Major-TOM/Core-S2L2A-249k-Clay-v1_5

Core-S2L2A-249k-Clay-v1.5

This dataset contains pre-computed image embeddings for the Core-S2L2A-249k satellite imagery subset, generated using the Clay v1.5 foundation model.

What is Clay?

Clay is an open-source foundation model for Earth observation. It uses a Masked Autoencoder (MAE) architecture with dynamic patch embeddings conditioned on sensor wavelengths. The model was trained on multi-spectral imagery from Sentinel-2, Landsat, NAIP, and other sensors.

• Model: Clay v1.5 (large)
• Architecture: ViT-Large MAE with dynamic embeddings
• Embedding dimension: 1024
• Input: 10 Sentinel-2 L2A bands (B02, B03, B04, B05, B06, B07, B08, B8A, B11, B12)
• Patch size: 8
• Input size: 384 × 384

How the embeddings were generated

1. Source imagery: MajorTOM Core-S2L2A-249k (~249k Sentinel-2 L2A chips, 384×384 pixels).
2. Preprocessing: Each chip was normalized using Clay's Sentinel-2 mean and std statistics.
3. Spatiotemporal inputs: The metadata-aware file conditions Clay on the acquisition time from product_datetime and the WGS84 centroid derived from each GeoTIFF's CRS and bounds.
4. Inference: The Clay encoder (without masking) was used to extract the CLS-token embedding from each chip.
5. Output: One 1024-dimensional embedding vector per chip, together with spatial and input-audit metadata.

Clay's metadata inputs use the official cyclical encodings:

• Time: [sin(2π week/52), cos(2π week/52), sin(2π hour/24), cos(2π hour/24)].
• Location: [sin(latitude), cos(latitude), sin(longitude), cos(longitude)], with angles in radians.

If a field cannot be recovered, only that four-value input is replaced with zeros. The source columns described below make this fallback auditable.

The generation script is generate_embeddings.py from the EarthEmbeddingExplorer repository.

Dataset structure

The dataset provides two GeoParquet variants:

Clay_crop_384x384.parquet
Clay_crop_384x384_with_space_time_input.parquet

File	Clay metadata input	Purpose
Clay_crop_384x384.parquet	Zero time and location vectors	Original visual-only index retained for reproducibility
Clay_crop_384x384_with_space_time_input.parquet	Real acquisition time and GeoTIFF footprint centroid	Recommended index for EarthEmbeddingExplorer

Metadata-aware file integrity

• Rows: 248,719
• Size: 1,083,918,437 bytes
• SHA-256: 658575606e327ef3e760b45682773aee4369a754d6e6abc8cd49f6e46ba3c92e
• Time source: parquet_product_datetime for all rows
• Location source: tiff_bounds for all rows

Columns

Column	Type	Description
unique_id	string	SHA-256 checksum of geometry + timestamp + product_id + embedding
embedding	float32[1024]	Clay v1.5 embedding vector
timestamp	string	Acquisition time (e.g., 20221115T161819)
product_id	string	Unique scene identifier
grid_cell	string	MajorTOM hierarchical grid code
grid_row_u	int16	Grid row index
grid_col_r	int16	Grid column index
geometry	geometry	WGS84 polygon of the chip footprint
centre_lat	float32	Center latitude
centre_lon	float32	Center longitude
utm_footprint	string	UTM footprint WKT
utm_crs	string	UTM CRS string
pixel_bbox	list	Pixel bounding box [x, y, x+w, y+h]
parquet_row	int64	Row index in the source imagery Parquet shard
parquet_url	string	URL to the source imagery Parquet shard
clay_time_input	float32[4]	Time vector passed to Clay
clay_latlon_input	float32[4]	Latitude/longitude vector passed to Clay
clay_time_input_source	string	tiff_tag, parquet_product_datetime, embedding_timestamp, or missing_zero_fallback
clay_latlon_input_source	string	tiff_bounds, embedding_center, or missing_zero_fallback

The four clay_* audit columns are present in the metadata-aware file only. Source precedence is TIFF time tag, source product_datetime, then embedding timestamp for time; TIFF bounds, then embedding center for location.

Usage

You can load the embeddings directly with pandas or geopandas:

import pandas as pd

df = pd.read_parquet("Clay_crop_384x384_with_space_time_input.parquet")
embeddings = df["embedding"].tolist()  # List of 1024-dim vectors

For cross-modal retrieval, pair this dataset with the EarthEmbeddingExplorer web application.

Acknowledgements

• Clay Foundation Model — Development Seed
• MajorTOM — European Space Agency

Citation

If you use this embedding dataset, please cite the EarthEmbeddingExplorer tutorial paper and the original Major-TOM paper:

@article{zheng2026earthembeddingexplorer,
  title={EarthEmbeddingExplorer: A Web Application for Cross-Modal Retrieval of Global Satellite Images},
  author={Zheng, Yijie and Wu, Weijie and Wu, Bingyue and Zhao, Long and Li, Guoqing and Czerkawski, Mikolaj and Klemmer, Konstantin},
  journal={arXiv preprint arXiv:2603.29441},
  year={2026},
  note={ICLR 2026 Workshop ML4RS Tutorial Track (oral)}
}

@inproceedings{francis2024majortom,
  title={Major TOM: Expandable Datasets for Earth Observation},
  author={Francis, Alistair and Czerkawski, Mikolaj},
  year={2024},
  booktitle={IGARSS 2024},
  eprint={2402.12095},
  archivePrefix={arXiv}
}

License

This dataset is released under the CC-BY-SA-4.0 license.