README.md

# 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](https://en.ifremer.fr/).

#### 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](https://www.galeio.fr)
- 📧 Email: [[email protected]](mailto:[email protected])
- 📍 Location: Paris, France
- 💼 LinkedIn: [Galeio](https://linkedin.com/company/galeio)

## 📚 Citations

If you use our models in your research, please cite:
```bibtex
@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}, 
}
```