IRSC/ELDOR-sample

ELDOR Sample

This repository is a compact sample companion to the full ELDOR dataset:

• Full dataset: https://huggingface.co/datasets/IRSC/ELDOR
• Paper: https://arxiv.org/abs/2605.15397
• It is intended for quick inspection of the imagery, labels, and released patch extraction without downloading the full release.

Citation

@article{cui2026eldor,
  title={ELDOR: A Dataset and Benchmark for Illegal Gold Mining in the Amazon Rainforest},
  author={Cui, Kangning and Bohara, Surendra and Prasai, Suraj and Shao, Zishan and Tang, Wei and Pillaca, Martin and Flores, Edwin and Yang, Zhen and Larsen, Gregory and Dethier, Evan and Lutz, David and Morel, Jean-Michel and Silman, Miles and Pauca, Victor and Yang, Fan},
  journal={arXiv preprint arXiv:2605.15397},
  year={2026}
}

What this sample contains

This sample contains one representative site and all of its released patches:

• Representative site: AcumulacionAaron2B (train split)
• Site-level aligned orthomosaic image and pixel-aligned label map
• All released 512x512 patches derived from that site with 256 stride and the public crop filter used in ELDOR

Two dataset configurations are exposed:

• site: the full aligned site-level image/label pair
• patches: every released patch derived from that site

How the sample was created

The sample was created deterministically from the public ELDOR release:

1. We selected AcumulacionAaron2B because it has the broadest per-site class coverage in the public release: 12 of the 14 foreground classes.
2. We copied the exact released aligned orthomosaic PNG and label PNG for that site from the full dataset.
3. We copied every released patch derived from that site from the public data-cropped export without additional filtering beyond the original ELDOR crop pipeline.

The public crop pipeline used for ELDOR is:

• patch size: 512x512
• stride: 256
• full windows only, no padding
• drop a patch if background exceeds 80%
• patch images stored as JPEG and patch labels stored as PNG

This sample includes the following foreground classes:

• Building
• Mining raft
• Primary Forest
• Heavy machinery
• Water bodies
• Agricultural crop
• Gravel mounds
• Type 1 natural regeneration
• Type 2 natural regeneration
• Bare ground
• Sluice
• Vehicles

Foreground classes present in the full dataset but not in this single-site sample:

• Compact mounds
• Grass

Sample size summary

• Site subset: 1 aligned image/label pair, 1.481 GB
• Patch subset: 9849 image patches and 9849 label patches, 1.352 GB
• Total sample size: 2.832 GB

File organization

site configuration

• site/images/AcumulacionAaron2B.png
• site/labels/AcumulacionAaron2B.png
• site/metadata.csv

patches configuration

• patches/images/AcumulacionAaron2B_<row>_<col>.jpg
• patches/labels/AcumulacionAaron2B_<row>_<col>.png
• patches/metadata.csv

Patch filename indices are 1-based row and column indices in the released crop grid.