Abstract
Recurrent flooding due to relative sea level rise and climate change is a growing concern for coastal communities. Forecasting when and where flooding might occur, at a spatial and temporal resolution that can best assist decision-makers, is needed to improve flood resiliency within these coastal communities. Conventional physics-based hydrodynamic models used by flood forecasters can simulate accurate flooding dynamics; however, their computational burden makes them impractical for real-time, hyper-resolution modeling at a city or regional scale. This research explores the potential of a machine learning method, Random Forest (RF), to create surrogates of physics-based hydrodynamic flood models more suitable for real-time, hyper-resolution flood forecasting. Additionally, new crowdsourced data products offer the potential to measure recurrent flooding impacts on communities across the globe. This data, due to its global scale and millions of data collectors, could offer coastal communities a systematic and periodic way to measure impacts, helping communities to better understand changes over time and to identify streets within transportation networks vulnerable to recurrent flooding. With this motivation in mind, the first contribution of this dissertation is to create surrogate flood models using RF to efficiently forecast the depths and extent of pluvial flooding from a high-fidelity, physics-based model in an urban environment. The second contribution is to tailor the surrogate flood model for combined pluvial and tidal flooding common in coastal communities. Finally, the third contribution is to create a method for utilizing crowdsourced traffic data to quantify roadway impacts caused by nuisance flooding based on traffic behavior. Key findings from this research are that (1) RF surrogate models can predict pluvial and tidal flooding with high accuracy while also differentiating between dominant flooding mechanisms (i.e., pluvial, tidal, or combined flooding); (2) RF surrogate models can predict water levels across the 16,914 street segments in the model domain within 4.2s±1.5s per event, which is 3,800 times faster than the original physics-based model, making it a suitable option for real-time flood forecasting; (3) according to crowdsourced traffic data, nuisance flooding during high tide (specifically, tides above 1.0 m NAVD) has increased one-way commuting time by more than 6 minutes per event for 10% of commuters on the flooded routes in Norfolk. Across these three objectives, this dissertation research advances coastal resiliency by creating transferable flood-predictive tools and methods for systematically modeling and measuring flood impacts using machine learning methods and crowdsourced data.
