Predicting a flood hours or days in advance can make the difference between evacuating in time and getting trapped by the water. Google says its advances in artificial intelligence now make it possible to generate forecasts in 150 countries, where more than 2 billion people live.
The company began testing its system in India in 2018, using real-time river data. Now, its Flood Hub platform also incorporates Groundsource, a methodology that turns millions of disaster-related news reports into useful data for studying sudden urban floods.
How Google uses AI to anticipate floods
The system combines large amounts of information about rainfall, river levels and terrain conditions. Using that data, several models calculate the likelihood of a flood and estimate which areas could be affected.
According to Google, the system can predict river floods up to seven days in advance. In the case of sudden floods in cities, alerts can arrive up to 24 hours beforehand.
Forecasts appear on Flood Hub, a tool that displays alerts on a map. They may also appear in Google Search when someone looks for information about flooding in their area. In addition, organizations can use the Flood Forecasting API to create their own alert and support systems.
The idea is not to replace local authorities, but to give them additional information so they can act before the water rises.
The data problem in cities
River floods are relatively easier to study because sensors have recorded water levels for years. These instruments work like rulers, helping us understand how a river behaves when it rains.
In cities, the situation is different. Sudden floods can appear on streets, in tunnels and in urban areas where there are few, if any, sensors. Without enough historical records, training an AI model becomes much more difficult.
Google needed to build a database from scratch. To do so, it used Gemini to analyze more than 5 million news reports published over a 20-year period. The result was a file containing 2.6 million flood events across more than 150 countries.
This dataset, called Groundsource, is used to power a new model focused on sudden urban floods. It is an interesting example of how AI can turn information that already exists—but is scattered and disorganized—into a tool for prevention.
Why it matters in regions with few sensors
Traditional models often need local historical data to work well. That puts many communities at a disadvantage because they do not have monitoring stations, even though they may be precisely the ones that most need to receive alerts.
AI makes it possible to use patterns observed in different parts of the world. That does not mean all territories are the same, but it does mean a model can learn from more scenarios and provide estimates in areas where there was previously not enough information.
From alerts to concrete assistance
Flood Hub is not used only to visualize maps. Researchers and humanitarian organizations also use it to decide where and when to send assistance.
Google highlights the case of Give Directly, an organization that used the Flood Forecasting API in Kogi, Nigeria. There, it delivered money to families before water levels rose.
According to the results cited by Google, that assistance allowed families to evacuate, protect their belongings and recover. Beneficiaries' incomes more than doubled, food insecurity fell by 90%, and 93% of people said they felt better prepared for future floods.
This case shows something important: a prediction is only truly useful when it becomes a decision. An alert can help move people, protect documents, distribute food or send money before an emergency blocks access routes.
What comes next for Groundsource
For now, the sudden-flood model is limited to urban areas because that is where the largest amount of collected data exists. Google is researching how to expand its capabilities to include sudden floods in rural areas and coastal events.
The company is also studying whether Groundsource could help analyze other disasters, such as heat waves and landslides. The logic would be similar: gather scattered public information, organize it and use it to identify patterns that support prevention.
Google says it has published its hydrology framework so national weather and hydrological services can combine their own data with the model. It has also made the Groundsource dataset and the Flood Forecasting API available to encourage new research.
AI cannot stop a flood, but it can help reduce its consequences. The difference lies in turning a prediction into time to act, especially in communities where traditional monitoring systems still do not exist.
Original source
https://blog.google/innovation-and-ai/technology/research/flood-prediction-ai