Dissecting Groundsource: When LLMs Become Scientific Instruments

1. What is Groundsource?

In February 2026, Google Research released Groundsource — an open-access global dataset of 2.6 million historical flood events extracted from news articles using Gemini LLMs. The dataset was published on Zenodo with a preprint on EarthArxiv.

The key claim: Google used Gemini to scan 5 million news articles across 80+ languages and generated 2.6 million geo-tagged flood events spanning 150+ countries. This is the training data behind Google's operational flash flood forecasting system, announced on the Google Blog and Google Research Blog.

We downloaded the full dataset, decoded every geometry, and verified the claims. Here's what we found.

2. What the Data Actually Shows

The dataset is a single 667 MB Parquet file containing exactly 2,646,302 flood events. Each event has:

What's notably absent

No country column. No language of source article. No confidence score. No link to the original news article. No event severity classification. The dataset is deliberately minimalist — just polygon geometries, dates, and areas. This makes it clean and privacy-preserving, but impossible to trace provenance or assess per-event reliability.

Geographic Distribution

We decoded all 2,646,302 WKB geometries into latitude/longitude centroids and classified by world region:

Exponential Temporal Growth

The dataset exhibits dramatic temporal skew:

2024 alone contributed 402,012 events — nearly double 2020's 198,201. This is a compound effect of more digitized global news, improved LLM extraction, and genuinely increasing flood frequency from climate change.

Event Characteristics

• Median area: 2.05 km² — extremely localized flash floods, exactly the type physical infrastructure misses
• Max area: ~5,000 km² — appears to be capped at this threshold
• 54.8% same-day events — flash floods by definition
• Max duration: 6 days — also capped; no multi-week events
• Monthly peak: July-September — Northern Hemisphere monsoon/storm season

3. Claim Verification

✅ "2.6 million geo-tagged events"

CONFIRMED. Exactly 2,646,302 events, all with polygon geometry and dates. Zero nulls, zero duplicates.

⚠️ "GDACS had roughly 10,000 entries"

⚠️ "5 million news articles across 80 languages"

CANNOT VERIFY FROM DATASET. No language column, no source article metadata, no article count. The Zenodo description says "spanning more than 150 countries" but the dataset itself provides no means to verify article counts or language coverage. The paper needs to provide this evidence.

⚠️ "22% recall and 44% precision for US NWS"

CANNOT VERIFY FROM DATASET. These are model evaluation metrics, not dataset properties. Flash flood prediction is genuinely hard — current literature puts F1-score ceilings around 0.3-0.5 for global models — so these numbers are plausible but need cross-checking against NOAA's own verification statistics.

✅ "82% practical precision" (from the paper)

The existing HF mirror's dataset card quotes the paper as reporting 82% practical precision in manual evaluations and 85-100% recall against GDACS severe events (2020-2026). These are paper claims that require the peer review process to validate.

✅ Africa coverage gap

CONFIRMED AND QUANTIFIED. Africa = 4.2% of events vs ~17% of world population. A 4× underrepresentation. More on this in Section 5.

4. The Real-Time Question

This is the most important conceptual question. The answer: Groundsource is training data, not forecast input.

The model studied 2.6 million historical events alongside the weather conditions present at each location at the time. It learned the patterns. For daily forecasting, it ingests live feeds from ECMWF, NASA, and NOAA and checks if today's weather matches a learned pattern.

The dataset doesn't need updating for real-time forecasting — just as ImageNet doesn't need daily updates for an image classifier to recognize new cats. The model learned what weather patterns precede floods from historical data. At inference, it checks if today's weather matches those patterns.

This architecture is confirmed by the RiverMamba paper, which follows the same paradigm: pretrain on GloFAS reanalysis (historical), forecast using ECMWF HRES (operational). Google's prior work (Nearing et al., Nature 2024) established the encoder-decoder LSTM architecture for global flood prediction.

