte radar rainfall dataset are used to create ``storm catalogs,'' which are used to examine rainfall structure and its interactions with watershed properties. These storm catalogs are used to estimate storm-centered Area Reduction Factors (ARFs). These are compared against conventional ARFs, demonstrating conventional ARFs do not accurately represent the properties of observed extreme rainfall, with important implications for flood risk assessment. ARFs also reveal that spatial rainfall properties vary significantly according to storm type.

Storm catalogs are coupled with a technique known as stochastic storm transposition (SST) for rainfall and flood frequency analysis. SST-based rainfall frequency analyses can reproduce the results of conventional intensity-duration-frequency relationships for a range of return periods, but, unlike conventional methods, preserve observed rainfall structure and variability. SST is coupled with the physics-based Gridded Surface Subsurface Hydrologic Analysis model to demonstrate the shortcomings of several design storm assumptions including the selection of rainfall duration and intensity. A simple storm classification system is used to demonstrate how flood response varies across drainage network spatial scales. Flood hazard at larger scales in Charlotte is the result of rainfall from tropical storms, while organized thunderstorm systems drive flood hazard in smaller watersheds. SST-based flood frequency analyses represent a robust alternative to design storm-based approaches and provides multi-scale estimates of flood hazard using relatively short rainfall records. It can also be used to translate climate-induced changes in rainfall into changes in flood hazards.

Changes in the frequency and intensity of extreme rainfall from tropical storms across the eastern United States between present and late-21st century conditions are compared using downscaled climate model results. The results point to the central role of model uncertainty and the challenge of differentiating between it, natural climate variability, and secular climate trends using relatively short simulation periods.