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

Dataset Summary

RSFAKE-1M is a large-scale dataset designed to advance the detection of forged remote sensing images, particularly those generated by diffusion models. It contains 1 million images in total — 500K real and 500K fake. The forged images are produced using 10 different diffusion models fine-tuned on remote sensing data, spanning six generation conditions, including text-to-image, structure-guided generation, and inpainting.

Remote sensing imagery plays a vital role in areas such as environmental monitoring, urban planning, and national security. However, with the rapid development of generative models, especially diffusion-based architectures, remote sensing images are increasingly vulnerable to realistic forgeries. Despite this, most existing benchmarks focus on GAN-based or natural image forgeries, leaving a critical gap in the remote sensing domain.

RSFAKE-1M addresses this gap by offering a comprehensive benchmark for training and evaluating forgery detection models under realistic and diverse conditions. Extensive experiments in our accompanying paper demonstrate that:

• Current state-of-the-art detectors struggle with diffusion-generated forgeries in remote sensing.
• Training on RSFAKE-1M significantly improves generalization and robustness across different forgery types.

We believe RSFAKE-1M serves as a solid foundation for the development of next-generation remote sensing forgery detection algorithms.

Dataset Structure

RSFAKE/
├── FAKE/
│   ├── generated_crsdiff
│   ├── generated_diffusion_sat
│   ├── generated_diffusion_sat_256
│   ├── generated_geosynth
│   ├── generated_geosynth_canny
│   ├── generated_geosynth_sam
│   ├── generated_mapsat
│   ├── generated_rsinpaint
│   ├── generated_RSSD_768
│   ├── generated_SDFRS
│
├── REAL/
│   └── fmow/
│       ├── train/
│       ├── val/
│       └── test/
│
├── SPLIT/
│   ├── RSFAKE_train_new.csv
│   ├── RSFAKE_val_new.csv
│   └── RSFAKE_test_new.csv

Real Image Construction

The real image subset is reconstructed from the publicly available fMoW dataset. To reproduce the real subset:

1. Download the original fMoW-rgb dataset to the REAL/fmow_process/ directory.

2. Prepare the environment:

   pip install pillow==11.2.1 pandas==2.2.3 tqdm==4.67.1
   

3. Run the cropping script:

   cd REAL/fmow_process/
   python crop.py
   

The processed output will be structured into train, val, and test under REAL/fmow/.

These scripts ensure that the real image set used for RSFAKE-1M evaluation is consistent and reproducible, while respecting the original data source’s license.

Disclaimer

RSFAKE-1M is a synthetic benchmark designed to facilitate research on forgery detection in remote sensing. The fake images are artificially generated and do not correspond to real-world scenes or locations. They must not be used for any purpose that could mislead, misinform, or be interpreted as real satellite or aerial data.

All model-generated content is based on publicly available generative models listed below. RSFAKE-1M does not distribute or modify the original models themselves — only images produced under fair-use conditions are included.

By using this dataset, you agree:

• The real images are reconstructed from the publicly available FMoW dataset and remain subject to its original license.
• The forged images are generated using publicly available diffusion models, whose licenses we fully acknowledge.
• We do not claim ownership of any third-party models or datasets used in RSFAKE-1M.
• This dataset is provided strictly for non-commercial research and educational use.
• Users must cite the RSFAKE-1M paper and comply with the licenses of all referenced resources.
• The authors bear no responsibility for any misuse or downstream consequences related to this dataset.

Citation

@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},
      url={https://arxiv.org/abs/2505.23283},
}

Acknowledgements

We would like to thank the creators of the following models and datasets, which served as the basis for generating the fake and real images in RSFAKE-1M:

• 🛰 Real Image Source

  • fMoW (Functional Map of the World) (FUNCTIONAL MAP OF THE WORLD CHALLENGE PUBLIC LICENSE)

• 🧨 Diffusion-Based Generative Models

  • CRSDiff
  • Diffusion-SAT (Apache 2.0)
  • GeoSynth (Apache 2.0)
  • MapSat (Apache 2.0)
  • RSPaint (Apache 2.0)
  • SDFRS (MIT License)
  • GeoRSSD (Apache 2.0)

We sincerely acknowledge the contributions of the above works. This dataset would not have been possible without their efforts in advancing generative modeling in the remote sensing domain.