Amazon Mining Watch

Updated 20 August 2026. Model 48px_v4.10b-18d-20g-21a-22bc-ensemble, annual periods 2018–2025 and quarters Q125–Q226.

This repository contains automated detections of artisanal gold mine scars in Sentinel-2 satellite imagery.

The data is licensed under the creative commons 4.0 international license which, in essence, only requires attribution.

Code, description of the machine-learning models that generated the detections, and additional versions of this data are available at the Amazon Mining Watch Github repository.

Data

Coverage is yearly from 2018. The current and previous calendar year are covered by provisional quarterly detections.

Patch detections

Patch detections are the primary product. They are derived from a machine learning model that makes probabilistic assessments on 480 m × 480 m patches of the Earth's surface as to whether traces of mining activity are visible within a patch. The model canvases over 100 million patches across the Amazon basin at each time step.

This file replaces the per-period cumulative files published earlier in 2026, which have been removed.

Mining scar masks

The mining scar raster masks are still in development. They provide mine footprints and mined area estimates for the Amazon Mining Watch website.

Confirmed and provisional detections

The patch detections are a cumulative historical record of lands that have been mined since the beginning of monitoring in 2018. A detection enters the record as confirmed once it appears in two annual periods within a two-year window. The period recorded is the first of those. A confirmed detection is never withdrawn by a later update.

Recent quarterly data are marked provisional and are held to a stricter confidence threshold. Provisional data will be replaced by annual data once the corroborating year arrives.

For the mining scar masks, pixel values indicate the time period. For example, 2024 takes the raster value 2024 and Q1 2025 takes the value 20251. Zero indicates no confirmed mining. The two-year persistence rule is applied to annual data also at the pixel level.

Segmentation is the least mature part of this pipeline and areas are known to run large compared to hand-drawn boundaries. Comparisons between periods are more trustworthy than any single total. More details can be found on our Github repository.

Single-period data

The individual yearly and quarterly assessments are for users who want the unaggregated model output:

• raw_detections/ — direct model output, every patch at or above a confidence of 0.4, before any filtering.
• postprocessed/ — the recommended single-period product. Mining activity clusters in space, so isolated candidate detections are held to a higher confidence than clustered ones (t_main = 0.43, t_iso = 0.75). This is also the set the segmentation model is prompted from.

Archive

archived/ data were published here in January, 2026, from an experimental model that has since been superseded by the current model version. We include them here for the record. The data have the limitation and the virtue of having been extensively cleaned by hand prior to publication.

Models

The primary mining activity detector is an ensemble of convolutional neural networks, trained on a dataset of approximately 20,000 locations labeled as mines or mine-free areas. This model was wholly redeveloped in 2026 and the data recomputed starting from year 2018.

Mining scar masks are computed by a fine-tuned SAM2 (Segment Anything) model applied in the near field of view surrounding the cumulative detections.

Contact

Using this? Want to work together? Ping us @ info@earthgenome.org or info@earthindex.ai