Major-TOM/Core-S2L2A-249k-SatCLIP

Core-S2RGB-249k-SatCLIP

Geospatial-vision embedding dataset computed from Core-S2L2A-249k using the SatCLIP model.

Overview

Property	Value
Source imagery	Core-S2L2A-249k (248,719 patches, 384 × 384 px)
Model	SatCLIP (ResNet-50 + location encoder)
Input bands	All 12 Sentinel-2 bands [B01, B02, B03, B04, B05, B06, B07, B08, B8A, B09, B11, B12]
Embedding dimension	512
Output format	GeoParquet
License	CC-BY-SA-4.0

Computation Pipeline

1. Pre-processing: Each 384 × 384 Sentinel-2 L2A patch is read from the source parquet files. All 12 spectral bands are stacked and normalised by dividing by 1e4 to convert digital numbers to reflectance.
2. Resize & Pad: The 12-band tensor is interpolated to the SatCLIP input size of 224 × 224 pixels. Because SatCLIP expects 13 input channels, a zero-filled B10 channel is padded at index 10.
3. Encoding: The 13-channel tensor is fed into the SatCLIP image encoder (a ResNet-50 trained with location-aware contrastive learning on satellite imagery) to extract a 512-dimensional image embedding.
4. Post-processing: No L2-normalisation is applied during dataset generation; normalisation is performed at retrieval time if required.
5. Geospatial metadata: The original UTM footprint is reprojected to EPSG:4326 (WGS-84) to obtain the geometry, centre_lat, and centre_lon fields. Additional metadata (product_id, grid_cell, timestamp, utm_crs, pixel_bbox) is preserved from the source dataset.

File Layout

.
├── SatCLIP_crop_384x384.parquet   # Main embedding GeoParquet (248,719 rows)
└── README.md

Schema

Column	Type	Description
unique_id	string	SHA-256 hash of geometry + timestamp + product_id + embedding
embedding	list	512-dim SatCLIP feature vector
timestamp	datetime	Acquisition timestamp
product_id	string	Original Sentinel-2 product identifier
grid_cell	string	Major-TOM grid cell identifier
grid_row_u	int16	Grid row index
grid_col_r	int16	Grid column index
geometry	geometry	WGS-84 polygon (footprint)
centre_lat	float32	Latitude of patch centre
centre_lon	float32	Longitude of patch centre
utm_footprint	string	Original UTM footprint as WKT
utm_crs	string	UTM CRS (e.g. EPSG:32633)
pixel_bbox	list	Pixel bounding box [x_min, y_min, x_max, y_max]
parquet_url	string	Source parquet file path in the image dataset
parquet_row	int64	Row index within the source parquet file

Usage

import pandas as pd

df = pd.read_parquet("SatCLIP_crop_384x384.parquet")
print(len(df), "embeddings")
print(df.iloc[0].embedding.shape)  # (512,)

Citation

If you use this embedding dataset, please cite the original Major-TOM paper and the SatCLIP 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}
}