zocar5807/sar-mining-detector

☿ Freddy Hg — SAR Mining Detector

Open-source detector of illegal fluvial gold dredges in the Colombian Amazon, based on Sentinel-1 SAR imagery and Google Earth Engine.

Model description

Freddy Hg detects metallic dredges over Amazon rivers using a thresholded backscatter algorithm on Sentinel-1 SAR GRD (C-band, VV polarization). Dredges (15–25 m vessels with metallic decks and dredging gear) act as dihedral corner reflectors, returning 17–30 dB more backscatter than the specular water surface around them.

Repository: https://github.com/ByZocar/Freddy-Hg

The system enriches each detection with:

1. ANM geofencing — checks if the point falls within an active mining concession (legal vs. presumptive illegal).
2. RAISG indigenous territory check — escalates alerts that intersect protected indigenous reserves (T-106/25 protocol).
3. Confidence level (1/2/3) combining SAR signal, historical change detection, and indigenous overlay.
4. SHA-256 chain of custody of source data + metadata for digital evidence admissibility under Colombian Law 1333/2009.
5. Mistral AI contextual enrichment — adds a natural-language summary and proxy impact metrics (Hg estimated, USD damage, people at risk).

This is not a deep learning model with downloadable weights. It is a deterministic geospatial pipeline whose code is open-source. The "model card" exists to document its algorithm, intended use, evaluation, and known limitations following the same transparency standards used for trained models.

Intended use

Primary use case

• Early-warning detection of illegal fluvial gold mining for environmental authorities (CARs) in the Colombian Amazon.
• Generation of legally-admissible evidence (PDF + SHA-256) for sanctioning procedures under Law 1333/2009.
• Push notifications via WhatsApp/SMS to indigenous guardians.

Out-of-scope use cases

• Terrestrial mining detection — Freddy Hg targets fluvial dredges specifically. Use Amazon Mining Watch or MapBiomas for deforestation signatures.
• Sub-meter detection — Sentinel-1 resolution is 10 m. Objects smaller than ~200 m² cannot be reliably detected.
• Real-time detection — Sentinel-1 revisit is 6 days (Sentinel-1C restored). The system is near-real-time, not live.
• Automatic legal action — alerts must be field-verified by competent authorities before any sanction is issued.

Algorithm overview

# Pseudocode of the core detection
roi = ee.Geometry.Rectangle([xmin, ymin, xmax, ymax])

# 1. Load Sentinel-1 IW GRD, VV polarization, last 60 days
s1 = (ee.ImageCollection("COPERNICUS/S1_GRD")
      .filterBounds(roi)
      .filterDate(start, end)
      .filter(ee.Filter.eq("instrumentMode", "IW"))
      .select("VV"))

# 2. Median composite reduces speckle and transient noise
median_current = s1.median()

# 3. Water mask: pixels where VV < -15 dB (specular reflector)
water_mask = median_current.lt(-15.0)

# 4. Bright targets: pixels where VV > -12 dB (corner reflector)
bright = median_current.gt(-12.0)

# 5. Candidates: bright AND on water AND ≥ MIN_PIXELS contiguous
candidates = bright.And(water_mask)
min_size_mask = candidates.connectedPixelCount(MIN_PIXELS + 1).gte(MIN_PIXELS)
candidates_filtered = candidates.And(min_size_mask)

# 6. Change detection vs 2018-2019 historical baseline
s1_baseline = (ee.ImageCollection("COPERNICUS/S1_GRD")
               .filterBounds(roi)
               .filterDate("2018-01-01", "2019-12-31")
               .select("VV")).median()
change = median_current.subtract(s1_baseline)
new_activity = change.gt(5)

# 7. Vectorize centroids of candidate clusters
detections = candidates_filtered.reduceToVectors(geometryType="centroid")

Full implementation: github.com/ByZocar/Freddy-Hg/blob/main/pipeline/freddy_detection.py

Training data

This is not a trained model. The detection rules are based on:

• Physical principles of SAR backscatter on dihedral corner reflectors in water environments.
• Empirical thresholds from peer-reviewed literature:
  • Schwartz et al. (2019) — Madeira basin, Brazil
  • Goncalves et al. (2021) — Tapajós basin, Brazil

The pilot ROIs used for system testing include ground truth from:

• MAAP #228 (May 2025): 56 dragas documentadas en el río Puré.
• Armada Nacional de Colombia (Feb–Mar 2026): 27 dredges destroyed in the Atrato river, providing position references.
• Sentencia T-106/25 Constitutional Court: zones affected in the Jaguares del Yuruparí macro-territory (Caquetá and Apaporis rivers).

