A recent study published in the renowned scientific journal Nature has unveiled an AI (artificial intelligence) model that shows promise for improving the accuracy of flood predictions. The research demonstrates that the AI system performs comparably or better than current state-of-the-art methods and may enable earlier warnings of extreme flooding events.
According to the paper, human-induced climate change is leading to an increase in the frequency of floods in some regions. Current forecasting methods rely on river gauges (monitoring stations along rivers) but these are unevenly distributed globally, posing limitations. This makes it more challenging to predict flows in ungauged rivers, a problem that primarily affects developing nations.
The study's lead author and corresponding author, Grey Nearing of Google AI in the United States, collaborated with colleagues to train an AI model on 5,680 existing gauges to predict daily streamflow in ungauged catchments for a seven-day forecast horizon. The AI model was compared to the Global Flood Awareness System (GloFAS), a leading global framework for forecasting short- to long-range flood conditions. The results revealed that the AI model issued flood alerts up to five days earlier, while maintaining comparable or higher accuracy.
Moreover, the AI model performed comparably or better in predicting extreme events with a five-year return period than GloFAS did for events with a one-year return period.
The authors conclude that their AI model demonstrates the potential to provide advanced warning of both small and extreme flood events in ungauged basins, with a longer lead time than previously possible, and promises to provide more reliable flood forecasts in developing regions.
