SatDiff: A Stable Diffusion Framework for Satellite Imagery Inpainting 🌍

Teerapong Panboonyuen

AGL (Advancing Geoscience Laboratory), Chulalongkorn University

MIT License

SatDiff leverages the power of Stable Diffusion Models to intelligently inpaint missing or corrupted regions in very high-resolution satellite imagery. Designed specifically for Earth observation, it preserves both texture and geospatial semantics—making it ideal for applications where visual and structural integrity matter most.

Built at the Advancing Geoscience Laboratory (AGL), our model incorporates a custom dual-branch architecture with attention-enhanced modules, tailored for aerial and satellite domains. It significantly outperforms classical and deep-learning inpainting approaches in PSNR, SSIM, and perceptual realism.

  • PSNR: 35.72 dB
  • SSIM: 0.952
  • Perceptual Realism: High structural fidelity & semantic awareness

📄 IEEE 2025 Publication

SatDiff Inpainting Result

🏗️ Architecture Overview

SatDiff Overview Dual-Branch Attention Design Diffusion Model Structure

🔧 Installation

git clone https://github.com/kaopanboonyuen/SatDiff.git
cd SatDiff
pip install -r requirements.txt

🚀 Usage

1. Prepare Configuration

Edit the config.yaml file with dataset paths and hyperparameters.

2. Train Model

python train.py

3. Evaluate Results

python evaluate.py

4. Run Inference

python inference.py --image_path path/to/image.png

📊 Example Output

SatDiff Output Example

🧑‍🔬 About the Author

Author Bio

📖 Citation

If you use SatDiff in your research, please cite the following paper:

@article{panboonyuen2025satdiff,
  title={SatDiff: A Stable Diffusion Framework for Inpainting Very High-Resolution Satellite Imagery},
  author=Panboonyuen, Teerapong, et al.,
  journal={IEEE Access},
  year={2025},
  publisher={IEEE}
}

🌐 Project Repository

View the source code and contribute at: 🔗 GitHub Repository