Flood frequency estimation at global scale: Exploring the power of machine learning combined with satellite precipitation estimates

Significant changes in precipitation extremes due to climate change coupled with continuous increase of urbanization in flood prone areas across the entire globe, necessitates the development of ways to improve resilience of our communities and infrastructure to flood hazards. Proper engineering design for flood mitigation measures relies on accurate flood frequency estimates that in turn require hydrologic information across scales and over long periods of time; the said information currently lacking in many parts of the world. However, recent advances in global precipitation datasets together with state-of-the-art flood prediction procedures relying on machine learning (ML) approaches, open up new horizons for flood prediction at global scale.

In this work, we develop and test a framework for global flood frequency estimation that is based on three main aspects. First, ML models for flood peak prediction that have been trained in data-rich regions, are applied in hydro-climatically similar ungauged areas. Second, global remotely-sensed and atmospheric reanalysis precipitation datasets are used as input for the global application of ML models. Third, the estimated flood peaks are used for flood frequency estimation following the Simplified Metastatistical Extreme Value (SMEV) framework that has demonstrated promising results when using simulated streamflow peaks. Results are contrasted against flood quantiles from in-situ streamflow observations for a number of catchments and regions including Brazil, UK, Chile and Australia. The performance accuracy of the proposed framework is discussed with respect to climatic and topographic factors.