Galeio

Galeio is a company specializing in developing custom foundation models for science and industrial applications.

🛰️ Our Models

The OceanSAR Family 🌊

Our OceanSAR models are specialized in ocean observation. The models were developed in collaboration with world-leading SAR experts from Ifremer.

Current Models

OceanSAR-1: Our super lightweight foundation model
- Trained on 10 years of Sentinel-1 Wave Mode data
- State-of-the-art performance on ocean analysis tasks
- Available architectures:
  - ResNet50 (75.5% TenGeoP accuracy in linear probing)
  - ViT-S/16 (78.6% TenGeoP accuracy in linear probing) (on-demand)
  - ViT-S/8 (82.1% TenGeoP accuracy in linear probing) (on-demand)
  - ViT-B/8 (83.6% TenGeoP accuracy in linear probing) (on-demand)
- Excels at:
  - Wind speed prediction (RMSE: 1.37 m/s in linear probing | best model ViT-B/8)
  - Wave height estimation (RMSE: 0.63 m in linear probing | best model ViT-B/8)
  - Geophysical phenomena classification
  - Many others (benchmarks coming soon)

OceanSAR Family

The OceanSAR model family is dedicated to ocean observation and analysis using SAR imagery.

OceanSAR-1: Our first foundation model trained on Sentinel-1 Wave Mode data:
- Specialized for ocean SAR imagery analysis
- Trained using dynamic dataset pruning for optimal performance
- Available in multiple architectures (ResNet50, ViT variants)

🔬 Research

Our research focuses on:

Self-supervised learning for Earth Observation
Dynamic dataset curation techniques
SAR image analysis
Ocean monitoring

✉️ Contact & Support

🌐 Website: galeio.fr
📧 Email: contact@galeio.fr
📍 Location: Paris, France
💼 LinkedIn: Galeio

📚 Citations

If you use our models in your research, please cite:

@article{kerdreux2025efficientselfsupervisedlearningearth,
      title={Efficient Self-Supervised Learning for Earth Observation via Dynamic Dataset Curation}, 
      author={Thomas Kerdreux and Alexandre Tuel and Quentin Febvre and Alexis Mouche and Bertrand Chapron},
      year={2025},
      eprint={2504.06962},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2504.06962}, 
}