Advancing Domain‐Wide Surrogate Modeling of Coastal Water Levels by Integrating Wet/Dry Classification
Water Resources Research · 2026-11 · 62 (11)
DOI: 10.1029/2025wr042841 ↗Kurz gefasst
Fachartikel in der Zeitschrift Water Resources Research, online erschienen am 17.11.2026. Verfasst von Sunghoon Han und 9 weiteren Autor:innen. Die Arbeit nennt 78 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Engineer Research and Development Center.
Abstract (Originalsprache)
Coastal flood modeling with high‐resolution hydrodynamic models such as ADCIRC + SWAN offers detailed predictions but is computationally intensive, limiting scalability for probabilistic or real‐time applications. This study develops a hybrid emulator framework that combines Gaussian Process Regression (GPR) for water levels with a binary classifier for wet/dry status to enable fast and accurate coastal flood predictions. To efficiently emulate water levels across a computational domain, nodes are categorized by a Wetness Ratio (WR) that quantifies how frequently each location is inundated in training simulations. This metric identifies persistently wet and intermittently wet regions, enabling targeted model architectures for each. The emulator is trained using spatial and physical forcing attributes derived from ADCIRC + SWAN simulations. GPU‐compatible libraries supported efficient domain‐wide training and inference. Cross‐validation results show that emulator accuracy improves with increasing training sample size, with RMSE values approaching an asymptotic minimum beyond 400 simulations. Classifier metrics show that Extreme Gradient Boosting (XGB) outperforms Random Forest in highly imbalanced cases, where wet samples are much fewer than dry (low WR) samples. To improve emulator robustness in transition zones, node‐specific probability thresholds were selected using Precision–Recall (PR) curves, with recall ≥0.9 and precision ≥0.6 as performance criteria. This approach suppressed false wet predictions while preserving flood‐extent realism. The unified GPR + XGB emulator achieves strong agreement with ADCIRC + SWAN simulations, and comparison with satellite‐derived waterlines provides independent spatial evidence that the framework reproduces simulator‐predicted inundation boundaries while emphasizing the role of bathymetry and mesh resolution in observational agreement.
Förderung
- Engineer Research and Development Center
Zitieren
APA
Han, S., Howe, C., Hill, D. F., Honegger, D. A., Williams, C., Ruggiero, P., Haller, M. C., Wang, Z., Graffin, M., & Bergsma, E. (2026). Advancing Domain‐Wide Surrogate Modeling of Coastal Water Levels by Integrating Wet/Dry Classification. Water Resources Research, 62(11). https://doi.org/10.1029/2025wr042841
Vancouver
Han S, Howe C, Hill DF, Honegger DA, Williams C, Ruggiero P, et al. Advancing Domain‐Wide Surrogate Modeling of Coastal Water Levels by Integrating Wet/Dry Classification. Water Resources Research. 2026;62(11). doi:10.1029/2025wr042841
BibTeX
@article{Han2026,
author = {Han, Sunghoon and Howe, Carter and Hill, David F. and Honegger, David A. and Williams, Carson and Ruggiero, Peter and Haller, Merrick C. and Wang, Zhenqiang and Graffin, Marcan and Bergsma, Erwin},
title = {Advancing Domain‐Wide Surrogate Modeling of Coastal Water Levels by Integrating Wet/Dry Classification},
journal = {Water Resources Research},
year = {2026},
volume = {62},
number = {11},
doi = {10.1029/2025wr042841}
}
EUVIMED