Core-S2L2A-249k-OlmoEarth-Base

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

What is OlmoEarth?

OlmoEarth is an Earth-system foundation model trained on Sentinel-2 L2A and 6 derived geospatial maps (OpenStreetMap, WorldCover, SRTM DEM, etc.). It uses a flexible multi-modal vision transformer architecture and excels at capturing spectral and spatial patterns from 12-band multispectral imagery.

• Model: OlmoEarth-Base
• Architecture: Flexible multi-modal ViT
• Embedding dimension: 768
• Input: All 12 Sentinel-2 L2A bands [B01, B02, B03, B04, B05, B06, B07, B08, B8A, B09, B11, B12]
• Pooling: Mean pooling over unmasked spatial tokens
• 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. Pre-processing: Each 384 × 384 Sentinel-2 L2A patch is read from the source parquet files. All 12 spectral bands are extracted and reordered from the MajorTOM format [B01, B02, B03, B04, B05, B06, B07, B08, B8A, B09, B11, B12] to the OlmoEarth expected format [B02, B03, B04, B08, B05, B06, B07, B8A, B11, B12, B01, B09]. The bands are normalized using the OlmoEarth pre-training normalizer.
3. Encoding: The 12-band tensor is fed into the OlmoEarth-Base encoder to extract per-token features. Mean pooling is applied over all unmasked spatial tokens to produce a 768-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, parquet_url, parquet_row) is preserved from the source dataset.

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

Dataset structure

The dataset is distributed as a single GeoParquet file:

OlmoEarth_Base_crop_384x384.parquet

Columns

Column	Type	Description
unique_id	string	SHA-256 checksum of geometry + timestamp + product_id + embedding
embedding	list	768-dim OlmoEarth 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

You can load the embeddings directly with pandas or geopandas:

import pandas as pd

df = pd.read_parquet("OlmoEarth_Base_crop_384x384.parquet")
embeddings = df["embedding"].tolist()  # List of 768-dim vectors

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

Acknowledgements

• OlmoEarth — Allen Institute for AI
• MajorTOM — European Space Agency

License

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