Regional Modeling of Storm Surges Using Localized Features and Transfer Learning

Abstract Storm surges induced by low pressure and high winds from tropical or extratropical cyclones are the main driver of major coastal flooding events. While tide gauges provide the most accurate sea level observations, their records are often short, spatially uneven, and contain gaps, posing challenges for a detailed analysis of surge characteristics continuously along the coastline. To perform extreme value analysis for hazard and risk assessment requires long time series of storm surges, which do often not exist. In this study, we use a regional long short‐term memory (LSTM) model to predict storm surges at multiple tide gauges simultaneously, while also accounting for localized features. Storm surge observations from 31 tide gauges across Florida are used as a predictand; atmospheric and oceanic data are used as dynamic predictors; and shelf width, nearshore slope, shelter factor, and tidal range are used as static predictors. Results demonstrate that incorporating local static attributes improves model performance. Pearson correlation increases by 20% and RMSE drops by 22%, on average, in the cross‐validation. The pretrained model captures localized surge