Enhanced Parquet Edition: RSFAKE-1M: A Large-Scale Dataset for Detecting Diffusion-Generated Remote Sensing Forgeries

This repository packages the official RSFAKE-1M release into Hugging Face Parquet configurations. It contains 1,000,000 detection samples: 500,000 real patches and 500,000 generated images. The additional fmow-sources configuration contains 421,596 referenced fMoW source images for provenance, separate from the detection samples. Original JPEG/PNG bytes are embedded in Parquet image columns at their existing resolutions.

Subsets

Config	Split	Rows	Image size	Parquet files
rsfake1m	train	600,000	Original sizes	121
rsfake1m	val	200,000	Original sizes	41
rsfake1m	test	200,000	Original sizes	41
fmow-sources	full	421,596	Original sizes	592

Columns

rsfake1m:

• image: original sample image, decoded with the Hugging Face Image feature.
• sample_id, source_image_path: original image path relative to the RSFAKE release root.
• label: image-level authenticity label, 0 = real, 1 = fake.
• dataset_name: original source or generator name.
• source_split, split_index: original RSFAKE split and zero-based CSV row position.
• fmow_source_id: join key matching fmow-sources.source_id for real samples.
• crop_x, crop_y, crop_size: original crop coordinates and square size in pixels.
• image_bytes: encoded original file size in bytes.

The fMoW source and crop fields are null for generated images.

fmow-sources:

• image: original fMoW source JPEG, at its existing resolution.
• source_id: source path as recorded in the crop CSV, beginning with fmow/.
• source_split: native fMoW split; source_image_path: path relative to the RSFAKE release root.
• sample_count: total number of selected real patches referencing this source.
• rsfake_train_samples, rsfake_val_samples, rsfake_test_samples: per-split reference counts.
• image_bytes: original JPEG file size in bytes.

Metadata Files

Each configuration includes stats.json with split counts, labels and source counts where applicable, and Parquet shard paths and sizes.

The root inventory.json collects these statistics and records the total row count, shard count, and encoded byte sizes.

provenance/ retains the original split CSVs, crop CSVs, cropping script, and upstream README.

Citation

If you reference RSFAKE-1M, please cite the original paper:

@misc{tan2025rsfake1mlargescaledatasetdetecting,
  title={RSFAKE-1M: A Large-Scale Dataset for Detecting Diffusion-Generated Remote Sensing Forgeries},
  author={Zhihong Tan and Jiayi Wang and Huiying Shi and Binyuan Huang and Hongchen Wei and Zhenzhong Chen},
  year={2025},
  eprint={2505.23283},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}

Dataset Export

This Parquet edition reorganizes the authors' released data and adds per-sample metadata and links to the corresponding fMoW source images. Please also reference this dataset repository when using the packaged version.

@misc{thanejoss_rsfake1m_2026,
  author={{ThaneJoss}},
  title={{RSFAKE-1M}},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/ThaneJoss/RSFAKE-1M},
  note={Dataset available on Hugging Face}
}