Characterizing Global Flood Extremeness Through Physically Informed Neural Networks

The tails of flood distributions provide key insights into the occurrence probability of extreme floods, which is commonly quantified by the shape parameter of an empirical Generalized Extreme Value (GEV) distribution fitted to annual maximum flood series. Despite the usefulness of fitting empirical GEV distributions to observations, considerable uncertainty remains in the estimated shape parameter across different parameter estimation approaches. In addition, most existing studies focus on regional scales, and a global-scale analysis is required to investigate the roles of varying climatic conditions and data quality in shaping extreme flood occurrence.

In this study, we first apply the L-moment method—an approach known for its robustness in extreme value statistics— to conduct a global analysis of extreme flood occurrence based on optimized GEV distributions. The Anderson–Darling test is used to evaluate the goodness-of-fit. We then integrate additional hydrological information, represented by up to 20 descriptors, into a supervised neural network (NN) model to construct a physically informed, data-driven framework for improving the estimation of GEV distribution parameters. A global-scale dataset comprising more than 6,600 river gauges, with record lengths ranging from 20 to 200 years, is used in this analysis.

Preliminary results indicate that the proposed framework can achieve flood distribution tail estimates comparable to those obtained from purely statistical methods (i.e., L-moment estimates), while providing additional physical insights into the estimation process. Overall, this study highlights the potential of integrating multi-dimensional common hydrological descriptors within a data-driven framework to support large-scale and consistent characterization of global flood extremeness.