Earth Index Embeddings V7

Chip-level vector embeddings of yearly, cloud-free Sentinel-2 mosaics, covering the globe between roughly 60.5°S and 70.3°N. They are produced by Earth Genome for Earth Index, a search engine for the Earth's surface: find places that look like a place you already know.

This is the V7 generation and supersedes earthindexembeddings. It holds two yearly periods, side by side:

Both periods use the same model and chip layout. The 2024 files were generated from V4 mosaics (a different compositing run from V7), so differences between the two periods can reflect imagery processing as well as real change on the ground. The imagery behind the 2024 embeddings is not part of the V7 imagery product.

This data is licensed under the Creative Commons Attribution 4.0 international license, which, in summary, only requires attribution.

Standards

The per-file metadata in this product follows the emerging geo-embeddings STAC embeddings specification (the STAC emb: extension, schema https://stac-extensions.github.io/embeddings/v0.0.1/schema.json). We validated it against the specification's JSON Schema:

• GeoParquet files embed a STAC item in their footer metadata under the emb key (next to the standard GeoParquet geo key), declaring the extension with emb:preprocessing, emb:runtime_parameters (bands, chip size, stride, intended uses), emb:quantization and emb:energy, and an embeddings asset. This item validates against the schema, so files stay self-describing when separated from the catalog.
• STAC items in ssl4eo_yearly_embeddings_v7 carry the same extension and fields and validate, with one caveat: the schema pins stac_version to 1.0.0, while our STAC API serves STAC 1.1.0.

The specification is a proposal and may change, so field names may be updated in later releases.

Resources

• Earth Index: search these embeddings in the browser
• Earth Genome
• STAC API: https://stac.earthgenome.org, collection ssl4eo_yearly_embeddings_v7. Search by place, MGRS tile or date, and follow each item's generated_from link to the imagery it came from.
• STAC Browser, a human-friendly interface to the same catalog
• Source imagery: earthindeximagery-v7

The href of each STAC asset points to Earth Genome's own S3 bucket (ei-ssl4eo-embeddings), not to Source Cooperative. The identical files are in this repository under the same name, so replace that prefix with https://data.source.coop/earthgenome/earthindex-embeddings-v7/ to read them from here.

Model

Embeddings come from the SSL4EO-S12 DINO ViT-S/16 model (self-supervised pretraining on Sentinel-2 L2A, from the SSL4EO-S12 project, loaded through TorchGeo). Each embedding is 384 dimensions, float32, not quantized.

This differs from the original earthindexembeddings product, which used the SoftCon model. Vectors from the two products are not comparable.

How the chips are made

• The mosaic is chipped with the MajorTom grid: cells of about 320 × 320 m, one embedding per cell.
• Each cell is read from the 10 m mosaic, upsampled bilinearly to 224 × 224 pixels, stacked into 13 bands (B01, B02, B03, B04, B05, B06, B07, B08, B8A, B09, B10, B11, B12), and divided by 10,000. Sentinel-2 L2A has no B10, so it is zero-filled.
• Cells only partly inside the tile, or touching no-data pixels, are skipped, and so are chips that are entirely empty.

Naming

Files are named by MGRS tile, then start and end date: 31TDF_2025-08-16_2026-08-16.parquet holds the embeddings for the Barcelona tile (31TDF) over the year ending 16 August 2026. The file name equals the STAC item id.

File format

Each file is GeoParquet 1.1 in WGS 84 (EPSG:4326), with one row group. Rows are not sorted by id. A full tile has up to roughly 235,000 rows (31TDF has 234,208); tiles that are partly ocean or no-data have fewer.

The Parquet key-value metadata carries a complete STAC item under the emb key, so every file is self-describing, following the embeddings STAC specification (see Standards). It includes preprocessing, bands, chip size, stride, quantization, and the energy used for inference.

Example

The same objects are readable over S3: s3://us-west-2.opendata.source.coop/earthgenome/earthindex-embeddings-v7/.

Contact

Using this? Want to work together? Ping us at info@earthgenome.org or info@earthindex.ai