Google researchers use AI to predict floods where there’s little data

By Leslie Kaufman

Climate change is making flooding more common and destructive globally. Artificial intelligence has the potential to help mitigate some of that damage by being trained to provide accurate warnings, even in flood basins lacking water gauges.

In fact, an AI model now operational in 80 countries, provides more accurate predictions of river flooding than the previous dominant system, according to a paper published Wednesday in Nature.

The model was developed by researchers at Google, who said in their paper that they have successfully been able to “improve the skill of forecasts in Africa to be similar to what is currently available in Europe.” That’s despite Africa having far fewer flood gauges. In addition, the real-time forecasts are free and publicly available.

Floods are the most common and most widely destructive natural catastrophes, causing an average of $50 billion in global economic damages annually, according to the paper. They are also difficult to predict, particularly in places with sparse or no data. Almost 90 per cent of the 1.8 billion people very susceptible to floods live in low- and middle-income countries, where there are fewer flood gauges than in rich countries — and sometimes none at all.

Google AI modelers tried to predict floods — including especially destructive events — in a river’s watershed without any gauges. Through a large, collaborative effort involving many academics and experts at the EU’s global flood forecasting system, known as GloFAS, which is the current gold standard, the scientists built a predictive AI model.

The Google model uses diverse, publicly available data sources, such as weather forecasts, satellite imagery, topography and soil type. It then uses AI to predict what areas will be affected by a flood and how deep the water will be. The model was tested and then improved based on feedback from 5,680 watersheds.

Researchers found that with AI, they could predict floods five days in advance in river basins without gauges with the same accuracy GloFAS could only do on the day of.

Beth Tellman, chief scientist at Floodbase, a company that develops technologies that can facilitate products that insure against floods in the developing world, concurred that Google's model showed significant improvement over GloFAS and added that it could have major implications for disaster preparation.

"If forecasts can be reliable, they could be used not just for early warning and evacuation to save lives, but to release strategic funding to save lives and property," she said. She listed a number of examples, including “using money to evacuate animals, pile sandbags along the river, harvest rice crops early enough to save them or even stockpile gas and food at cheaper prices before the flood happens and prices spike."

With climate change intensifying precipitation patterns, the need to improve flood forecasts is growing more urgent. Flood risk has more than doubled since the turn of this century. Bringing flood warning systems in developing countries up to the standards of developed countries would save 23,000 lives annually, the World Bank estimates.

Yossi Matias, vice president of engineering and research at Google, said that since the model has been deployed, it has helped predict flooding in Colombia and India. The model does not yet handle other types of flooding, such as urban and coastal flooding, but the team plans to tackle that next.