Closing the Gap in Pluvial Flash Flood Prediction: A Generalizable AI Model for Global Flash Flood Forecasting

Flash floods, characterized by rapid onset time, often within six hours of a causative meteorological event, constitute a significant global hazard, causing fatalities comparable to riverine floods globally. Despite their impact, these events are frequently inadequately captured by global flood forecasting models, which predominantly rely on riverine discharge data at daily resolutions. Accurate and timely warnings are paramount for effective flash flood risk mitigation.

This study introduces an AI-based forecasting model specifically targeting the prediction of pluvial flash floods. Utilizing a Recurrent Neural Network (RNN) architecture, the model integrates recent regional weather conditions, static characteristics of the region, and hourly weather forecast data to predict the probability of flash flood occurrence within the subsequent 24-hour period for any given region. Training was conducted using the US Storm Events Database (by NOAA, National Centers for Environmental Information). Performance assessments demonstrate the model reliably predicts flash flood events, with accuracy metrics comparable to current state-of-the-art warning systems in the US, specifically the NWS Flash Flood Warnings.

Due to the lack of a suitable global counterpart to the Storm Events dataset, the model was trained exclusively on US data. However, despite this regional focus in training, the model's generalization potential was evaluated globally using global weather data. Evaluation in regions outside the training domain revealed promising generalization capabilities and significant predictive skill, particularly for extreme events previously missed by state of the art global flood forecasting models. This approach shows potential to substantially increase the coverage of detected flood events while maintaining comparable precision, suggesting the viability of AI for establishing robust global flash flood warning systems.