Evaluation metrics

Empirically validated metrics for the Colombian Amazon basin are not yet available. The system is in pilot. Proxy metrics from analogous studies (Brazilian Amazon, same sensor, same vessel type):

Source	Region	TPR	FPR	Threshold
Schwartz et al. 2019	Madeira	78%	22%	VV > -10 dB
Goncalves et al. 2021	Tapajós	74%	31%	VV > -10 dB

Calibration work in progress with Corpoamazonía and CDA Guainía pilots. First field-validation report expected Q3 2026.

Limitations and biases

Known false positives

Source	Mitigation
Sandbars in braided rivers (dry season)	Historical baseline (2018-2019) change detection
Legitimate boats in transit	Median multi-temporal compositing
Permanent riverside infrastructure (docks, camps)	Excluded by baseline comparison
Layover in narrow forested canyons	Documented; affected pixels excluded
Flooded vegetation with double-bounce	Strict water mask (VV < -15 dB)

Known false negatives

• Dredges smaller than ~200 m² — below Sentinel-1 resolution.
• Dredges actively masked by tree canopy — possible in narrow tributaries.
• First-pass dredges in zones not covered by baseline — change detection misses them in the first 6-day cycle.

Geographic bias

The system has only been tuned for three Colombian Amazon basins:

• Caquetá / Apaporis (Amazonas-Caquetá)
• Inírida (Guainía)
• Atrato (Chocó)

Performance outside these basins is unknown. Different river morphology, sediment loads, or vegetation density may require threshold recalibration.

Environmental bias

The model implicitly assumes:

• Tropical wet climate with characteristic Amazon hydrology.
• Dredges with metallic decks ≥ 200 m².
• No prior major hydraulic interventions (large dams change SAR signature).

Ethical considerations

This model identifies coordinates where illegal activity may be occurring. That data is sensitive:

• Field verification is mandatory before sanctions or armed interventions — false positives could expose innocent communities.
• Indigenous territory alerts (Level 3) activate human-rights protocols. The system never replaces FPIC (Free, Prior and Informed Consent) procedures.
• Guardian alerts are sent without storing recipient phone numbers identifiable in logs. The hash → secret-ref design protects guardians in conflict zones.
• The pipeline is open-source to allow independent audit by journalists, prosecutors, and academia.

Recommendations

For environmental authorities (CARs):

• Always verify in field before issuing cautionary measures.
• Conserve the SHA-256 hash from the PDF as legal proof.
• Document confirmed false positives back to the system for retraining.

For journalists and researchers:

• Public alerts at /public are 30 days delayed (CC BY 4.0).
• Cite by alert ID + SHA-256 + Sentinel-1 scene ID.
• Read /docs/accuracy before publishing impact figures.

Citation

@software{freddy_hg_2026,
  title  = {Freddy Hg: SAR-based detection of illegal fluvial gold dredges
            in the Colombian Amazon},
  author = {Cardozo, Andrés Felipe and contributors},
  year   = {2026},
  url    = {https://github.com/ByZocar/Freddy-Hg},
  license = {Apache-2.0},
  version = {1.1.0}
}

Contact and contributions

• Repository: https://github.com/ByZocar/Freddy-Hg
• Live demo: https://freddy-hg.vercel.app
• API documentation: https://freddy-hg-backend-production.up.railway.app/docs
• Issues: https://github.com/ByZocar/Freddy-Hg/issues

Contributions are welcome under Apache 2.0. See CONTRIBUTING.md in the repository for guidelines.

Acknowledgments

• ESA / Copernicus for free open access to Sentinel-1 SAR imagery.
• Google Earth Engine for free compute for conservation projects.
• ANM Colombia for the public mining concession registry.
• RAISG for the indigenous territory polygons of the Amazon basin.
• Mistral AI for context enrichment via OpenRouter.
• Corpoamazonía, CDA Guainía, FCDS, Gaia Amazonas, OPIAC as pilot partners.

---

Last updated: May 2026 · v1.1.0