Google paper claims AI can better predict floods in low-data areas

Flooding is becoming more frequent and severe as the climate crisis intensifies, and it’s only likely to get worse. That can make the job of predicting where and when these disasters will hit challenging, especially in areas with little historical or real-time data from which to draw.

A new paper from researchers at Google posits that machine learning could hold some answers. The tech giant claims an AI model, which is trained on publicly available data, can predict river-based flooding up to five days in advance with a reliability “similar to or better” that of current systems.

Google VP of Engineering and Research Yossi Matias said what’s particularly notable about this model is that it’s broadly applicable on a global scale, which could provide more accurate forecasting for developing areas underserved by stream gauges, or water measurement stations along rivers.

“The real question was, how can we build a model that is not going to be built for one locale, but is a global model, where we can take whatever data [on what] happens anywhere in the world and put that into the prediction modeling,” Matias told Tech Brew. “The breakthrough is really to, one, show that we can actually take such a global model and do that and, two, to actually measure that in a systematic way.”

The findings come as Google has been expanding its Maps and Search tools to provide consumers and developers alike with more data on climate-related phenomena such as wildfires, air quality, and flooding.

“This is a cross-company effort,” Matias said. “Everything that we do, we’re looking not only into building models, but then to see how they actually impact people through our platforms, and through other platforms.”

Google first piloted its flood forecasting work in 2018 in the Indian city of Patna, located in the flood-prone state of Bihar, Matias said. From there, the program gradually expanded across the country and into Bangladesh. But the location-specific nature of the model initially made it difficult to scale the approach into areas without streamflow data, according to a blog post written by Matias.

That’s where the globally focused model has been useful, Matias said. The company has already implemented it into a platform called Flood Hub that now offers alerts to 80 countries and encompasses about 460 million people, according to Matias. It can also help with coverage as the climate crisis creates flooding risks in areas that haven’t historically been prone to flood, Matias said.

“It turns out that when we started thinking about global models, this was in a situation of, ‘How can we actually predict floods in places that don’t have a lot of history?’” Matias said. “What happened was that because of these new floods that we see in places that never actually experienced floods, it turns out that this model is even more important than we anticipated.”

Right now, Google’s alert system is focused specifically on riverine flooding, which Matias said is the “most common, in terms of the impact on society.” But more research is still necessary to understand and predict other types, like flash, coastal, and urban flooding, he said.

“When you think about other natural disasters, again, with the progress of technology and the ability to actually solve hard problems, ideally, I’d like to get us to the place that nobody is surprised by any crisis or any sort of extreme [weather] event, just because we’re going to have the enough technology that is going to be able to give prediction and alerts in due time…I’m very optimistic that it’s a matter of time before we’re going to be able to do that."