Flash floods are notoriously difficult to predict, but Google might have a novel solution. The company just revealed Groundsource, a prediction tool for flash floods that uses Gemini to source data from old news reports. This is the first time it has used a language model for this type of work.
Google tasked Gemini with sorting through 5 million news articles from around the world and isolating flood reports. It transformed this data into a geo-tagged series of chronological events. Next, researchers trained a model to ingest current weather forecasts and leverage the Groundsource data to determine the likelihood of a flash flood in a given area.
We don't have any concrete information as to how accurate Google's forecast model is, though that should come over time. One trial user did say it helped his organization respond quicker to localized weather events. For now, the company is highlighting risks for urban areas in 150 countries via its Flood Hub platform. Google is also sharing its data with emergency response agencies in these locations.
There are some limitations here. The model can only identify risk across a 20-square-kilometer area. It's also not quite as precise as the US National Weather Service's flood alert system, because Google's model doesn't integrate local radar data. This data typically enables real-time tracking of precipitation. However, the platform's been designed to work in areas that don't typically have access to that kind of weather-sensing infrastructure.
Juliet Rothenberg, a program manager on Google's Resilience team, hopes that this technology can eventually be used to predict other tricky phenomena. This includes stuff like heat waves and mudslides.
"We’re aggregating millions of reports,” she told reporters this week. "It enables us to extrapolate to other regions where there isn’t as much information."
This is Google's first use of a language model for weather forecasts, but not its first time it has relied on AI for this type of thing. The company's DeepMind WeatherNext 2 forecasting model has proven to be extremely accurate.
