README.md

---
pipeline_tag: image-feature-extraction
license: mit
datasets:
- IGNF/FLAIR-HUB
---

# RAMEN: Resolution-Adjustable Multimodal Encoder for Earth Observation

[Paper](https://huggingface.co/papers/2512.05025) | [Code](https://github.com/nicolashoudre/RAMEN)

RAMEN is a resolution-adjustable multimodal encoder that learns a shared visual representation across Earth Observation (EO) data in a fully sensor-agnostic manner. It treats modality and spatial/temporal resolutions as key input features, enabling coherent analysis across modalities. Its main methodological contribution is to define spatial resolution as a controllable output parameter, giving users direct control over the desired level of detail at inference and allowing explicit trade-offs between spatial precision and computational cost.

<p align="center">
  <img src="https://github.com/nicolashoudre/RAMEN/raw/main/.figures/Intro_RAMEN.png" alt="RAMEN workflow" width="400"/>
</p>

## Key features

- 🛰️ **Sensor-agnostic foundation model**: RAMEN supports any kind of multispectral, SAR or elevation maps modalities. Just specify input shape, channels and original spatial resolution (GSD) !
- 🔧 **Adjustable feature map resolution**: Customize the resolution of feature maps to suit specific downstream tasks and computational constraints.
- 🌍 **Multimodal data fusion**: Effectively combine data from multiple modalities into a unified representation.

## PANGAEA Bench evaluation

All downstream tasks results presented in RAMEN were conducted using the [PANGAEA](https://github.com/VMarsocci/pangaea-bench) Benchmark. We report here the main results obtained on eight tasks.

| Model | BurnSr | MADOS | PASTIS | Sen1Fl11 | DEN | CTM-SS | SN7 | AI4Farms | Avg. mIoU | Avg. Rank |
|-------|---------|--------|--------|----------|------|--------|------|-----------|-----------|-----------|
| CROMA | 82.42 | 67.55 | 32.32 | 90.89 | 38.29 | 49.38 | 59.28 | 25.65 | 55.72 | 6.50 |
| DOFA | 80.63 | 59.58 | 30.02 | 89.37 | 39.29 | 51.33 | **61.84** | 27.07 | 54.89 | 7.50 |
| TerraMind-B | 82.42 | 69.52 | 40.51 | 90.62 | 37.87 | **55.80** | 60.61 | 28.12 | 58.18 | 4.25 |
| TerraMind-L | 82.93 | **75.57** | **43.13** | 90.78 | 37.89 | 55.04 | 59.98 | 27.47 | 59.10 | 3.75 |
| **RAMEN (ours)** | **85.02** | 69.72 | 42.29 | **91.03** | **39.85** | 53.27 | 60.31 | **38.78** | **60.03** | **2.63** |

More informations on how to reproduce results and implement RAMEN in PANGAEA can be found in the [`pangaea-bench`](https://github.com/nicolashoudre/RAMEN/tree/main/pangaea-bench) folder.

## Citation

If you use RAMEN, please cite our paper:

```bibtex
@article{RAMEN,
  title={{RAMEN}: Resolution-Adjustable Multimodal Encoder for Earth Observation},
  author={Nicolas Houdré and Diego Marcos and Hugo Riffaud de Turckheim and Dino Ienco and Laurent Wendling and Camille Kurtz and Sylvain Lobry},
  journal={arXiv preprint arXiv:2512.05025},
  year={2025}
}
```