---
license: cc-by-4.0
task_categories:
- feature-extraction
language:
- en
tags:
- climate
- remote_sensing
- multi-modal
- NIR
- nDSM
- self-supervised
size_categories:
- 100B<n<1T
pretty_name: l
---

# M3DRS: Multi-Modal Multi-Resolution Remote Sensing Dataset


This repository hosts the M3DRS dataset, a comprehensive collection of 5-channel remote sensing images (RGB, NIR, nDSM) from Switzerland, France, and Italy. The dataset is unlabelled and specifically designed to support self-supervised learning tasks. It is part of our submission to the NeurIPS 2025 Datasets and Benchmarks Track. The dataset is organized into three folders, each containing ZIP archives of images grouped by location or in batches of 500. The dataset supports research in multi-modal learning, semantic segmentation, and geospatial analysis.

---

## 🧪 Benchmarking & Code

The M3DRS dataset is accompanied by a benchmark suite with ScaleMAE for pretraining and baseline models available at our GitHub repository:

🔗 [https://github.com/swiss-territorial-data-lab/proj-vit](https://github.com/swiss-territorial-data-lab/proj-vit)

This includes tools for data preprocessing, model training, and evaluation.

For dataset configuration or details, please use a [discussion](https://huggingface.co/datasets/heig-vd-geo/M3DRS/discussions) in this repository.

For benchmark questions or feedback, please open an issue in the [GitHub repository](https://github.com/swiss-territorial-data-lab/proj-vit).

---

## 📁 Dataset Structure

```
├── swiss/
│   ├── scratch_{aquization_date}_{tile_idx}.zip
│   └── ...
├── flair/
│   ├── batch_1.zip
│   ├── batch_2.zip
│   └── ...
├── italy/
│   ├── batch_1.zip
│   ├── batch_2.zip
│   └── ...
├── NOTICE
└── README.md
```

* **`swiss/`**: Contains images from Switzerland, grouped by specific locations.
* **`flair/`**: Contains images from France, grouped in batches of 500 images per ZIP file.
* **`italy/`**: Contains images from Italy, grouped in batches of 500 images per ZIP file.
* **`NOTICE`**: Details the licenses and attributions for the source data.

---

## 📊 Dataset Composition

| Source       | Country     | Area (km²) | Resolution | Images  | Size (GB) |
| ------------ | ----------- | ---------- | ---------- | ------- | --------- |
| Swisstopo    | Switzerland | 2,172      | 10/25 cm   | 282,243 | 346       |
| Ferrara City | Italy       | 95         | 10 cm      | 39,907  | 49        |
| FLAIR #1     | France      | 810        | 20 cm      | 77,762  | 96        |
| **Total**    |             | 3,077      |            | 399,912 | 491       |

---

## 📥 Download Instructions

You can download the dataset directly from this Hugging Face repository. Each folder (`swiss`, `flair`, `italy`) contains ZIP files as described above.

⚠️ Warning: This dataset is large, the full download size is approximately 380 GB. Unzipped files are approximately 500 GB Make sure you have sufficient disk space and a stable internet connection before downloading.

```
from huggingface_hub import snapshot_download
local_dir = snapshot_download(
    repo_id="heig-vd-geo/M3DRS",
    repo_type="dataset",
    local_dir="" # where to replicate the file tree
)
```

---

## ⚖️ License

The dataset is composed of public open data sources. Please refer to the `NOTICE` file in the root of this repository for detailed licensing information and attributions for each data source.

---

## 📚 Citation

If you use the M3DRS dataset in your research, please cite our NeurIPS 2025 Datasets and Benchmarks Track submission:

```bibtex
@inproceedings{m3drs2025,
  title={M3DRS: Multi-Modal Multi-Resolution Remote Sensing Dataset},
  author={Shanci Li, Antoine Carreaud, Adrien Gressin},
  booktitle={NeurIPS Datasets and Benchmarks Track},
  year={2025}
}
```

---