A Hybrid Framework for Generating Probabilistic Compound Flood Inundation Maps

Flood risk assessment is primarily performed by a single flood driver at a specific location. A significant flaw in this approach is the oversight of the nonlinear interactions between various flood drivers (e.g., river flooding, tides, storm surges, and fluvial regimes), potentially resulting in compound flooding. This oversight can lead to underestimating the socioeconomic consequences of compound floods, which often surpass the risks posed by individual drivers acting alone. This study employs a deep learning model, mainly Long Short-Term Memory to predict water levels in tidal rivers under the influence of various flood drivers. The model is used to predict water levels at specific locations, using upstream river discharge, downstream water levels, and initial water levels as input variables. To account for coincidence/concurrence of drivers, we use Copula functions as a probabilistic approach to model the correlation between peak river discharge and coastal water levels as input features for the DL Model. The application of the proposed method is illustrated by applying it to a case study in the Buffalo Bayou area near Houston, TX. The results show that, for a 50-year flood, considering prior water level conditions represented by the 10th and 90th percentile baseflow scenarios, the projected flood inundation area can vary significantly, ranging from 30% to 70% for the same return period. The proposed methodology advances flood hazard assessment in coastal regions by capturing the complex interplay of different flood drivers and offering a robust yet practical flood inundation mapping approach.