Google expands AI-powered flood prediction

Google now uses artificial intelligence to forecast floods in 150 countries, home to more than 2 billion people. Its models can predict river flooding up to seven days in advance and, starting in 2026, are also beginning to detect flash floods in cities.

The initiative began in 2018 with a pilot project in India. Since then, Google has added more data and models to expand its system’s coverage. It is available through Flood Hub, in some Google Search results, and through an API for organizations.

How AI predicts a flood

The system combines information about rainfall, river levels, soil conditions, and weather forecasts. It then uses several models to calculate how much water will reach a river and which areas could be covered.

For river flooding, the process has two main parts:

• The hydrological model estimates how much water will flow through the river based on weather and ground conditions.
• The flood model calculates which areas could be affected by that flow.

The results appear on a Flood Hub alert map. They can also be integrated into the systems used by public agencies and aid organizations to warn people or act before the water rises.

The difference from many traditional models is that they usually need years of local data, such as historical measurements of a river’s level. Google’s AI can use information from other regions to generate predictions even in places with few sensors. That is a significant advantage in areas with limited resources and measurement infrastructure.

The problem of urban flooding

Predicting a river’s rise is easier than anticipating a flash flood in a city. Rivers often have sensors that record their levels over many years. By contrast, instruments that measure water accumulating on streets, in tunnels, or across urban neighborhoods are uncommon.

Google needed data that did not exist in a reliable global archive. To fill that gap, it created Groundsource, a methodology announced in March 2026.

First, the system used Gemini to analyze more than 5 million news articles published over two decades. From that material, it extracted information about 2.6 million historical floods in more than 150 countries. That dataset was then added to a model designed to predict flash floods in urban areas, with a lead time of up to 24 hours.

That does not mean every flood can be predicted with the same accuracy. Current urban coverage is limited to places with enough data, and Google is still researching how to extend the system to rural and coastal areas.

What changes for people

An early warning can give people time to evacuate, protect their belongings, or deliver aid before roads become blocked. In Nigeria, the organization GiveDirectly used the flood prediction API to send money to families in Kogi before the water rose.

According to data shared by Google, those payments helped families evacuate, protect their belongings, and recover. Their incomes more than doubled, food insecurity fell by 90%, and 93% of beneficiaries said they felt better prepared for future floods.

Google has also published its hydrology framework, the Groundsource dataset, and the prediction API. Weather services and other organizations can combine them with their own data to create alerts tailored to each region.

The next step will be to improve predictions for rural and coastal flash floods. Google is also studying whether the same approach to gathering scattered data could help anticipate heat waves or landslides. The central idea is concrete: when sensors and records are scarce, AI can turn scattered public information into a prevention tool.