That said, periodic retraining with fresh data would improve performance — especially for novel weather patterns from climate change. The Zenodo record is a v1 snapshot. Whether Google plans periodic releases remains unclear.

5. The Africa Gap: Quantified

Africa represents 4.2% of events (111,053) despite holding ~17% of world population and experiencing severe flood vulnerability. This is a 4× underrepresentation.

The paper itself acknowledges this: "Many countries in Africa are still lacking in ground truth beyond Groundsource, making it difficult to accurately estimate the accuracy of our model."

Why Africa Is Underrepresented

The gap is structural, not accidental:

1. Fewer digitized news sources. Many African news outlets aren't indexed by Google News. Local radio — the primary information medium in rural Africa — is invisible to text mining.
2. Language gap. Africa has ~2,000 languages. Even if Gemini handles 80, the long tail of African languages (Hausa, Amharic, Yoruba, Igbo, Swahili dialects) is largely uncovered. The LEMONADE dataset showed LLM extraction F1 drops severely for low-resource languages.
3. Urban reporting bias. News articles disproportionately cover urban floods. Rural flash floods affecting small communities may never appear in any outlet.
4. Digital divide. Smartphone penetration and internet access are lower, meaning fewer citizen journalism sources, fewer photos/videos that trigger news coverage.

6. The Methodology Is The Story

The most important thing about Groundsource is not the flood data. It's the demonstration that LLMs can convert the world's unstructured text into structured scientific ground truth at global scale.

The Core Recipe

The existing paradigm for scientific ground truth requires physical sensors (expensive, sparse, rich-country bias), expert annotation (slow, small-scale), or citizen science (unreliable). The Groundsource paradigm requires news articles (exist wherever humans report events) and an LLM. This is infrastructure-independent ground truth.

Related Work That Validates The Approach

This methodology has already been demonstrated in adjacent domains:

7. Where Else Can This Go?

The Groundsource pipeline is domain-agnostic in principle. Here's where it could transfer, and where it breaks down:

The Critical Insight

The methodology works best for binary, acute, widely-reported events that can be paired with continuously-available physical observations. The more the phenomenon resembles flash floods (sudden, localized, binary, widely reported), the better this approach will work.

Gold Mining / Mercury Pollution — An Interesting Case

Artisanal gold mining in Africa causes mercury pollution, deforestation, and water contamination — all poorly monitored. News articles report on illegal mining operations, environmental damage, and health effects. A Groundsource-like pipeline could create the first systematic database of artisanal mining impacts, paired with satellite change detection (deforestation, river sediment). The limitation: mining operations are often chronic (mine pollutes for years) rather than acute (flood lasts one day). The pipeline needs adaptation for long-duration events.

Air Quality — The Hybrid Approach

Direct air quality prediction from news text is limited — articles say "air quality was terrible" not "PM2.5 reached 152 μg/m³." But news-extracted pollution events (industrial accident, wildfire) could serve as supplementary features in a model that primarily uses satellite (Sentinel-5P/TROPOMI) and sensor data. The text provides the "what happened" context that satellites can't capture. The AirPhyNet paper shows physics-guided neural networks already achieve strong air quality predictions — adding event context from text could push performance further.

8. Fixing The Africa Gap: Concrete Approaches

A. Multi-Source Data Fusion Most Promising

Instead of relying solely on news, combine: Satellite SAR imagery (Sentinel-1 works everywhere), community reporting platforms (Ushahidi, WhatsApp-based reports), and local radio monitoring (transcribe and mine broadcasts in African languages). Microsoft's AI4G-Flood already mapped 10 years of global floods from Sentinel-1 SAR — this provides coverage independent of news.

B. Satellite-Only Ground Truth

The Kuro Siwo dataset provides 33 billion m² of manually annotated flood extent from SAR imagery. Fine-tuning a geospatial foundation model like TerraMind on multimodal Sentinel-1/Sentinel-2 data could generate flood ground truth for Africa without relying on text at all.

