Pan-ASGM

Pan-ASGM is a multi-temporal Sentinel-2 dataset of artisanal and small-scale gold mining (ASGM). It covers nine mining sites in five countries, with one image per year for 2016-2025 and seven-class labels for 2016, 2020, and 2024. The labels separate active ore-washing ponds from inactive ones.

Download

hf download oryota/pan-asgm --repo-type dataset --local-dir data/pan_asgm

import cv2, tifffile
image = tifffile.imread("data/pan_asgm/images/LaPampa9/LaPampa9_2020_r1c2.tif")  # (13, 200, 200) uint16
mask = cv2.imread("data/pan_asgm/masks/LaPampa9/LaPampa9_2020_r1c2.png", 0)     # (200, 200) uint8

Files

images/{site}/{site}_{year}_{tile}.tif   Sentinel-2 L1C, 13 bands (B01-B12, B8A after B08), 10 m, 200 x 200 px
masks/{site}/{site}_{year}_{tile}.png    labels for 2016, 2020, 2024
splits/loso.yaml                         leave-one-site-out folds (fold -> test site)
metadata.csv                             site, country, year, tile, composite month

Each site is a 6 km x 6 km area split into 3 x 3 tiles (r{row}c{col}). Image values are top-of-atmosphere reflectance x 10000.

Mask values: 0 = unlabeled, 1 = vegetation, 2 = bare_ground, 3 = active_pond, 4 = inactive_pond, 5 = artifact, 6 = cloud, 7 = cloud_shadow.

See the paper for how the data were collected and labeled.

Known limitations

• Nine sites and three labeled years. Mining types and regions outside these sites are not covered.
• Tiles carry no geographic coordinates, and site locations are given only at the country level.
• artifact is very rare (0.1% of labeled pixels).

License

The labels, splits, and metadata are released under CC BY-NC 4.0. The images contain modified Copernicus Sentinel data (2016-2025) and are subject to the Copernicus data terms.

Citation

@inproceedings{okumura2026tracking,
  title     = {Tracking Gold Mining Activity Behind Tropical Deforestation with Satellite Time Series},
  author    = {Okumura, Ryota and Hirose, Kazuyo and Wang, Binze and Matsubae, Kazuyo and Yamakata, Yoko},
  booktitle = {NeurIPS 2026 Workshop on Tackling Climate Change with Machine Learning},
  year      = {2026}
}