Large-Sample Hydrology For A Shift In Paradigm From Regional To Global Flood Frequency Analysis

Regional flood frequency analysis (R-FFA) for flood prediction has been often deployed as a tool to extend information available from gaged basins to ungaged ones with similar characteristics. The original context for R-FFA approaches was often characterized by short flow records at single gaging stations and poor capabilities for basin characterization, at a time when only rudimentary regression models were available to study the relationship between basin characteristics and the occurrence of floods with different return periods. Given these premises, regionalization offered some interesting advantages, including the possibility of (i) fitting statistical models to an enlarged dataset of peak flow events, pooled from multiple homogenous watersheds, and (ii) breaking down the full extent of spatial variability in the probability distribution of floods, focusing on a limited range of variation, at the regional scale, that could still be effectively explained using simple linear regression models, using relatively limited watershed characterizations as an input. On the other hand, this approach relied on the limiting assumption that floods from homogenous basins must follow the same normalized probability distribution, and it inevitably introduced subjectivity when defining and identifying homogenous regions. Despite its intrinsic limitations, regionalization remains one of the most used techniques in Hydrology since it first appeared in the technical literature 70 years ago, still enjoying great popularity today. Since those days however, data availability for R-FFA has dramatically increased, both in terms of longer flow records at an increasing number of instrumented watersheds and available information for basin characterization, ranging from climate to land use and geomorphology. Besides, advancements in machine learning (ML) combined with exponentially greater computational power provide more suitable alternatives to the traditional regression models to study complex, non-linear hydrological relationships between basin and flood characteristics. Put together, all these developments could lead one to suggest that regionalization may be considered obsolete, as the limitations that it originally addressed have been overcome in the meantime. Furthermore, switching from a regional to a global approach for FFA may represent an opportunity to achieve greater generalization of ML models by training on a wider variety of dynamics between watershed characteristics and hydrologic response, as reflected in the probability distribution of peak flow events. We aim to develop approaches for global FFA (G-FFA), in contrast with traditional R-FFA, to make continuous predictions across widely different sites, considering a large hydrologic dataset of hundreds of natural watersheds in the U.S., testing different machine learning algorithms and deploying optimization techniques to identify catchment characteristics with greatest explanatory power. The advantages of the alternative, global approach involve (i) eliminating subjectivity in the definition of homogenous regions and (ii) relaxing the assumption that similar basins have exactly the same normalized distribution of flood probability.