How do you predict an urban flash flood when nobody recorded the last one?
In Bangkok, days of heavy rain have left streets and neighbourhoods under water. The flooding has disrupted travel and daily life, even as water begins to recede in some areas. For people living through it, there is no doubt where the water went. For researchers trying to predict future floods, establishing a precise record of when and where it happened is a separate task.
A river can be monitored by gauges that record rising water. Sudden flooding in a city may occur far from one. Water collects on roads, enters buildings and sometimes recedes before anyone measures its extent. Afterwards, there may be photographs and local news coverage, but little data in a form a forecasting model can readily use.
That gap is the starting point for Groundsource, a project from Google Research. It uses artificial intelligence (AI) to extract details of past floods from news reports. Google has used the resulting dataset to train a separate model that forecasts urban flash flood risk up to 24 hours ahead. Those forecasts are available through its Flood Hub.
Finding floods that gauges missed
Google says Groundsource analysed more than five million news reports covering roughly 20 years. Its system processed articles in 80 languages and used Google’s Gemini model to identify accounts of floods that had actually happened. A report warning that a flood might occur, for instance, should not be mistaken for a record of one that did.
The system then had to establish when and where. “Last Tuesday” only means something alongside an article’s publication date. A neighbourhood mentioned in a report has to be located on a map. From this process, Google assembled an openly available dataset containing 2.6 million records of reported flood events across more than 150 countries.
Note that these are records extracted from reports, not 2.6 million floods independently confirmed by instruments. Nevertheless, they offer evidence of events that conventional monitoring may have missed.
How does Google’s flash flood forecast work?
Gemini’s role is to help build the historical record. It does not predict tomorrow’s flood by reading today’s headlines.
For that, Google trained a separate forecasting model on the flood records, together with weather and geographic data. Given a forecast and the characteristics of an area, the model estimates the risk of an urban flash flood over the next 24 hours. Google focused its initial rollout on more populated areas, partly because news coverage provides more training examples there.
Even a broad warning could give emergency teams time to monitor conditions and prepare a response. But this is a forecast of risk, not an instruction to evacuate. The model currently works in areas approximately 20 by 20 kilometres across. It cannot identify which underpass will fill first or tell a household whether its front door is safe. Local observations and warnings from the responsible authorities remain essential.
Other ways to close the warning gap
Google is not alone in forecasting sudden floods. In Europe, the Copernicus Emergency Management Service already provides urban flash flood information through its European Flood Awareness System. One product uses radar observations and short-term rainfall predictions to highlight risk in urban grid cells of about one square kilometre over the next few hours. It offers finer local detail over a shorter period, in places covered by the radar data. Google’s model looks further ahead across a much larger area.
Researchers associated with the European Centre for Medium-Range Weather Forecasts (ECMWF) are exploring an idea closer to Groundsource’s: learning from reports of past flash flood impacts. Their 2026 preprint describes a machine-learning model trained with impact reports and weather data, and tests whether learning from well-documented regions might help forecasts elsewhere.
That study remains a preprint. Its main quantitative evaluation is based on a data-rich US study region, so it does not yet demonstrate reliable performance around the world. What it does show is why records of past impacts are attracting attention: they may offer a way to study floods where direct measurements are sparse.
Which floods make the record?
News is an inventive source of historical data, but it is not an impartial measuring instrument. A flood in a busy city centre may generate several articles. One in a less populated area may never appear in a searchable archive.
An AI model trained on reports therefore risks learning something about where floods get reported, as well as where they occur. Reviewers of the ECMWF-led preprint have raised a related concern about the effect of population and reporting patterns on that study’s results. The two projects differ, but both must contend with the gap between a flood happening and someone recording it.
The extracted records have uncertainties of their own. In Google’s manual review, 60% had both the correct location and timing. The company judged 82% accurate enough for practical analysis, for example because they identified the right district or placed the event within a day of its reported peak. Those figures assess the historical dataset. They do not tell us how often a forecast shown in Flood Hub will be correct.
Missing reports also make forecasts difficult to assess. An alert may appear to be a false alarm when flooding occurred but was never documented in the records used to check it. Google says it is working to extend the model towards rural areas, where the shortage of reports presents a particular challenge.
Groundsource cannot recover measurements that nobody took. It can find evidence in accounts written for another purpose. A report filed after water entered a street may help forecasters recognise a similar threat somewhere else. How useful that warning becomes will depend on the quality of the record, the forecast and the local response.
Further reading on MoveTheNeedle.news: