Overview
- Google used its Gemini language model to parse roughly 5 million news articles into a geolocated dataset called Groundsource.
- An LSTM forecasting model processes global weather predictions to estimate flash‑flood probability, now displayed for urban areas across about 150 countries and shared with emergency agencies.
- Company‑reported results say 60% of extracted events were precise in time and location, 82% were usable, and severe GDACS floods from 2020–2026 were captured at 85–100%.
- The tool’s limits include coarse ~20 km² resolution and no local radar data, making it less precise than national radar‑based alert systems such as those in the U.S.
- Early users reported practical value—an SADC emergency official said it sped responses—and anyone can check local risk by searching for Google Flood Hub.