Google researcher details how AI is pushing flood forecasts beyond sensor-sparse regions

Google on Tuesday offered a behind-the-scenes look at how it is using artificial intelligence to improve flood forecasting, including warnings in data-scarce regions where traditional models struggle. In an interview-style blog post, Deborah Cohen, a senior staff research scientist on Google Research’s Climate Crisis Resilience team, said the company has worked for nearly a decade to evolve its flood models and expand where they can predict hazards in time to help communities respond.

Google said its system uses AI to forecast river flooding up to seven days before it strikes and urban flash flooding up to 24 hours before it hits. The company said the resulting alerts reach more than 2 billion people across 150 countries, and it describes Flood Hub as the primary interface for viewing prediction alerts on a map. Google also said the forecasts show up in Search for users looking for information about floods near them.

The company said its Flood Hub approach is built on a collection of models that process large-scale global data inputs including rainfall, river levels, and ground surface conditions. For river floods, Cohen described two AI models: a Hydrologic Model that forecasts how much water will flow through a river using weather and land conditions, and an Inundation Model that predicts which areas will be affected using streamflow data. Google emphasized that there are many physical gauges in rivers that measure water levels over time, which helps support these predictions.

Cohen said the major challenge for urban flash floods is the lack of sensors and historical records. Google said Flood Hub originally focused on river flooding and only later expanded to urban flash flooding because “almost none” of the global data available for flash floods comes from monitored urban areas. She said the company had to create its own dataset to train and improve the urban flash flood modeling.

Google introduced what it calls Groundsource in March 2026, describing it as a new AI-powered methodology that turns public disaster information into a high-quality data archive, starting with flash floods in urban settings. Cohen said Groundsource uses Gemini to read more than 5 million news reports about flooding over a 20-year period. Google said this process produced a dataset of 2.6 million historical flood events across more than 150 countries, which was then integrated into an urban flash flood model now live in Flood Hub.

Google also positioned its work as both an operational warning tool and a research platform. It said researchers and aid organizations can use a Floods API, described as a way for organizations to access flood predictions so they can warn and support people before severe floods. Cohen cited Give Directly as an example, saying that last year the organization used Google’s Flood Forecasting API to deliver cash to people in Kogi, Nigeria before water levels rose. Google reported that beneficiaries experienced outcomes including higher incomes, reduced food insecurity, and a majority saying they felt better prepared for future floods.

On the product side, Google said Flood Hub draws on “leading weather data” from around the world and uses AI to generate prediction alerts mapped to specific locations. It described the tool as a leap beyond conventional flood models that usually require local historical water-level data for calibration. In Google’s view, that difference matters because the communities most in need of early warnings often live in regions with limited monitoring infrastructure.

Beyond building models, Cohen said Google is trying to lower barriers for other meteorological agencies and researchers. The company said it open sourced its hydrology framework so National Meteorological and Hydrological Services (NMHSs) and other agencies can integrate their own data with the model to generate tailored forecasts. Google also said it made the Groundsource dataset and the Flood Forecasting API publicly available to support additional flood research.

Some limitations remain, according to Google’s own framing. Cohen said the current models are limited to predicting flash floods in urban areas, where the data quality is best. She described ongoing research to improve coverage for rural flash flooding and to include coastal flooding, noting that the existing datasets make those expansions difficult. She also said the Groundsource methodology is being explored for other disasters, including heat waves and mudslides, though Google did not provide timelines or results for those areas.

Going forward, attention is likely to focus on whether Google can extend reliable forecasts beyond urban flash flooding as it continues to fill data gaps, and whether the open-source tools lead to broader adoption by national agencies. For organizations that integrate external forecast data into planning and assistance, the more immediate question will be how to operationalize these warnings in real-time decision-making as models are refined.