Data-Driven Modeling of Global Storm Surges

Abstract

In many areas, storm surges caused by tropical or extratropical cyclones are the main contributors to critical extreme sea level events. Storm surges can be simulated using numerical models that are based on the underlying physical processes, or by using data-driven models that quantify the relationship between the predictand (storm surge) and relevant predictors (wind speed, mean sea-level pressure, etc.). This study explores the potential of data-driven models to simulate storm surges globally. A multitude of predictors (obtained from remote sensing and climate reanalysis) along with predictands (from tide gage observations and storm surge reanalysis) are utilized to train and validate data-driven models to simulate daily maximum surge for the global coastline. Datadriven models simulate daily maximum surge better in extratropical and subtropical regions [average correlation and root-mean-square error (RMSE) of 0.79 and 7.5 cm, respectively], than in the tropics (average correlation and RMSE of 0.45 and 5.3 cm, respectively). For extreme events, the average correlation decreases to 0.54 (0.33) and RMSE increases to 14.5 (13.1) cm for extratropical (tropical) regions. Models forced with remotely sensed predictors showed a slightly better performance (average correlation of 0.69) than models forced with predictors obtained from reanalysis products (average correlation of 0.68). Results also highlight a significant improvement (i.e., average correlation increases from 0.54 to 0.68; RMSE reduces from 11 to 7 cm) over the Global Tide and Surge Reanalysis (GTSR), derived from the only global hydrodynamic model. For approximately 70% of tide gages, mean sea-level pressure is the most important predictor to model daily maximum surge. Our results highlight the added value of datadriven models in the context of simulating storm surges at the global scale, in addition to existing hydrodynamic numerical models.

A database of global storm surge reconstructions

“…The statistical models are trained with predictors from climate reanalysis and observed water levels at tide gauge locations. Both paradigms, data-driven methods as well as hydrodynamic models, have their own advantages and disadvantages (see Tadesse et al 10 for more information).…”

“…We present here the Global Storm Surge Reconstruction (GSSR) database making use of already developed data-driven models 10 and multiple satellite-era as well as longer-term atmospheric reanalysis products, going as far back as 1836. The long-term surge reconstruction we present here can be used for more robust extreme value analysis, especially in locations where observational records are short, as well as to better understand the trends and longer-term variations in the storm surge climate (e.g., intensity and frequency of storm surge events) from the mid-nineteenth century until present (please see Usage Notes).…”

Probabilistic Characterization of Extreme Storm Surges Induced by Tropical Cyclones

“…The non‐tidal residual component is considered as the SSL in this study, which represents the difference between the recorded sea level and the predicted astronomical tide (Cid et al., 2018; Pawlowicz et al., 2002; Tadesse et al., 2020). Since tropical cyclones do not necessarily lead to significant storm surge events, we focus on the maximum SSL during the passage of each of tropical cyclones (and tropical depression) that necessitate the issuance of tropical cyclone warning signals.…”

“…To justify the robustness of the proposed data‐driven framework, it is necessary to quantitatively compare with previous approaches. The principal components regression (PCR) and the random forest (RF) regression were selected as comparative benchmarks since they have been successfully used in the SSL characterization by linking with the surrounding atmospheric conditions (Cid et al., 2018; Tadesse et al., 2020). The PCR is a multivariate linear regression model that relates the SSL with the principal components of the climatological forcing factors, while the RF is a supervised machine learning algorithm that combines predictions from multiple machine learning algorithms based on the concepts of classification and regression trees, thereby leading to a more reliable prediction than a single model.…”

“…Tadesse et al. (2020) used a random forest regression model to simulate storm surges globally. These parametric data‐driven approaches can yield reasonable SSL simulations, but they commonly require the assumption of normality and linearity.…”