On March 20, 2024, Google published a Nature paper, “Global prediction of extreme floods in ungauged watersheds,” describing an AI river-flood model that works even where there are no local stream gauges - the situation across much of Africa and Asia, where flood risk is highest and observational infrastructure is thinnest.
The model is built on long short-term memory (LSTM) networks that ingest historical and forecasted weather and produce probabilistic streamflow predictions. It was trained on streamflow records from thousands of gauged watersheds worldwide, and crucially generalizes to ungauged basins. Google reported that the AI achieves reliability at up to a five-day lead time comparable to, or better than, the zero-day nowcasts of the leading prior global system, and matches accuracy on five-year-return-period events that earlier systems only reached for far more common one-year events. In effect it extended reliable warning, on average, from zero days to five.
The model powers Flood Hub, Google’s operational early-warning system, which provides free, publicly available river forecasts up to seven days ahead across over 80 countries, covering hundreds of millions of people exposed to flooding. It is a leading example of machine learning delivering climate-adaptation infrastructure at global scale.