H21E-04 Learning Surrogate Extreme Rainfall-driven Inundation Models with Few Data

Flood hazard assessment demands fast and accurate predictions. Hydrodynamic models are detailed but computationally intensive and impractical for quantifying uncertainty or identifying extremes. In contrast, machine learning surrogates can be quick, but training on scarce simulated or observed extreme data can also be ineffective.

Here, we show that priming deep learning with a conditional Gaussian Process (GP) and bias correcting afterward yields an effective flood hazard surrogate model. Our workflow couples precipitation downscaling, surrogate inundation modeling, and bias correction, exploiting an ensemble-approximated GP in each. In particular, it uses GP-primed, physics-coupled adversarial learning to downscale rainfall. It then uses an ensemble-approximated conditional GP with a Resnet-18 to estimate flood depth. Subsequently, it bias-corrects against measurements using ensemble optimal estimation.

Rainfall (Daymet), reanalysis model-simulated rainfall (ERA5), and hydrodynamic simulations (LISFLOOD) dataset assesses surrogate model skill, training on a rolling window between 1981 and 2019 while testing the year ahead. That required a few training iterations (~100) and data. Inundation depths are estimated rapidly at run-time (~0.006 s/event). Results in the current climate over Chicago deliver an R² value of $0.98$ over multi-year testing, with median flood depth confidence intervals of about 1 cm.

The surrogate hazard chain, primarily applied to inland pluvial flooding, is adaptable. It can be applied to coastal areas (e.g., using SFINCS) and is applicable in nearly any region (e.g., with CHIRPS) with appropriate training or retraining. It is designed to apply to both present and future climate scenarios (e.g., CMIP6/HighResMIP), making it a versatile tool for modeling flood damages, estimating losses, and facilitating proactive disaster management in a changing climate. We illustrate its use for insurance applications assessing inland flood risk.