Google turns news reports into data with Gemini to predict flash floods

Google just turned 5 million news articles into a flood prediction system. The research team built Groundsource, a pipeline using Gemini to extract flood data from global news.

Google Research has developed Groundsource, a framework that utilizes Gemini AI to convert millions of unstructured global news articles into a structured database of natural disasters.

The AI was used to sort through 5 million news articles from around the world, isolating reports of 2.6 million different floods, and turning those reports into a geo-tagged time series.

Gemini extracts flood data from news reports by following a strict, four-step analytical pipeline engineered through a sophisticated prompt:

1. Classification: Gemini checks if the text describes an actual, ongoing, or past flood event, filtering out articles about future warnings, risk modeling, or policy discussions.

2. Temporal Extraction: It identifies the specific dates the flood occurred, using the article’s publication date to calculate relative timelines like “last Tuesday”.

3. Spatial Extraction: It extracts granular, specific locations (such as streets or neighborhoods) that were explicitly flooded, ignoring overly broad regions or areas only considered “at risk”.

4. Location Reconciliation: It matches the extracted location names against standardized geographic identification codes so they can be accurately mapped to physical coordinates.

Here’s a snippet of what the actual Gemini prompt used:

You are a meticulous flood event analyst. Your task is to analyze the provided article text (text),
URL (url), and publication date (date) to extract information about a single, specific flood
event and map locations to a reference database. You must only respond with a single, clean
JSON object. Do not include any markdown formatting, text, or explanations outside the JSON.
Step-by-Step Instructions:
Phase 1: Flood Event Extraction & Verification
Analyze the article text (en_text), URL (url), and publication date (date) to determine if a single
flood event occurred and extract verifiable details.
Step 1 - Initial Analysis:
Carefully read the article text, taking note of the publication date.
Step 2 - Core Task - Classify Article Type (The "Gate"):
You must first determine if the article describes a single, actual, ongoing, or past flood event.
An actual flood is an event that the text describes as a fact that has happened or is currently
happening.
Crucial Distinction: An article is NOT about an actual flood if it only discusses:
- Warnings/Predictions: Flood warnings, advisories, forecasts, or statements about potential or
future risk (e.g., "floods may occur," "government warns of floods," "14 states are at risk,"
"heavy rainfall could cause flooding").
- Policies/Preparations: Flood-related policies, defense projects, community preparations, or
government meetings.
- Multiple Flood Events: The article describes separate flood events in different locations (e.g.,
a flood in Brazil and a flood in Italy).
- Other: future risk modeling, or general discussions.

This initiative addresses the critical shortage of historical data needed to train accurate hydrological models, especially for localized urban flash floods that satellite networks often miss.

You can download the actual dataset of Groundsource here. You can use this online tool to view the parquet file.

This is what the dataset looks like:

It has 6 columns:

1. uuid: A unique identifier for each record

2. area_km2: Area of the reported location polygon.

3. geometry: The spatial boundary of the reported location, utilizing the standard WGS 84 coordinate reference system (EPSG:4326). Depending on the extraction and geocoding process, this geometry may represent a complex polygon (e.g., an administrative district boundary) or a buffered point (e.g., representing a specific street intersection).

4. start_date: The initial day (formatted as YYYY-MM-DD) for which there is documented textual evidence of an ongoing flood.

5. end_date: The final consecutive day (formatted as YYYY-MM-DD) for which there is documented evidence of the flood.

6. _index_level_0__: row counter

Has a total rows of 2,646,302 and file size is 636.2 MB.

Here’s a global map showing the density of flood events. Red dots indicate floods from GDACS.

Google is now rolling out these forecasts in Flood Hub platform. It now provides urban flash flood forecasts up to 24 hours in advance.

Here’s a preview of what Google’s Flood Hub looks like:

For example if you live within the Petaling District, Selangor in Malaysia or nearby areas, you have to be careful of potential flash floods within the next 24 hours.

Existing limitations

The model itself faces operational and forecasting constraints:

Forecasting Constraints

• Low Resolution: The model currently identifies flood risks across relatively broad 20-square-kilometer areas.

• No Real-Time Radar: It lacks the precision of systems like the U.S. National Weather Service because it does not incorporate local radar data to track live precipitation.

Dataset (Groundsource) Limitations

• AI Extraction Errors: Manual reviews found that 18% of the AI-extracted events were unusable due to errors like mapping a city to the wrong country or hallucinating incorrect dates based on vague reporting.

• Recency Bias: Because the system relies on available digitized online news, 64% of the recorded global floods occurred between 2020 and 2025.

• Geographic and Linguistic Blind Spots: The text-extraction tool only supports 80 languages, meaning regions with limited digital news infrastructure or unsupported local languages are systematically underrepresented in the data.

References:

• Blog post: https://research.google/blog/introducing-groundsource-turning-news-reports-into-data-with-gemini

• Paper: https://eartharxiv.org/repository/view/12083

• Dataset: https://zenodo.org/records/18647054

---

Join 250k+ developers staying ahead in AI. We curate the latest models, repos, and research — so you don’t miss what matters: AlphaSignal.ai

---