Abstract
Coastal cities are experiencing increasingly frequent and severe flooding driven by the combined influences of heavy rainfall, high tides, storm surge, and sea level rise. Effective management of these evolving hazards requires an integrated approach that brings together real-time observation, rapid predictive capabilities, and high-fidelity hydrodynamic modeling. This dissertation advances such an approach by developing a multi-component framework for monitoring, forecasting, and analyzing compound flooding in coastal urban environments.
The first component introduces a novel computer-vision methodology that leverages publicly available traffic cameras as opportunistic sensors for real-time flood monitoring. A deep-learning–based image segmentation pipeline was designed to detect roadway inundation under complex visual conditions, and a geometric calibration method was developed to convert segmentation outputs into quantitative water-depth estimates. This approach provides an inexpensive, scalable alternative to traditional sensor networks, greatly expanding observational capacity in data-sparse urban areas.
The second component of the dissertation presents a hyper-resolution spatiotemporal surrogate model that integrates Convolutional Neural Networks (CNNs) with Long Short-Term Memory (LSTM) architectures to emulate physics-based hydrodynamic simulations. Trained on outputs from physics-based flood models, the surrogate predicts high-resolution water levels across Norfolk up to one hour in advance. The results demonstrate that data-driven surrogates can reproduce complex coastal–urban flood dynamics including tide-rainfall interactions and watershed-scale responses, while achieving computation times suitable for operational early warning systems. This provides a practical pathway for municipalities seeking to supplement or accelerate traditional flood forecasting tools.
The third component develops a one-way modeling framework that applies a storm surge model (Delft3D) as the boundary condition for an urban flood model (TUFLOW). Application to the Park Place neighborhood reveals substantial amplification of flood depths during compound events and highlights the importance of representing both inland and coastal drivers of flooding. The results underscore the limitations of uncoupled models and demonstrate the value of jointly modeling multiple hydrologic and hydrodynamic mechanisms for coastal urban flooding.
Together, these three components form an integrated framework that spans observation, rapid prediction, and process-based simulation. The dissertation provides new tools for understanding compound flood behavior, enhances the ability to monitor and forecast flooding in real time, and offers a physically grounded basis for future planning and adaptation. The findings contribute to the advancement of resilient coastal flood management and offer practical methodologies that can be extended to other coastal cities facing similar challenges under a changing climate.
