Year: 2026
Type: preprint
Abstract: Flash floods account for approximately 85% of flooding cases worldwide and exhibit the highest mortality rates among hydrological hazards.Yet, effective early warning systems remain largely absent in the Global South due to a lack of dense observational networks.In this study, we present an AI-based global urban flash flood prediction system that generalizes effectively across unseen regions.Our approach leverages recent advances in deep learning, which allow for a paradigm shift: For the first time, we can directly train models on global ground truth data extracted from millions of news articles, as opposed to scarce and expensive measures of physical quantities.We demonstrate that this deep learning approach achieves results comparable to the state-of-the-art National Weather Service Flood and Flash Flood Warnings in the United States, despite using lower-quality inputs that are available globally.The system is operational and provides free and open real-time forecasts in over 150 countries (https://g.co/floodhub)-asignificant step toward democratizing access to high-quality disaster warnings and building global socioeconomic resilience in the face of an accelerating hydrological cycle.
DOI: 10.31223/x5177s
Institution: Google (United States)
Language: English
Subfield: Global and Planetary Change
Field: Environmental Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7134286497