---
license: cc-by-4.0
task_categories:
- image-segmentation
language:
- en
tags:
- sentinel-2
- remote-sensing
- earth-observation
- zarr
- geospatial
pretty_name: Sentinel-2 Land Cover Dataset (Zarr Format)
---
# Sentinel-2 Land Cover Dataset (Zarr Format)

<!-- Provide a quick summary of the dataset. -->
## Dataset Summary
This dataset contains land-cover labels derived from the [AI4LCC benchmark dataset](https://doi.theia.data-terra.org/ai4lcc/?lang=en) and corresponding Sentinel-2 satellite imagery.
The original labels and Sentinel-2 were converted into Zarr format to enable efficient storage and scalable access for large geospatial datasets. Sentinel-2 image tiles corresponding to the label locations were added to create a paired dataset suitable for machine learning tasks such as land-cover classification and semantic segmentation.
The dataset is intended for research and development in remote sensing, Earth observation, and machine learning.

<!--It has been generated using [this raw template](https://github.com/huggingface/huggingface_hub/blob/main/src/hug<>gingface_hub/templates/datasetcard_template.md?plain=1).-->


## Dataset Details 

### Dataset Source
The label data originates from the AI4LCC benchmark dataset: [https://doi.theia.data-terra.org/ai4lcc/?lang=en](https://doi.theia.data-terra.org/ai4lcc/?lang=en)
The Sentinel-2 imagery was retrieved from Planetary Computer for the same spatial locations and time periods (along 2020) corresponding to the labels.


### Supported Tasks
The dataset can be used for:
- Land-cover classification
- Semantic segmentation
- Remote sensing representation learning
- Geospatial deep learning research

### Dataset Structure 
The dataset is stored in **Zarr format**, where each `.zarr` folder contains a single tile with all Sentinel-2 spectral bands, the corresponding label, and spatial coordinates. This format allows efficient storage and scalable access for large geospatial datasets.<br>
Current available granules: 
- 30TXN, 30TXP, 30TXQ, 30TXR, 30TXS,
- 30TXT, 30TYN, 30TYP, 30TYQ, 30TYR,
- 30TYT, 31TCJ, 31TCK, 31TCL, 31TCM,
- 31TCN, 31TDK, 31TDL, 31TDM, 31TFN,
- 31UEP, 31UEQ
<!--
Example repository structure:
dataset/ <br>
 ├sentinel2_{granule}_reduced2021_withlabelclass1.zarr/ <br>
 │ └── B01 <br>
 │   └── B02 <br>
 │   └── B03 <br>
 │   └── B04 <br>
 │   └── B05 <br> 
│   └── B06 <br>
│   └── B07 <br>
│   └── B08 <br>
│   └── B8A <br>
│   └── B09 <br>
│   └── B11 <br>
│   └── B12 <br>
│   └── label <br>
│   └── spatial_ref <br>
│   └── x <br>
│   └── y <br>
├── data1.zarr/ <br>
-->

### Dataset Creation
- Retrieving Sentinel-2 imagery (2020) within ground-truth bounding boxes using the Planetary Computer API (sentinel-2-l2a) and odc-stac 
- Generating annual composites using median reduction with Xarray 
- Simplifying the original 14 land-cover classes into 5 classes as in Table 1 ref [Wenger et al., 2022](https://www.germain-forestier.info/publis/isprs2022.pdf)
- Combining imagery and labels
<p align="center">
<img src="images/typology_AI4LCC.png" alt="typology" width="400">
</p>

### Intended Use
This dataset is intended for:
- machine learning research
- benchmarking remote sensing models
- land-cover classification studies
- geospatial AI experimentation
It may also serve as a benchmark dataset for scalable data pipelines using Zarr.

### Limitations
Potential limitations include:
- Temporal mismatch between label creation and satellite acquisition
- Spatial resolution differences between label sources and imagery
- Class imbalance in land-cover categories
Users should validate dataset suitability for their specific application.

### Licensing and Attribution
The label data originates from the AI4LCC benchmark dataset.
Users must comply with the licensing terms of the original dataset.
If using this dataset, please also cite the original AI4LCC dataset: [https://doi.theia.data-terra.org/ai4lcc/?lang=en](https://doi.theia.data-terra.org/ai4lcc/?lang=en)

## Acknowledgements
This dataset builds upon the AI4LCC benchmark dataset and publicly available Sentinel-2 satellite imagery.