Outrunning flash floods: improving global medium-range forecasts for better preparedness

Pillosu, F. M. (2026) Outrunning flash floods: improving global medium-range forecasts for better preparedness. PhD thesis, University of Reading. doi: 10.48683/1926.00128953

Abstract/Summary

The UN’s ”Early Warning for All” initiative, supported by WMO, prioritises flash floods due to their high mortality rates, widespread exposure, major economic impacts, and climate driven exacerbation. Global medium-range predictions are essential for protecting vulnerable populations. However, despite recent advances in medium-range numerical weather prediction and hydrological modelling, developing medium-range predictions of areas at risk of flash floods over a continuous global domain remains severely constrained by computational requirements, data availability, and the inherent challenge of predicting localised extreme events. This thesis aims to develop a first proof-of-concept of medium-range (up to day 5) predictions of areas at risk of flash floods over a continuous global domain. To achieve this goal, three interconnected research objectives are addressed using the Continental United States (CONUS) as the primary study region. The selection of CONUS leverages the Storm Event Database, which provides a comprehensive long-term record of flash flood impact reports essential for robust model development and validation. First, a flash-flood-focused verification framework, directly comparing rainfall predictions and flash flood impact reports, is developed. This framework is used throughout the thesis, but it is first used to assess whether rainfall forecasts from global NWP models can identify areas at risk of flash floods up to medium-range lead times. The ERA5 reanalysis and ERA5 forecasts, post-processed with the ecPoint technique, show good reliability and discrimination ability up to day 5. Second, multiple data-driven models—including random forest, gradient boosting, and neural networks—are evaluated to determine their capacity to extract predictive signals using severely imbalanced observational datasets. These models integrate hydro-meteorological variables from the ERA5 reanalysis and forecasts with flash flood impact reports to identify patterns indicative of flash flood risk. When compared to rainfall-based predictions alone, the data-driven hydro-meteorological predictions demonstrate superior performance while maintaining computational efficiency suitable for operational deployment. Among the tested models, the XGBoost implementation of gradient boosting emerges as the best performer. Third, systematic sensitivity analysis demonstrates that it is possible to deploy such a regionally-trained data-driven model with global hydro-meteorological forecasts to produce medium-range predictions of areas at risk of flash floods over a continuous global domain. This research helps to establish methodologies to extract valuable predictive information from limited observational datasets and lower-resolution hydro-meteorological forecasts that could enhance preparedness and emergency management strategies, contributing to the UN’s ”Early Warnings for All” initiative.

Item Type	Thesis (PhD)
URI	https://centaur.reading.ac.uk/id/eprint/128953
Identification Number/DOI	10.48683/1926.00128953
Divisions	Science > School of Archaeology, Geography and Environmental Science > Department of Geography and Environmental Science
Date on Title Page	July 2025
Download/View statistics	View download statistics for this item