Google is using old news reports and AI to predict flash floods
2 days ago
- #weather-forecasting
- #machine-learning
- #flash-floods
- Flash floods kill over 5,000 people annually and are hard to predict due to their short-lived and localized nature.
- Google used its Gemini LLM to analyze 5 million news articles, identifying 2.6 million flood reports to create the 'Groundsource' dataset.
- The Groundsource dataset was used to train an LSTM neural network model to predict flash floods globally.
- Google's flash flood forecasting model now covers urban areas in 150 countries via the Flood Hub platform.
- The model has limitations, such as low resolution (20-square-kilometer areas) and lack of local radar data integration.
- The project aims to assist regions lacking expensive weather-sensing infrastructure or extensive meteorological records.
- Google hopes to apply this LLM-based approach to other phenomena like heat waves and mudslides.
- Experts highlight data scarcity as a major challenge in geophysics, praising Google's creative solution.