Google is using old news reports and AI to predict flash floods
A new way to solve data scarcity: Turning qualitative reports into quantitative data with an LLM.
TL;DR
- Google is using its Gemini large language model to process 5 million news articles and identify 2.6 million flood events.
- This data forms "Groundsource," a geo-tagged time series used to train a Long Short-Term Memory (LSTM) neural network for flash flood prediction.
- The model is being integrated into Google's Flood Hub platform, providing risk information for urban areas in 150 countries.
- This approach aims to overcome data gaps in predicting localized and short-lived flash floods, especially in regions with limited weather infrastructure.
- The project's success could lead to similar LLM-based data creation for forecasting other phenomena like heat waves and mudslides.