Google uses Groundsource to expand urban flash-flood forecasting with 2.6 million events

Google on Mar 12, 2026 introduced Groundsource, a new Gemini-powered methodology designed to close one of the hardest data gaps in disaster forecasting: the lack of historical records for urban flash floods. Google says the system analyzes decades of public reports and uses Google Maps to turn scattered accounts into a structured geographic record that can train forecasting models.

According to the company, Groundsource identified more than 2.6 million historical flood events spanning over 150 countries. Google then used Maps to determine geographic boundaries for each event and built a dataset focused on urban flash floods. Using that dataset, Google says it trained a new model that can make progress toward forecasting urban flash floods up to 24 hours in advance.

Key points

• Groundsource uses Gemini and Google Maps to build historical disaster datasets from public reports.
• Google says it identified more than 2.6 million flood events across over 150 countries.
• The resulting model aims to forecast urban flash floods up to 24 hours in advance.
• Google says the approach could also be adapted to hazards such as landslides or heat waves.

That forecast is now being surfaced in Google Flood Hub alongside the company existing riverine flood forecasts. Google says the riverine system covers 2 billion people in more than 150 countries for the most significant river flooding events, and that the urban flash-flood extension broadens the usefulness of the platform for communities that previously lacked detailed historical data. In other words, the advance is not only a research result; it is also a distribution update into an existing public-facing resilience product.

Google also presents Groundsource as more than a single-use flood project. The company says the method can create a massive open-source benchmark for partners and scientists, and that the same AI-driven approach could potentially be adapted to landslides, heat waves, and other hazards that suffer from incomplete historical records. If that transfer works, the practical value of the work could extend well beyond flood forecasting.

For the AI industry, Groundsource is a notable example of a model being used to manufacture better training data rather than only generate end-user answers. For public agencies and crisis-response organizations, the value proposition is more immediate: better data can improve earlier warnings, especially in dense urban areas where flash floods can escalate quickly and existing records are often fragmented.

Source: Google