Google uses Gemini AI and 5 million old news reports to predict flash floods across 150 countries

Google Tackles Flash Flood Prediction with AI-Powered News Analysis

Flash floods kill more than 5,000 people each year, making them among the deadliest weather events globally1

1

. Yet these catastrophic events remain notoriously difficult to forecast. Google believes it has found an unconventional solution by teaching AI to read the news. The company used its Gemini AI model to analyze 5 million news articles from around the world, isolating reports of 2.6 million different floods and transforming them into a geo-tagged dataset called Groundsource1

1

. This marks the first time Google has deployed language models for this type of crisis prediction work, according to Gila Loike, a Google Research product manager1

1

.

Addressing Data Scarcity Through Old News Reports

The core challenge in predicting flash floods stems from their short-lived and localized nature. While humans have assembled extensive weather data, flash floods are too ephemeral to be measured comprehensively the way temperature or river flows are monitored over time1

1

. This data gap has long prevented deep learning models from accurately forecasting these events. "High-fidelity data for certain disasters like flash floods simply did not exist," explained Yossi Matias, VP and Head of Google Research5

5

. News coverage, however, proved broader and longer-lasting than traditional meteorological records2

2

. By cross-referencing flood reporting with weather data, Google created a foundation for training predictive models where none existed before.

How the Flash Flood Forecasting Model Works

With Groundsource as a real-world baseline, researchers trained a model built on a Long Short-Term Memory (LSTM) neural network to ingest global weather forecasts and generate the probability of flash floods in specific areas1

1

. Google Maps determined precise geographic boundaries for each historical flood event to create a dataset focused on urban flash floods4

4

. The AI-powered system can now make predictions up to 24 hours in advance4

4

5

5

. Google's flash flood forecasting model now highlights risks for urban areas in 150 countries on the company's Flood Hub platform, alongside existing riverine flood forecasts that cover 2 billion people1

1

4

4

.

Emergency Response Agencies See Faster Results

Google is sharing its data with emergency response agencies around the world, with early trials showing promising results. António José Beleza, an emergency response official at the Southern African Development Community who tested the forecasting model with Google, said it helped his organization respond to floods more quickly1

1

. For communities worldwide, this means better preparedness before natural disasters strike4

4

. The project was specifically designed to work in places where local governments can't afford to invest in expensive weather-sensing infrastructure or don't have extensive meteorological records1

1

.

Limitations and Future Applications

The model has notable constraints. It operates at fairly low resolution, identifying risk across 20-square-kilometer areas1

1

2

2

. It's also not as precise as the US National Weather Service's flood alert system, partly because Google's model doesn't incorporate local radar data, which enables real-time tracking of precipitation1

1

3

3

. "Because we're aggregating millions of reports, the Groundsource data set actually helps rebalance the map," said Juliet Rothenberg, a program manager on Google's Resilience team. "It enables us to extrapolate to other regions where there isn't as much information"1

1

2

2

. The team hopes that using language models to develop quantitative datasets from qualitative sources could be applied to forecasting other phenomena, including heat waves and mudslides1

1

2

2

. Marshall Moutenot, CEO of Upstream Tech, called Google's contribution "a really creative approach" to addressing one of the most difficult challenges in geophysics1

1

. The Groundsource methodology could potentially transform Crisis Resilience efforts by turning verified reports into datasets that enable improved global preparedness for multiple types of natural disasters4

4

.