A Stochastic Deep Learning Approach for Projecting Storm Surges in the Context of Climate Change

Projections of sea level rise are vital for assessing the impacts of climate change, especially in coastal regions. Present sea level rise projections are primarily focused on monthly water levels, but tend to underrepresent the critical role of storm surges. There are some studies that also provide projections of storm surges along global coastlines based on numerical models using meteorological forcing data from Global Climate Models (GCMs). However, those applications are limited by coarse meteorological inputs as well as the computational demands placed by running numerical models for an ensemble of different GCMs and climate change scenarios.

We propose a stochastic deep learning model trained on model output from numerical surge models. It is designed to capture the spatial and temporal dependencies that are characteristic of storm surge time series. Our approach generates potential storm surge scenarios that are consistent with GCM outputs but are not directly determined by those meteorological inputs. A second advantage is that the trained machine learning model has lower computational demands than traditional numerical models which makes it possible to explore different GCMs and climate change scenarios.