Google’s Groundsource Uses AI to Predict Flash Floods

Google has launched Groundsource, a new AI-powered methodology leveraging its Gemini large language model to analyze decades of public reports and predict urban flash floods up to 24 hours in advance. This initiative addresses a critical data gap in flash flood forecasting, particularly in regions lacking robust weather-sensing infrastructure.

The Challenge of Flash Flood Prediction

Flash floods are among the world’s deadliest disasters, accounting for approximately 85% of flood-related fatalities worldwide, claiming over 5,000 lives annually [1]. Early warning systems (EWS) are crucial for mitigating damage and saving lives, with even a 12-hour lead time potentially reducing flash flood damage by 60% [1]. But, many developing countries lack access to these life-saving systems.

How Groundsource Works

Groundsource tackles this challenge by utilizing Gemini to sift through approximately 5 million news articles globally, identifying and extracting reports of past flood events [3], [4]. This data is then transformed into a geo-tagged, chronological dataset. Researchers then trained a model to combine this historical data with current weather forecasts to assess the likelihood of flash floods in specific areas [3].

Flood Hub and Global Coverage

The flash flood predictions generated by Groundsource are now available through Google’s Flood Hub, initially focusing on urban areas in 150 countries [3]. Google is also sharing this data with emergency response agencies to improve preparedness and response efforts.

Limitations and Future Applications

Currently, the Groundsource model identifies risk within a 20×20 kilometer area [2]. It’s also less precise than systems utilizing real-time radar data, as it doesn’t currently integrate this type of information. However, it’s designed to function effectively in regions where such infrastructure is limited.

Google researchers envision expanding this AI-powered methodology to predict other challenging phenomena, including heat waves and mudslides [3]. Juliet Rothenberg, a program manager on Google’s Resilience team, emphasized the value of aggregating millions of reports to extrapolate insights to data-scarce regions [3].

Groundsource and Gemini: A First

This marks Google’s first application of a large language model like Gemini for weather forecasting [3], building upon previous AI-driven initiatives within Google’s Crisis Resilience efforts [1].