Google uses old news reports and AI to predict floods

Flash floods are among the world’s deadliest weather events, killing more than 5,000 people each year. They are also among the most difficult to predict. But Google thinks it has circumvented that problem in an unexpected way – by reading the news.

Although humans have compiled a lot of weather information, floods are too short-lived and localized to be measured completely, the way temperature or even river flow is observed over time. That data gap means that deep learning models, which are increasingly able to predict weather, cannot predict floods.

To solve that problem, Google researchers used Gemini — Google’s giant language model — to sift through 5 million news articles from around the world, classify 2.6 million flood reports, and turn those reports into a geo-tagged time series called “Groundsource.” It’s the first time the company has used language models for this type of work, according to Gila Loike, Google Research product manager. The study and data set were shared publicly on Thursday morning.

With Groundsource as a real-world basis, the researchers trained a model built on a long-term short-term memory (LSTM) neural network to import global weather forecasts and generate local flood probabilities.

Google’s flood forecasting model now highlights the vulnerability of urban areas in 150 countries on the company’s Flood Hub platform, and shares its data with emergency response agencies around the world. António José Beeleza, an emergency response officer at the Southern African Development Community who tested the forecasting model with Google, said it helped his organization respond quickly to floods.

There are still limitations to the model. For one, it is a very low resolution, which identifies the risk in 20-square-kilometer areas. It’s also not as accurate as the US National Weather Service’s flood warning system, in part because Google’s model doesn’t include local radar data, which allows for real-time tracking of rainfall.

Part of the point, however, is that the project is designed to work in areas where local governments can’t afford to invest in expensive weather-sensing infrastructure or don’t have extensive records of weather data.

“Because we aggregate millions of reports, the Groundsource dataset helps recalibrate the map,” Juliet Rothenberg, program manager in the Google Resilience group, told reporters this week. “It allows us to go to other regions where there is not much information.”

Rothenberg said the team hopes that using LLMs to develop quantitative data sets from documented, high-quality sources can be used in efforts to build data sets about other ephemeral-but-important-prediction events, such as heat waves and mudslides.

Marshall Moutenot, CEO of Upstream Tech, a company that uses similar deep learning models to forecast river flows for clients such as hydroelectric companies, said Google’s offering is part of a growing effort to integrate data into deep learning-based weather forecasting models. Moutenot was co-founded by dynamical.org, a group that processes a collection of machine learning-friendly climate data for researchers and startups.

“Data scarcity is one of the most difficult challenges in geophysics,” Moutenot said. “At the same time, there’s a lot of Earth data, and when you want to test it against reality, it’s not enough. This was a really creative way to get that data.”