---
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 for Clay (Precomputed)
size_categories:
- 10K<n<100K
---

# Sentinel-2 Land Cover Dataset for Clay (Precomputed)

## Dataset Summary
This dataset is a derived version of the [Sentinel-2 Land Cover Dataset (Zarr Format)](https://huggingface.co/datasets/wtr001/S2_AI4LCC) prepared specifically for use with the [Clay foundation model](https://github.com/Clay-foundation/model/tree/main) for Earth observation. The dataset was precomputed and reformatted so that it can be used directly in Clay training or inference pipelines without additional preprocessing. Operations that normally occur during data loading—such as tiling, data and label filtration, normalization—have already been applied and stored in the dataset. The goal of this dataset is to provide an efficient, ready-to-use representation of Sentinel-2 land-cover data optimized for large-scale machine learning workflows.

## Source Dataset
This dataset is derived from the [Sentinel-2 Land Cover Dataset (Zarr Format)](https://huggingface.co/datasets/wtr001/S2_AI4LCC), which itself combines:
- land-cover labels originating from the [AI4LCC benchmark dataset](https://doi.theia.data-terra.org/ai4lcc/?lang=en)
- corresponding Sentinel-2 satellite imagery
Only preprocessing and structural changes were applied; the semantic content of the labels remains unchanged.

## Preprocessing
To make the dataset compatible with Clay pipelines, several preprocessing steps were applied:
- **Tiling**: Large Sentinel-2 scenes were divided into fixed-size tiles to enable efficient loading and batch processing during model training.
- **Data filtering**: Satellite observations affected by quality issues (e.g., corrupted or unreliable data) were removed to improve dataset quality.
- **Label filtering**: Samples dominated by the background or majority class (class 0) were filtered out to reduce class imbalance and improve the usefulness of training samples.
- **Clay-compatible formatting**: The dataset was reorganized into a structure compatible with Clay data pipelines and model input requirements.
- **Zarr storage**: The processed data were stored in Zarr format to enable scalable, chunked access and efficient loading for large-scale machine learning workflows.

These preprocessing steps remove the need to perform expensive transformations during training. More about the preprocessing, please visit my github repo particularly the [ClayIterableDataset](https://github.com/witra/ModularGeoFM/blob/main/src/modulargeofm/datamodules/clay_datamodule.py)

## Data Fields
Typical fields within each Zarr tile include: <br>
- **pixels**\
  Sentinel-2 multi-spectral imagery with dimensions [samples, bands, height, width] = [num_chips, 12, 224, 224]. The bands dimension corresponds to the 12 Sentinel-2 spectral bands used in the dataset.
- **label**\
  Land-cover class labels aligned with the imagery, with dimensions [samples, height, width] = [num_chips, 224, 224]. Each pixel contains the land-cover class associated with the corresponding Sentinel-2 observation.
- **latlon**\
  Geographic bounding box coordinates for each sample, with dimensions [samples, 4] = [num_chips, 4]. The four values represent the spatial extent of the tile in the following order: [ymin, xmin, ymax, xmax].
- **time**\
  Temporal metadata describing the acquisition time of each sample, with dimensions [samples, 4] = [num_chips, 4].
  The time representation consists of: [year, month, week, hour].
- **attributes**\
  Additional dataset-level metadata describing sensor characteristics, including:\
  gsd – ground sampling distance (spatial resolution)\
  waves – central wavelengths corresponding to the Sentinel-2 spectral bands.

Currently there are 19367 and 2304 chips for training and validation dataset respectively

## Example Samples

Below is an example of a data patch and its corresponding land-cover label.

Each sample consists of a **224 × 224 Sentinel-2 multi-spectral tile** and the aligned **land-cover label mask**.

**Example data patch**

* Input tensor shape: `[12, 224, 224]` represented as R(B04), G(B03), B(B02)
* Bands correspond to Sentinel-2 spectral channels.

**Example label patch**

* Label tensor shape: `[224, 224]`
* Each pixel represents the land-cover class in level 1 topology as in [Wenger et al., 2022](https://www.germain-forestier.info/publis/isprs2022.pdf) associated with the corresponding satellite observation.

Example visualization:
<p align="center">
<img src="images/patches.png" alt="typology" width="600">

<img src="images/typology_AI4LCC.png" alt="typology" width="600">
</p>

## Intended Use
This dataset is designed for:
- training Clay-based remote sensing models
- benchmarking foundation models for Earth observation
- land-cover classification and segmentation tasks
- geospatial machine learning research
  
It is not intended to serve as a general-purpose Sentinel-2 dataset but rather as a model-ready representation optimized for Clay workflows.

## Limitations
Users should consider the following limitations:
- The dataset inherits biases and limitations from the original label source.
- Temporal differences between labels and Sentinel-2 imagery may exist.
- The preprocessing choices were optimized for Clay and may not be optimal for other models.

## Licensing and Attribution
This dataset derives from the Sentinel-2 Land Cover Dataset and ultimately from the AI4LCC benchmark dataset. Users must respect the licensing terms of the original data sources. Please cite the original dataset when using this derived version.