C. Synthetic Data Augmentation

SAGDA demonstrates synthetic data generation can overcome Africa's data scarcity for agriculture. The same principle could apply: generate synthetic flood scenarios for African river basins using physics-based models (LISFLOOD), then use these as additional training labels alongside sparse Groundsource events.

D. Transfer Learning

RiverMamba demonstrates this already works: pretrain globally on GloFAS reanalysis, then the model generalizes to ungauged locations including Kenya-Tanzania floods. DengueNet showed satellite imagery can predict dengue in resource-limited countries — same transfer paradigm.

E. Low-Resource Language LLM Improvement

Fine-tune extraction models specifically for African language news. Cross-lingual crisis sentence embeddings and CrisisTransformers show that crisis-domain fine-tuning dramatically improves multilingual performance. Targeted investment in Hausa, Amharic, Swahili, Yoruba extraction could significantly close the gap.

9. Tutorial: Working with the Enriched Dataset

We've published an enriched version of Groundsource on Hugging Face with decoded coordinates and derived columns. Here's how to use it:

Basic Loading

from datasets import load_dataset

ds = load_dataset("rdjarbeng/groundsource-enriched")
df = ds['train'].to_pandas()

print(f"Total events: {len(df):,}")
print(f"Columns: {list(df.columns)}")
# ['uuid', 'area_km2', 'start_date', 'end_date', 
#  'longitude', 'latitude', 'year', 'month', 
#  'duration_days', 'region']

Analyze the Africa Gap

africa = df[df['region'] == 'Africa']
print(f"African events: {len(africa):,} ({100*len(africa)/len(df):.1f}%)")
print(f"\nYearly growth in Africa:")
print(africa.groupby('year').size().tail(10))

# Compare event density per km² by region
import numpy as np
region_stats = df.groupby('region').agg(
    events=('uuid', 'count'),
    median_area=('area_km2', 'median'),
    mean_area=('area_km2', 'mean')
).sort_values('events', ascending=False)
print(region_stats)

Create a Simple Map

import matplotlib.pyplot as plt

fig, ax = plt.subplots(figsize=(15, 8))
sample = df.sample(50000, random_state=42)

# Color by region
colors = {
    'Africa': 'red', 'Europe': 'blue', 'South Asia': 'green',
    'Southeast Asia': 'orange', 'North America': 'purple',
    'South America': 'cyan', 'East Asia': 'magenta',
    'Oceania': 'brown', 'Other': 'gray'
}
for region, color in colors.items():
    mask = sample['region'] == region
    if mask.sum() > 0:
        ax.scatter(
            sample[mask]['longitude'], sample[mask]['latitude'],
            s=0.5, alpha=0.3, c=color, label=region
        )

ax.set_xlabel('Longitude')
ax.set_ylabel('Latitude')
ax.set_title('Groundsource: 2.6M Global Flood Events')
ax.legend(markerscale=10, loc='lower left')
plt.tight_layout()
plt.savefig('groundsource_map.png', dpi=150)
plt.show()

Time Series Analysis

# Monthly event counts by region over time
df['date'] = pd.to_datetime(df['start_date'])
df['yearmonth'] = df['date'].dt.to_period('M')

monthly = df.groupby(['yearmonth', 'region']).size().unstack(fill_value=0)
# Focus on recent years
monthly_recent = monthly[monthly.index >= '2015-01']

fig, ax = plt.subplots(figsize=(15, 6))
for region in ['South Asia', 'Europe', 'Southeast Asia', 
               'North America', 'Africa']:
    if region in monthly_recent.columns:
        monthly_recent[region].plot(ax=ax, label=region, alpha=0.7)

ax.set_title('Monthly Flood Events by Region (2015-2026)')
ax.set_ylabel('Events per month')
ax.legend()
plt.tight_layout()
plt.savefig('groundsource_timeseries.png', dpi=150)

Accessing the Original WKB Geometry

If you need the full polygon boundaries (not just centroids), use the original dataset mirror:

# Original with WKB geometry
ds_original = load_dataset("stefan-it/Groundsource")

# Or download directly from Zenodo
# wget https://zenodo.org/records/18647054/files/groundsource_2026.parquet

10. Experiment: Testing the Methodology on Disease Outbreaks

To test whether the Groundsource methodology actually transfers, we ran a complete replication on a different domain: epidemic surveillance from WHO Disease Outbreak News.

