README.md

---
license: cc-by-4.0
---

# SSL4EO-EU-Forest
<table>
  <tr>
    <td align="center">
      <img src="https://cdn-uploads.huggingface.co/production/uploads/66f8758c91a4135d96b80176/Qi5QpnIMUI1AdX5hSv9zN.png" alt="Sentinel-2 L2A winter scene" width="200"/>
<br/>
      <sub>Sentinel-2 winter scene</sub>
    </td>
    <td align="center">
      <img src="https://cdn-uploads.huggingface.co/production/uploads/66f8758c91a4135d96b80176/4Nrz90xyXWoeRJC-K4eSm.png" alt="Sentinel-2 L2A winter scene" width="200"/>
<br/>
      <sub>tree (green) segmentation</sub>
    </td>
    <td align="center">
      <img src="https://cdn-uploads.huggingface.co/production/uploads/66f8758c91a4135d96b80176/N5tYT2rSHK_k5y8U7z3EH.png" alt="tree (green) segmentation" width="200"/><br/>
      <sub>Sentinel-2 summer scene</sub>
    </td>
    <td align="center">
      <img src="https://cdn-uploads.huggingface.co/production/uploads/66f8758c91a4135d96b80176/sEsiAbhag9XPqdE7c_7Jq.png" alt="tree (green) segmentation" width="200"/><br/>
      <sub>tree (green) segmentation</sub>
    </td>
  </tr>
</table>

## Introduction

The dataset is an add-on to the globally sampled SSL4EO-S12 pre-training dataset (https://doi.org/10.1109/MGRS.2023.3281651) where `v1.1` is available at https://huggingface.co/datasets/embed2scale/SSL4EO-S12-v1.1 with downstream tasks in https://huggingface.co/datasets/embed2scale/SSL4EO-S12-downstream .

## Data Characteristics

*SSL4EO-EU-Forest* assembles an additional downstream task (85 GB) for binary semantic segmentation of forest land cover as follows:
- `images/` sub-directory: It curates 16K spatial patches (264x264 pixels) of Sentinel-2 L2A imagery (`UInt16` GeoTIFF, 12 bands upsampled to 10m pixel resolution) over European forests. Each geolocation contains four Sentinel-2 timestamps from all seasons (winter, spring, summer, fall) of year 2018. The cloud cover in each scene is below 10%.
- `masks/` sub-directory: Binary forest masks (`1`: tree, `0`: background, `Byte` GeoTIFF) have been derived from the 2018 HRL dataset (https://doi.org/10.2909/486f77da-d605-423e-93a9-680760ab6791) available through the Copernicus Land Monitoring Service, cf. https://land.copernicus.eu/en/products/high-resolution-layer-forests-and-tree-cover/dominant-leaf-type-2018-raster-10-m-europe-yearly .
- Data of a geolocation are placed in numbered sub-directories (e.g. `0017618`) of `images/` and `masks/`, respectively. Each of those for the `images/` directory contains four directories indicating the four seasons as per convention by Sentinel-2. For example, `20180520T094031_20180520T094033_T35VMH` denotes the sensing start time as approx. May 20, 2018 at 9.41am UTC (spring scene) for geolocation defined by tile ID `T35VMH` (which the patch was cropped from). Additional details in https://sentiwiki.copernicus.eu/web/s2-products .

## Spatial Coverage

<img src="https://cdn-uploads.huggingface.co/production/uploads/66f8758c91a4135d96b80176/7dcsK6k5MchJcmikkI-dC.png" alt="Spatial Distribution of SSL4EO-EU-Forest" width="700"/>

## Funding

This work was carried under the *EvoLand* project, cf. https://www.evo-land.eu . This project has received funding from the European Union's Horizon Europe research and innovation programme under grant agreement No. `101082130`.