Integration of Node Classification in Storm Surge Surrogate Modeling

Abstract

Surrogate models, also referenced as metamodels, have emerged as attractive data-driven, predictive models for storm surge estimation. They are calibrated based on an existing database of synthetic storm simulations and can provide fast-to-compute approximations of the expected storm surge, replacing the numerical model that was used to establish this database. This paper discusses specifically the development of a kriging metamodel for the prediction of peak storm surges. For nearshore nodes that have remained dry in some of the synthetic storm simulations, a necessary first step, before the metamodel calibration, is the imputation of the database to address the missing data corresponding to such dry instances to estimate the so-called pseudo-surge. This imputation is typically performed using a geospatial interpolation technique, with the k nearest-neighbor (kNN) interpolation being the one chosen for this purpose in this paper. The pseudo-surge estimates obtained from such an imputation may lead to an erroneous classification for some instances, with nodes classified as inundated (pseudo-surge greater than the node elevation), even though they were actually dry. The integration of a secondary node classification surrogate model was recently proposed to address the challenges associated with such erroneous information. This contribution further examines the above integration and offers several advances. The benefits of implementing the secondary surrogate model are carefully examined across nodes with different characteristics, revealing important trends for the necessity of integrating the classifier in the surge predictions. Additionally, the combination of the two surrogate models using a probabilistic characterization of the node classification, instead of a deterministic one, is considered. The synthetic storm database used to illustrate the surrogate model advances corresponds to 645 synthetic tropical cyclones (TCs) developed for a flood study in the Louisiana region. The fact that various flood protective measures are present in the region creates interesting scenarios with respect to the groups of nodes that remain dry for some storms behind these protected zones. Advances in the kNN interpolation methodology, used for the geospatial imputation, are also presented to address these unique features, considering the connectivity of nodes within the hydrodynamic simulation model.