Simulating spatial multi-hazards with generative deep learning

Abstract

When natural hazards coincide or spread across large areas they can create major disasters. For accurate risk analysis, it is necessary to simulate many spatially resolved hazard events that capture the relationships between extreme variables, but this has proved challenging for conventional statistical methods, particularly in high-dimensional settings. In this article we show that generative deep learning models – when combined with specific transformations to the training data – offer a useful alternative method for stochastically sampling realistic multi-hazard events. Our framework combines generative adversarial networks with extreme value theory in a hybrid approach that can capture complex dependence structures in gridded multivariate weather data, while providing a theoretical basis for extrapolation to new extremes. We apply our method to jointly model fields of strong winds, heavy precipitation, and low atmospheric pressure (∼ 12 000 variables) during storms in the Bay of Bengal, demonstrating that our model learns the spatial and multivariate extremal dependence structures of the underlying data and captures the distribution of storm severities. For the Bay of Bengal case study, we validate our approach against a popular model for multivariate climate extremes, and demonstrate improved performance in capturing the extremal correlation structure.