Flash floods are among the many deadliest climate occasions on the earth, killing greater than 5,000 folks annually. They’re additionally among the many most troublesome to foretell. However Google thinks it has cracked that downside in an unlikely means — by studying the information.
Whereas people have assembled a number of climate information, flash floods are too short-lived and localized to be measured comprehensively, the way in which the temperature and even river flows are monitored over time. That information hole signifies that deep studying fashions, that are more and more able to forecasting the climate, aren’t in a position to predict flash floods.
To unravel that downside, Google researchers used Gemini — Google’s massive language mannequin — to kind by means of 5 million information articles from all over the world, isolating experiences of two.6 million completely different floods, and turning these experiences into a geo-tagged time series dubbed “Groundsource.” It’s the primary time that the corporate has used language fashions for this sort of work, in line with Gila Loike, a Google Analysis product supervisor. The analysis and information set was shared publicly Thursday morning.
With Groundsource as a real-world baseline, the researchers trained a model constructed on a Lengthy Quick-Time period Reminiscence (LSTM) neural community to ingest climate world forecasts and generate the chance of flash floods in a given space.
Google’s flash flood forecasting mannequin is now highlighting dangers for city areas in 150 nations on the corporate’s Flood Hub platform, and sharing its information with emergency response companies all over the world. António José Beleza, an emergency response official on the Southern African Growth Group who trialed the forecasting mannequin with Google, mentioned it helped his group reply to floods extra shortly.
There are nonetheless limitations to the mannequin. For one, it’s pretty low decision, figuring out threat throughout 20-square-kilometer areas. And it isn’t as exact because the US Nationwide Climate Service’s flood alert system, partially as a result of Google’s mannequin doesn’t incorporate native radar information, which allows real-time monitoring of precipitation.
A part of the purpose, although, is that the undertaking was designed to work in locations the place native governments can’t afford to spend money on costly weather-sensing infrastructure or don’t have in depth information of meteorological information.
Techcrunch occasion
San Francisco, CA
|
October 13-15, 2026
“As a result of we’re aggregating thousands and thousands of experiences, the Groundsource information set really helps rebalance the map,” Juliet Rothenberg, a program supervisor on Google’s Resilience group, instructed reporters this week. “It allows us to extrapolate to different areas the place there isn’t as a lot info.”
Rothenberg mentioned the group hopes that utilizing LLMs to develop quantitative information units from written, qualitative sources might be utilized to efforts to constructing information units about different ephemeral-but-important-to-forecast phenomena, like warmth waves and dirt slides.
Marshall Moutenot, the CEO of Upstream Tech, an organization that makes use of comparable deep studying fashions to forecast river flows for purchasers like hydropower firms, mentioned Google’s contribution is a part of a rising effort to assemble information for deep learning-based climate forecasting fashions. Moutenot co-founded dynamical.org, a bunch curating a set of machine learning-ready climate information for researchers and startups.
“Information shortage is without doubt one of the most troublesome challenges in geophysics,” Moutenot mentioned. “Concurrently, there’s an excessive amount of Earth information, after which if you wish to consider towards reality, there’s not sufficient. This was a extremely artistic strategy to get that information.”
