Google鈥檚 artificial intelligence flood prediction system can now forecast floods four days in advance in data-poor regions, like South America and Africa, as well as data-rich areas like Europe and the US.
Most of the world鈥檚 waterways don鈥檛 have accurate measurements for water flow, which makes predicting floods hard. Lower-income countries are disproportionately affected by a lack of data, whereas higher-income countries are more likely to have well-measured rivers and lakes and so more accurate flood predictions.
In 2018, Google introduced an AI-powered flood forecasting system for India and Bangladesh, two of the most flood prone countries in the world. It later expanded the number of countries available and launched a website, . The system had high predictive power for the few countries covered and helped alert people living there to possible floods.
Advertisement
Now, at Google and his colleagues have tested their AI prediction system on data from more countries around the world.
To develop the AI model, Nearing and his colleagues fed it publicly available stream flow data gathered by the , as well as ground and satellite-based weather forecasts from up to a year before. The model worked out the relationship between these variables and water levels from real bodies of water. It could then use this to predict future events when given new data.
鈥淥ne of the limitations of traditional hydrology models is that they鈥檙e really accurate mostly in places where they鈥檙e well calibrated,鈥� says Nearing. 鈥淎I models, even though they鈥檙e data driven, learn general hydrological behaviours and are better able to move from one location to another.鈥�
The researchers tested their model on more than 5000 water flow measurements from across the world, taken between 1984 and 2021, and found that, on average, the prediction level was as accurate four days in advance as the best conventional systems can predict on the day.
鈥淚f you go to our Flood Hub, and you look at the predictions four days from now, you鈥檙e gonna get information that鈥檚 about as accurate as you would have gotten from the existing system if you went and looked for today,鈥� says Nearing.
The system could predict flooding events between four and six days in advance in regions with sparse water data, such as in South America and Africa, and sometimes as far ahead as seven days. The researchers assessed the flood predictions at the continent level rather than at the country level.
While the data in the study was restricted to publicly available data, Google currently has data for more than 80 countries available in its Flood Hub and Google Search. 鈥淎nyone can go and look. And if you鈥檙e in an area that we cover, then you鈥檒l see the information from this model in real time,鈥� says Nearing.
The system has been operational and sending out flood alerts in 80 countries since October 2022.
鈥淭he fact that this improves on our current global benchmarks is really significant for disaster forecasting,鈥� says at the University of Bristol, UK.
arXiv
