Amazon Mining Watch

Amazon Mining Watch maps the spread of artisanal and small-scale gold mining, legal and illegal, across all nine Amazonian countries. Earth Genome builds its models and data in partnership with the Pulitzer Center's Rainforest Investigations Network and Amazon Conservation Association, for journalists, activists, researchers, policymakers and local organisations working to understand and curb mining's impacts.

The detector is an ensemble of convolutional neural networks trained from scratch on 13-band Sentinel-2 imagery. It scores 480 m x 480 m patches across the Amazon basin, more than 100 million per period, using cloud-reduced median composites built on Google Earth Engine. A fine-tuned SAM2 segmentation model then delineates mining scars at 10 m resolution to estimate the area affected. Detections run yearly from 2018 and quarterly from 2025: a detection is confirmed once it reappears in a second annual period, and recent quarterly detections are marked provisional. Earth Genome reports 99% specificity and 90% sensitivity at working thresholds on its labelled test sets.

The web map shows mining over time by country, subnational jurisdiction, protected area and Indigenous territory, with an illegality-risk breakdown and socio-environmental cost estimates from Conservation Strategy Fund's Mining Impacts Calculator, in English, Spanish and Portuguese. The detections (GeoJSON), mining-scar masks (GeoTIFF) and model code are free to download from Source Cooperative and GitHub under CC BY 4.0 and MIT licences.