What We Did

1. Scraped 3,177 WHO Disease Outbreak News articles (2004-2026) via the WHO API
2. Used Qwen2.5-72B-Instruct (via HF Inference API) to extract: disease name, country, event date, case count, death count, severity
3. Geocoded extracted countries to lat/lon coordinates
4. Evaluated against both title-derived ground truth and the JRC paper's published metrics

Results

The LLM Normalizes Intelligently

The LLM doesn't just copy — it cleans and normalizes messy titles into proper disease names:

• Title: "International food safety event: Infant formula and products containing arachidonic acid oil contaminated with cereulide toxin" → LLM: "Cereulide toxin poisoning"
• Title: "Mpox: recombinant virus with genomic elements of clades Ib and IIb – Global situation" → LLM: "Mpox"
• Title: "Trends of acute respiratory infection, including human metapneumovirus" → LLM: "Acute Respiratory Infections"

Africa Coverage Flips: 50.7%

A striking finding: 50.7% of WHO DON events are in Africa — the complete opposite of the Groundsource flood dataset (4.2%). This makes sense: WHO specifically targets regions with high disease burden and weak surveillance. Top African diseases: Cholera (26), Ebola (14), Marburg (10), Yellow fever (10).

This means the methodology's Africa gap is data-source-dependent, not inherent. Choose the right text source, and the geographic bias shifts.

Sample Extractions

Measles - Bangladesh → 19,161 cases, 166 deaths (2026-04-14) severity: high
Marburg virus disease - Ethiopia → 19 cases, 14 deaths (2026-01-25) severity: critical
Cholera - Senegal → 3,475 cases, 54 deaths severity: high
Typhoid fever - DR Congo → 42,564 cases, 214 deaths severity: high

→ Dataset: rdjarbeng/who-epidemic-events (213 geo-tagged events with extraction pipeline code)

11. Resources & References

Dataset & Paper

• 📊 Enriched Dataset on HF Hub — This work: decoded coordinates, region classification
• 📊 Original HF Mirror — Raw dataset with WKB geometry
• 💾 Zenodo (Original) — CC-BY 4.0
• 📄 EarthArxiv Preprint
• 📰 Google Blog Announcement
• 🔬 Google Research Blog

Flood Forecasting SOTA

• RiverMamba — State space model for global river discharge and flood forecasting (Mamba blocks, 0.05° grid, 7-day lead time)
• Nearing et al. 2024, Nature — Global prediction of extreme floods in ungauged watersheds (the anchor paper for Google's system)
• Microsoft AI4G-Flood — 10 years of global flood mapping from Sentinel-1 SAR
• ECMWF AIFS — ECMWF's data-driven weather forecasting system

Text-to-Ground-Truth Methodology

• Epidemic IE from ProMED/WHO — LLMs for epidemic surveillance (F1 up to 0.954)
• eKG from WHO DONs — Epidemiological Knowledge Graph via ensemble LLMs
• CrisisTransformers — Pre-trained models for crisis text
• Cross-lingual crisis embeddings — Multilingual sentence encoders for crisis text

Africa Gap Solutions

• SAGDA — Synthetic Agriculture Data for Africa
• DengueNet — Satellite-based disease prediction for resource-limited countries
• Kuro Siwo — 33B m² annotated SAR flood data
• TerraMind/FloodsNet — Geospatial foundation models for flood mapping

Air Quality & Other Domains

• AirPhyNet — Physics-guided air quality prediction
• UK Air Pollution Gap-Filling — ML framework for monitoring network gaps
• AirCast — Multi-variable air pollution forecasting
• IncidentAI — NER from industrial safety incident reports