A1140
Title: Modeling spatio-temporal extremes via conditional variational autoencoders
Authors: Likun Zhang - University of Missouri (United States) [presenting]
Abstract: Extreme weather events are widely studied in the fields of agriculture, ecology, and meteorology. Enhanced scientific comprehension of the spatio-temporal dynamics of these events could significantly improve policy formulation and decision-making within these domains. In this paper, we propose a novel approach to model spatio-temporal extremes by integrating climate indices using conditional variational autoencoders (extreme-CVAE). The alignment between modeled and true extremal dependence structures showcases the model's ability to be a spatio-temporal extreme emulator. Along with the decoding path, a convolutional neural network was built to investigate the relationship between climatological dynamics and latent-space parameters, thereby inheriting the underlying temporal dependence structures. The extensive simulation validated the effectiveness and time efficiency of the proposed model. Furthermore, we apply our method to analyze monthly maximum Fire Weather Index (FWI) values over eastern Australia from 2014 to 2024, using GEOS-5 data from the Global Fire Weather Database (GFWED). This case study highlights the model's practical utility and performance in real-world scenarios.
