Using surrogate modeling to predict storm surge on evolving landscapes under climate change

Abstract

Planners managing coastal flood risk under a constrained computational budget face a tradeoff. Simulating many time periods or scenarios limits how many storm simulations can be run on each landscape. In this analysis, we present a deep learning model to predict storm surge as a function of storm parameters but also landscape features and boundary conditions (e.g., sea level). It is trained on peak surge elevations from Advanced Circulation (ADCIRC) hydrodynamic simulations of coastal Louisiana in a 2020 baseline and decadal periods from 2030 to 2070 under two morphological and climate scenarios. Leave-one-landscape-out cross-validation yielded a 0.086-m RMSE and 0.050-m MAE over 90 storms per landscape and 94,013 geospatial locations. A two-sided Kolmogorov-Smirnov test comparing annual exceedance probability (AEP) estimates from the model predictions to ADCIRC simulations rejected the null hypothesis that the predicted and ADCIRC AEP values were drawn from the same distribution only 1.1% of the time.