Enhancing Compound Flood Modeling through Efficient Data-Driven, Physics-based, and Transfer Learning Frameworks

Abstract

Compound flooding in coastal and estuarine systems typically occurs during tropical and extratropical cyclones, yet its evolution in space and time remains difficult to predict using data-driven or physics-based modeling approaches. Compound flooding arises from either concurrent or successive occurrences of flood drivers and their nonlinear interactions, including tides, storm surge, waves, precipitation, and river flow. These synergistic flood drivers are modulated by site-specific morphology and uneven availability of observations, leading to impacts larger than those from each driver in isolation. Physics-based hydrodynamic models can represent nonlinear interactions among multiple flood drivers with high fidelity, but their computational cost and domain-specific configuration requirements limit model transferability to adjacent and/or distant coastal systems. Conversely, advanced data-driven approaches (e.g., deep learning) offer substantial advantages in terms of computational efficiency with faster model predictions across model domains. Yet, data-driven approaches are prone to overfitting issues that emerge when applied to other domains whose hydro-geomorphic characteristics, storm types, or observation records differ from those used in the training process. This dissertation enhances compound flood modeling by integrating data-driven, physics-based, and transferable approaches in novel frameworks for extreme water level and compound flood predictions. The proposed frameworks are introduced in three main studies that collectively aim to (i) prioritize storm-relevant extreme patterns in the learning of water level signal, (ii) embed spatial propagation consistent with coastal-estuarine hydrodynamics, and (iii) organize transfer learning around physically meaningful frameworks rather than purely statistical probability. The first study establishes the foundational deep learning architecture by introducing an attention-augmented mechanism that prevents the neural network from allocating most of its capacity to repetitive training data trends and instead prioritizes periods of extreme water level variability; thus, enabling a constrained application of storm-relevant patterns to limit negative transfer, i.e., information from the training domain may not be entirely applicable to target domains. The second study extends the foundational architecture from temporal prediction at individual tide-gauge stations to spatially distributed responses under cyclonic forcing by adopting a cluster-based formulation. Such a formulation conditions attention on spatial subdomains (e.g., spatial clusters), enabling spatially heterogeneous yet dynamically coherent water-level responses. Specifically, it demonstrates how customized deep learning architectures can encode lagged effects of extreme water level evolution along extended coastal reaches. The third study generalizes the cluster-based formulation to a spatiotemporal transfer learning framework that maps sparse observations to a denser grid, node, or cell prediction domain suitable for compound flood applications. Importantly, this novel framework shows that, once the deep learning model has learned to generalize extreme water level variability and represent spatial propagation, it can achieve a zero-shot transfer to unseen coastal domains with differing hydro-morphological configurations. Collectively, the three studies show that combining attention-guided learning, spatially informed model structuring, and physically grounded transfer criteria provides a viable pathway to an efficient and generalizable end-to-end compound flood modeling framework.