Outrunning flash floods: improving forecasts for better preparedness

Flash floods are one of the most devastating natural hazards. Every year, they cost thousands of lives and millions of dollars in damaged infrastructure. They can occur in large or small catchments, rural or urban areas, close or away from rivers, and with little to no warning. Some regions might have adapted to protect infrastructure and people against this hazard; however, with climate projections suggesting that extreme rainfall might increase in intensity and frequency, "residual risk" might increase in protected areas while unprotected ones might experience unseen severe losses. Hence, relying on forecasts that offer good predictions of areas at risk of flash floods, with enough lead time to extend preparedness and action time windows, is becoming increasingly important.
This presentation will show the most recent developments in data-driven prediction of areas at risk of flash floods, over a continuous global domain and up to one week ahead. The method uses global reanalysis and medium-range post-processed rainfall forecasts to improve the detection of extreme localised rainfall events. It then tests different machine learning algorithms to learn the complex, non-linear relationships between hydro-meteorological parameters to determine the areas at risk of flash floods. The presentation will also focus on cross-validation, hyperparameter tuning, and ensemble approaches to address the issues that arose due to the severely imbalanced dataset we had to work with.
We will finally explore the added value of these data-driven forecasts and reflect on what this all might mean for decision-makers to extend their preparedness and action time window when the next low-probability, high-impact flash flood event strikes.