Google’s new flood AI can predict risk even where no data flows

Floods are one of the most common natural disasters. Without warning, a river swells, bursting its banks and sweeping through in a furious torrent. Homes are lost, lives are upended, and a community is left grappling with the aftermath of an unexpected disaster. This scenario, unfortunately, is not uncommon at all. Across the globe, floods kill over 7,000 people annually and the risk of floods is increasing steadily each year.

Communities, especially in the developing world, face the devastating impacts of floods with little to no warning. This is of greatest concern in regions without advanced monitoring systems.

If scientists can improve flooding forecasting, they could save numerous lives and prevent property damage. In a new study, this is exactly what researchers demonstrated.

Predicting disaster

To understand how AI can help, it’s essential to grasp the concept of ungauged watersheds. These are areas lacking in the necessary tools and infrastructure — such as streamflow gauges — to collect data on water levels. This makes traditional flood forecasting methods nearly impossible. With over 90% of flood-related disasters occurring in developing countries, the lack of data poses a significant challenge to predicting and preparing for such events.

Because current forecasting methods largely rely on these stream gauges and this is such a big issue, Grey Nearing and colleagues from the Flood Forecasting team at Google Research thought about using an AI model for forecasts in ungauged areas.

They trained the model on 5,680 existing gauges to predict daily streamflow in ungauged watersheds over a 7-day forecast period. The model was then compared against state-of-the-art systems for both short-term and long-term forecasting.

Computed forecasting

After its training, the AI system was able to provide five-day flood forecasts that are as reliable as (or better than) current same-day predictions. Furthermore, the five-year forecasts for extreme events were also as good (or better) than current one-year forecasts.

“Here we show that artificial intelligence-based forecasting achieves reliability in predicting extreme riverine events in ungauged watersheds at up to a five-day lead time that is similar to or better than the reliability of nowcasts (zero-day lead time) from a current
state-of-the-art global modelling system (the Copernicus Emergency Management Service Global Flood Awareness System),” the researchers note in the study.

“In addition, we achieve accuracies over five-year return period events that are similar to or better than current accuracies over one-year return period events. This means that artificial intelligence can provide flood warnings earlier and over larger and more impactful events in ungauged basins.”

Already, the model has been incorporated into an operational early warning system that produces publicly available (free and open) forecasts in real time in over 80 countries. This platform can be accessed here. Furthermore, fully functional pretrained models have been released by the team — you just need to input your data and run the models.

As promising as this method is, however, there are also limitations.

The AI models’ performance heavily depends on the quantity, quality, and diversity of data available for training. Where historical data are sparse or non-existent, the model’s ability to learn and predict accurately can be significantly hampered. Moreover, inaccuracies in the data, whether from measurement errors or inconsistencies in data collection methods across different regions, can affect model reliability.

Furthermore, AI models, such as the one developed in this study, show promise in generalizing their predictions to ungauged basins. However, there’s always a concern about how well these models can perform across varied and complex geographical and climatic conditions that weren’t sufficiently represented in the training data.

Nevertheless, this is a good example of how AI can make a real impact in the world. By harnessing the power of machine learning and vast datasets, we take a more proactive step in the era of disaster preparedness and response. Unfortunately, with climate change on the rise, this era is about to get more and more challenging.

The study was published in Nature. https://doi.org/10.1038/s41586-024-07145-1