Google’s flood prediction AI, dubbed the “Flood Hub,” can now forecast potential flooding events four days in advance, even in regions where water flow data is scarce, such as South America and Africa.
In some cases where abundant data is available, Google’s flood forecasting AI can predict floods some seven days before disaster strikes.
Historically, the accuracy of flood predictions has been tied to the availability of precise water flow data. While affluent regions often have access to this critical information, many areas, especially in lower-income countries, don’t.
As described in a recent paper, Google’s latest methods have vastly boosted the system’s effectiveness where data is scarce.
Flood Hub was initially created in 2018 for flood-prone countries like India and Bangladesh.
Grey Nearing and his colleagues at Google have since improved the system by integrating publicly available stream flow data provided by the World Meteorological Organization and ground and satellite-based weather forecasts.
Nearing highlighted the tool’s progression, “One of the limitations of traditional hydrology models is that they’re really accurate mostly in places where they’re well calibrated.”
“AI models, even though they’re data driven, learn general hydrological behaviours and are better able to move from one location to another.”
Nearing and his team conducted extensive tests on the AI system using over 5,000 water flow measurements spanning from 1984 to 2021. Remarkably, predictions made four days ahead were as accurate as what most current systems could estimate on the same day.
Nearing noted, “If you go to our Flood Hub, and you look at the predictions four days from now, you’re gonna get information that’s about as accurate as you would have gotten from the existing system if you went and looked for today.”
You can explore the tool here.
Currently, Google’s Flood Hub provides data for more than 460 million people in over 80 countries and has been dispatching flood alerts since October 2022.
How Google’s flood modeling has improved over time
Flooding is the most frequently occurring natural disaster, costing about $10bn in damages annually and affecting the lives of countless individuals, Google stated in 2021.
That was before the Pakistan floods, which cost some $10bn in their own right.
Forecasting water analysis
The cornerstone of flood forecasting is predicting potential river floods. Hydrologic models, which predict water levels or river discharges based on inputs like precipitation or upstream gauge measurements, are indispensable.
Another model, the inundation model, simulates the behavior of the water as it moves across the floodplain. This enables a more localized analysis of where floods might strike.
Neural networking
In 2020, Google introduced HydroNets, a unique deep neural network structure tailored for water level predictions.
By structuring itself around the analyzed river network, HydroNets enables upstream locations to pass critical data to downstream models.
Google recently developed an AI model to reduce airplane pollution, underscoring the company’s commitment to developing novel models for primarily non-commercial purposes.
