Particle.news

Google Puts AI-Built Flood Forecasts Into Public View With Flood Hub

Trained on 2.6 million news‑derived flood reports, the Gemini‑based system provides up to seven‑day risk outlooks for data‑poor regions.

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.