A1925
Title: Conditional EVT-GAN for extreme precipitation risk mapping
Authors: Kang Sora - Chonnam National University (Korea, South) [presenting]
Na Myung Hwan - Chonnam National University (Korea, South)
Yoon Sanghoo - Chonnam National University (Korea, South)
Abstract: As climate change and environmental variability intensify extreme events, mapping regional risks has become essential for climate and environmental risk assessment, return-period estimation, disaster prevention, and policy planning. Extreme value theory (EVT) models block maxima or high-concentration extremes using the generalized extreme value (GEV) distribution, but conventional EVT is limited in representing gridded risk patterns. Generative adversarial networks (GANs) can learn spatial fields, but standard GANs often fail to preserve extreme upper-tail characteristics. EvtGAN combines EVT and GAN by fitting GEV distributions to block maxima, transforming data into a uniform space, and restoring samples through inverse GEV transformation. However, existing EvtGAN mainly relies on r=1 block maxima, which limits training samples under short observation periods. Moreover, GEV parameters are used only in transformation steps, not explicitly in the generator or loss function. An improved conditional EvtGAN is proposed that uses pixel-wise r-th order statistics and incorporates GEV parameters into an EVT-based loss function. The goal is to generate spatial risk maps based on annual maxima or upper-order statistics, rather than simultaneous event fields. This approach is expected to support environmental extreme risk mapping, risk simulation under climate and environmental change, and regional risk assessment.
