Google is using old news reports and AI to predict flash floods

Google researchers used Gemini to parse 5 million news articles, extracting 2.6 million flood reports to create a geo-tagged time series dataset called 'Groundsource.' This dataset was used to train an LSTM-based model that predicts flash flood probabilities from global weather forecasts. The model now provides risk assessments for urban areas in 150 countries via Google's Flood Hub platform. The approach addresses data scarcity in regions lacking expensive weather infrastructure, and researchers hope the LLM-to-quantitative-data pipeline can be extended to other hard-to-measure phenomena like heat waves and mudslides.