IRSC/ELDOR

ELDOR Data

• —Public dataset name: ELDOR
• —Paper: ELDOR: A Dataset and Benchmark for Illegal Gold Mining in the Amazon Rainforest
• —PDF: https://arxiv.org/pdf/2605.15397
• —Local dataset root: GoldMDD/data
• —Output folders: train/image, train/label, val/image, val/label, test/image, test/label (copied from GoldMining/Data/Orthomosaic and GoldMining/Data/Label).
• —Source folders: Drone_Orthomosaic_V1/Orthomosaic_org and Drone_Orthomosaic_V1/Just_labels
• —Matching rule: label to orthomosaic by filename mapping, then crop/reproject label to orthomosaic grid.
• —Canonical labels: 14 semantic classes (1..14) plus 0=Background after class merging in ELDOR.
• —Split rule in this dataset: train/val/test assignment is listed in the Split column of the spatial metadata table.

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}
}

Spatial metadata table (source files + aligned output)

Unified class mapping (ELDOR merged classes)

Label color plan (6-digit HEX)

Per-site classes using ELDOR merged mapping

Generation workflow

• —Source orthomosaic/label alignment and original canonical labels come from GoldMining/Data.
• —ELDOR labels are remapped from the original 19-class canonical IDs into a 14-class merged scheme (stored only under GoldMDD/data/*/label).
• —Merge rules:
• —Heavy machinery = original IDs 5 (Front loader), 8 (Excavator), 9 (Heavy machinery), 19 (Dump truck)
• —Vehicles = original IDs 17 (Vehicles), 18 (Small vehicles)
• —Building = original IDs 1 (Urban area), 4 (Mining camp)
• —Heatmap output: GoldMDD/site_class_pixel_counts_heatmap_merged.png
• —Heatmap CSV: GoldMDD/site_class_pixel_counts_merged.csv
• —Train/val/test distribution plot: GoldMDD/train_val_test_class_distribution_merged.png
• —Train/val/test distribution CSV: GoldMDD/train_val_test_class_distribution_merged.csv

Cropped patch dataset (data-cropped)

• —Output root: GoldMDD/data-cropped
• —Patch size: 512x512
• —Stride: 256
• —Filtering rule: drop a patch if background pixels in the merged label (label==0) are >80%.
• —Windowing rule: full windows only (no padding).
• —Patch naming: matching basenames per pair, e.g. image AcumulacionAaron2B_2_3.jpg and label AcumulacionAaron2B_2_3.png (1-based row/col indices).
• —Storage format: image patches = JPEG, label patches = PNG (lossless class IDs).
• —Folder structure: train/image, train/label, val/image, val/label, test/image, test/label.

Crop summary by split

Crop summary by site

• —Summary CSVs:
• —GoldMDD/data-cropped/crop_summary_by_split.csv
• —GoldMDD/data-cropped/crop_summary_by_site.csv