Fast generation of widespread extreme events based on extreme-value theory

Severe climate events are becoming more frequent, leading to many fatalities, significant economic damage and disruptions to vital infrastructure. As a result, accurately estimating the frequency and potential consequences of widespread extreme events has become a critical need. However, the limited availability of observations of extreme events poses a major challenge for impact studies, and even large sets of climate simulations often lack sufficient extreme or record-breaking events for thorough analysis. In contrast, weather generators adapted to extreme observations can efficiently produce a large number of plausible extreme events, even those with unprecedented intensity levels. 

Using fundamental principles from spatial extreme-value theory, we adapt traditional Fourier-based phase-randomisation to specifically generate high-resolution synthetic datasets of rare extreme events. The key feature is that the stochastically generated datasets exhibit the same spatial tail dependence as the observed extreme events. Compared to other existing methods for modelling spatial extremes, our approach is distinguished by speed, easy implementation and scalability to higher dimensions.

Using high-resolution datasets for precipitation and temperature, we show that our algorithm produces realistic spatial patterns of extreme events.  We successfully generated datasets with 10,000 grid points, and this number can be easily increased. Given the need for high-resolution climate data in many impact models, our algorithm is particularly useful for robust impact and vulnerability assessments.

References

- Van de Vyver, H. (2024) Fast generation of high-dimensional spatial extremes, Weather Clim. Extrem. 46, 100732. https://doi.org/10.1016/j.wace.2024.100732.