The system now covers more than 200 million people at risk for flooding across India as well as large portions of neighboring Bangladesh, a country where an average of 5,000 people each year are killed in floods.

Changes in the technology underpinning the system have allowed Google to double the warning time it is now providing, giving people detailed alerts up to 48 hours before flooding occurs.

Floods affect an estimated 95 million to 240 million people worldwide annually, killing between 6,000 and 8,000 of them and causing up to $33 billion in economic damage. Those figures are expected to rise as climate change makes flooding, owing to stronger rainstorms and glacial melting, more frequent and severe.

Google began its Flood Forecasting Initiative in 2017, covering the area around Patna, the capital of the Indian state of Bihar, historically the country’s most flood-prone region. In 2019, Bihar experienced some of the worst floods in a quarter-century, which killed more than 130 people.

The Silicon Valley technology giant has worked with the Indian government’s Central Water Commission to improve the forecasts it relies on. It has also worked with the agency to improve the way it sends alerts to citizens warning them of danger.

Since then Google has steadily expanded the program, and the company says it has helped send more than 27 million flood alerts in India to date.

Sella Nevo, a senior software engineer at Google who leads the flood forecasting project, said part of its improvement in forecasting in India has involved working with the Indian government to improve how it collects data on water levels. This has reduced both erroneous water-level readings and delays that hampered forecasting in the past.

While some of these problems are specific to flood monitoring in India and other developing nations, some of the techniques Google has pioneered in India could change flood forecasting worldwide.

Nevo said even state-of-the-art flood forecasting had previously relied on hydrologic models that were based largely on maps of local topography and conceptual principles derived from physics. Each watershed was thought to be unique—leaving little ability to create a model that would work equally well across different river basins, Nevo explained.

Google, instead, took an approach largely based on A.I., in which software analyzes historical flood data taken from several different river basins in different parts of the world and trains itself to make accurate predictions for almost any river basin.

“One assumption that was presumed to be true in hydrology is that you cannot generalize across water basins,” Nevo said. “Well, it’s not true, as it turns out.” He said Google’s A.I.-based forecasting model has performed better on watersheds it has never encountered before in training than classical hydrologic models that were designed specifically for that river basin.

Of course, issuing these forecasts is one thing. Figuring out how to alert people based on them is another. And so far, exactly how people react to government-issued flood warnings and what kinds of alerts work best are topics that have been understudied.

Google said it is currently working with researchers from Yale University to try to answer some of these questions. Preliminary work by Yale and Google in India has shown that receiving an alert doubles the chance that someone will take action to protect themselves, with about 65% of all people who receive flood warnings taking some protective steps.

But the company has been working to improve these figures. This year, it said it overhauled its alerts to provide information in nine different local languages as well as in a visual formats, which can help people intuitively grasp the warning.

It is also providing people with more information about exactly how far the water is likely to rise in their specific village or area at specific times, based on the Google forecast.
This story has been updated to correctly describe the method Google has used to improve the quality of data fed into its flood forecasts. It has not used electronic sensors as part of the initiative and instead has relied on other ways to improve the timeliness and accuracy of the data.