Google Uses AI and News Reports to Improve Flash Flood Forecasting

Flash floods are among the deadliest and most difficult-to-predict weather events, causing over 5,000 fatalities each year. Google is tackling this challenge by leveraging artificial intelligence and an unconventional data source: news reports. This innovative approach aims to provide more accurate and timely flash flood warnings, particularly in regions with limited weather infrastructure.

The Power of “Groundsource”

Google researchers utilized Gemini, its large language model, to analyze 5 million news articles from around the world. The system identified and extracted data from 2.6 million distinct flood events, creating a geo-tagged time series dataset named “Groundsource.” This marks the first time Google has used language models for this type of scientific data extraction, according to Gila Loike, a Google Research product manager.

How it Works: From Headlines to Hazard Maps

With Groundsource as a foundation, researchers trained a Long Short-Term Memory (LSTM) neural network to ingest global weather forecasts and calculate the probability of flash floods in specific areas. This model is now integrated into Google’s Flood Hub platform, providing flash flood risk alerts for urban areas in 150 countries. The data is also shared with emergency response agencies globally. António José Beleza, an emergency response official at the Southern African Development Community, reported that the system helped his organization respond to floods more quickly.

Addressing Data Scarcity

A key benefit of this approach is its ability to address data scarcity. Traditional deep learning models require extensive historical data for accurate predictions, but flash floods are localized and short-lived, often occurring in areas with limited monitoring equipment. By aggregating millions of news reports, Google’s system helps “rebalance the map,” enabling forecasts for regions with less available information. Juliet Rothenberg, a program manager on Google’s Resilience team, explained that this method allows for extrapolation to areas with limited data.

Beyond Flash Floods: Future Applications

Google’s team envisions broader applications for using large language models to convert qualitative data into quantitative datasets. This could potentially extend to forecasting other speedy-moving hazards, such as heat waves and mudslides.

Flood Hub: A Comprehensive Resource

The Flood Hub provides AI-driven flood forecasts and data to support communities stay safe. It visualizes Google’s flood forecasting models on an easy-to-use map, offering:

• Water trends
• Real-time flood forecasts
• Local inundation maps
• Flood alerts

All information on Flood Hub is free and updated daily, providing a vital resource for disaster response and individual safety. Flood Hub provides life-saving riverine flood forecasts up to 7 days in advance.

Limitations and Future Development

The current system has a resolution of 20 square kilometers, which may be too coarse for highly localized precision. It also doesn’t incorporate local radar data, unlike the U.S. National Weather Service’s flood alert system. However, the project’s focus on regions lacking extensive weather infrastructure makes this trade-off intentional. The Flood Hub provides users with locally relevant flood data and flood forecasts up to 7 days in advance.

As Marshall Moutenot, CEO of Upstream Tech, noted, Google’s operate is part of a larger trend toward assembling training data for AI-driven weather models, addressing a significant challenge in geophysics: the scarcity of reliable data.