Core-S2L2A-249k-Clay-v1_5

# Core-S2L2A-249k-Clay-v1.5 This dataset contains **pre-computed image embeddings** for the [Core-S2L2A-249k](https://modelscope.cn/datasets/Major-TOM/Core-S2L2A-249k) satellite imagery subset, generated using the **[Clay v1.5](https://github.com/Clay-foundation/model)** 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. **Inference**: The Clay encoder (without masking) was used to extract the CLS-token embedding from each chip. 4. **Output**: One 1024-dimensional embedding vector per chip, together with spatial metadata. The generation script is `generate_embeddings.py` from the [EarthEmbeddingExplorer](https://github.com/OpenGeoScope/EarthEmbeddingExplorer) repository. ## Dataset structure The dataset is distributed as a single GeoParquet file: ``` Clay_crop_384x384.parquet ``` ### 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 | ## Usage You can load the embeddings directly with pandas or geopandas: ```python import pandas as pd df = pd.read_parquet("Clay_crop_384x384.parquet") embeddings = df["embedding"].tolist() # List of 1024-dim vectors ``` For cross-modal retrieval, pair this dataset with the [EarthEmbeddingExplorer](https://github.com/OpenGeoScope/EarthEmbeddingExplorer) web application. ## Acknowledgements - [Clay Foundation Model](https://github.com/Clay-foundation/model) — Development Seed - [MajorTOM](https://github.com/ESA-PhiLab/MajorTOM) — European Space Agency ## Citation If you use this embedding dataset, please cite the EarthEmbeddingExplorer tutorial paper and the original Major-TOM paper: ```bibtex @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)} } ``` ```bibtex @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](https://creativecommons.org/licenses/by-sa/4.0/) license.