Amazon Mining Watch

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, to summarize, 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

Current data include yearly assessments of mining activity starting from 2018, and both yearly and quarterly estimates from 2025. The models make probabilistic assessments on 480m x 480m square patches of the Earth's surface as to whether traces of mining activity are visible within a patch. The models assess over 100 million patches at each time step across the Amazon basin.

cumulative_detections/ are the data on the Amazon Mining Watch website. The detections are registered yearly or quarterly, and aggregated across years beginning from 2018.

For instance, the file Amazon_ACA_48px_v4.10b-18d-20g-21a-22bc-ensemble_t0.55_d5_3km_t-iso0.8_cumulative2018-2023.geojson is a record of all mining scars detected by our models in years 2018 - 2023, inclusive.

Amazon_ACA_48px_v4.10b-18d-20g-21a-22bc-ensemble_t0.55_d5_3km_t-iso0.8_cumulative2018-Q226.geojson aggregates yearly data from 2018 - 2025 and the first two quarters of 2026.

Mine scars and their appearance in satellite imagery change over time, making detection variable. To our mind, cumulative detections best establish an historical record of lands that have been mined since the beginning of the monitoring period in 2018.

--

The individual yearly or quarterly assessments are in the single_periods/ subfolders:

• raw_detections/ are direct outputs of the machine learning model, for all patches meeting a minimum confidence score of 0.4.

Mining activity clusters in space. Following that observation, we impose a higher confidence threshold on isolated candidate detections, as a postprocessing step to reduce scattershot noise.

• postprocessed_t0.43_d5_3km_t-iso0.75/ uses relatively relaxed thresholds, tuned for single-year detections.

• postprocessed_t0.55_d5_3km_t-iso0.8/ use more stringent thresholds for the cumulative detections presented on the website. (Coverage and errors compound year to year.)

At lower confidence thresholds, the model captures more mining activity, at the cost of additional false detections. These additional views on the data accommodate different preferences with respect to that tradeoff.

archived/ data were published here in January, 2026, from an experimental model that has since been superceded 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.

Finally, mining_scar_masks/ give pixelwise estimates of the delinated boundaries of mining scars, based on 10m Sentinel-2 pixels. The masks are used for mined-area calculations on Amazon Mining Watch.

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 surronding the cumulative detections.

Contact

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