fmow-fake-small: a prototype dataset for remote sensing deepfake and image forgery localization

Dataset Details

Dataset Description

This dataset consists of 30 real and 30 manipulated satellite images. For the fake images, three types of image manipulations are used including random copy-paste splices, object copy-paste splices, and diffusion model inpainting. For the diffusion model inpainting, we use the pretrained diffusion model RSPaint.

• Curated by:
• Language(s) (NLP): English
• License: cc-by-4.0

Dataset Sources [optional]

• Repository: https://huggingface.co/datasets/geodf/fmow-fake-small

Uses

Direct Use

This dataset is intended to be used for image forensics research as well as evaluating deepfake detection and image forgery detection models. It includes ground truth masks indicating pixels that have been manipulated or modified by copy-paste splices or diffusion model inpainting. With the ground truth masks, one can evaluate deepfake and image forgery localization methods.

Out-of-Scope Use

Dataset may not be suitable for training detection models.

Dataset Structure

The dataset is organized as follows:

fmow-fake-small/
├── train/                            # train split
│   ├── real/                         # real images
│   │   ├── {filename}_img.tif        # RGB image
│   │   ├── {filename}_seg.tif        # ground truth mask
│   │   └── {filename}_meta.json      # metadata for RGB image
│   └── fake/                         # Fake images 
│       ├── simple-splice/            # random simple-splice/copy-paste samples
│       │   ├── {filename}_img.tif
│       │   ├── {filename}_seg.tif
│       │   └── {filename}_meta.json
│       ├── object-splice/            # object-splice/copy-paste samples
│       │   ├── {filename}_img.tif
│       │   ├── {filename}_seg.tif
│       │   └── {filename}_meta.json
│       └── inpaint/                  # diffusion model inpainting samples
│           ├── {filename}_img.tif
│           └── {filename}_seg.tif
├── val/                             # validation split
│   ├── real/
│   │   ├── {filename}_img.tif
│   │   ├── {filename}_seg.tif
│   │   └── {filename}_meta.json
│   └── fake/
│       ├── simple-splice/
│       │   ├── {filename}_img.tif
│       │   ├── {filename}_seg.tif
│       │   └── {filename}_meta.json
│       ├── object-splice/
│       │   ├── {filename}_img.tif
│       │   ├── {filename}_seg.tif
│       │   └── {filename}_meta.json
│       └── inpaint/
│           ├── {filename}_img.tif
│           └── {filename}_seg.tif
│
└── metadata.csv

A single sample consists of the following assets:

• suffix _img.tif is the RGB image. 8-bit, 3 channel.
• suffix _seg.tif is the mask image. 8-bit, 1 channel (pixel value 0 = non manipulated, pixel value 255 = manipulated)
• suffix _meta.json is the image acquisition metadata for source and base image

Dataset Creation

Curation Rationale

There is a lack of existing remote sensing image forgery localization datasets. This dataset attempts to fill this gap to enable research in deepfake localization.

Source Data

The dataset is derived from the functional map of the world (fmow) dataset: https://github.com/fMoW/dataset. We use the three channel (RGB), 8-bit, pansharpened images in the fMoW training set.

Data Collection and Processing

For real images, we randomly sample 30 images from the fmow training set. For fake images, we randomly sample from the fmow training set and manipulate them using copy-paste splice from different images in the fmow training set or using a pretrained diffusion model.

Who are the source data producers?

The dataset is derived from the fmow dataset. We augment this dataset by providing fake images.

Bias, Risks, and Limitations

Recommendations

Citation

@article{arndt2026towards,
  title={Towards a satellite image manipulation and deepfake localization benchmark dataset},
  author={Arndt, Jacob and Varshney, Debvrat and Dias, Philipe and Nukavarapu, Nivedita},
  journal={arXiv preprint arXiv:2608.04840},
  year={2026}
}

Glossary

Dataset Card Contact

Jacob Arndt, arndtjw@ornl.gov