Core-S2L2A-UniverSat

 *Global coverage of the embeddings, coloured by the top-3 principal components of the 768-d UniverSat vectors (mapped to RGB). Distinct colours mark distinct embedding neighbourhoods — deserts (yellow), vegetation (green), ice & boreal regions (cyan).* # Core-S2L2A-UniverSat 🛰️ | Dataset | Modality | Number of Embeddings | Sensing Type | Embedding Dim | Source Dataset | Source Model | Size | |:--------:|:--------------:|:-------------------:|:------------:|:--------------:|:--------------:|:--------------:|:--------------:| | Core-S2L2A-UniverSat | Sentinel-2 Level 2A | 2,245,884 | Multispectral (L2A surface reflectance) | 768 | [Core-S2L2A](https://huggingface.co/datasets/Major-TOM/Core-S2L2A) | [UniverSat](https://huggingface.co/g-astruc/UniverSat) | 6.4 GB | This dataset provides a dense, global set of **whole-image embeddings** for the [**Major TOM Core-S2L2A**](https://huggingface.co/datasets/Major-TOM/Core-S2L2A) collection. Each Sentinel-2 Level 2A fragment is encoded into a single **768-dimensional** vector with [**UniverSat**](https://huggingface.co/g-astruc/UniverSat), a resolution- and modality-agnostic transformer backbone for Earth Observation. It is an embedding expansion of Major TOM in the spirit of the [Major TOM Global Embeddings](https://huggingface.co/collections/Major-TOM/major-tom-global-embeddings-6748110149103fbfdad4b210) project. ## Content | Field | Type | Description | |:-----------:|:------:|-------------------------------------------------------------------| | embedding | array | Raw UniverSat whole-image embedding (768 × float32) | | grid_cell | string | Major TOM grid cell (e.g. `922D_249L`) | | timestamp | string | Sensing timestamp of the source product (e.g. `20230119T161811`) | | product_id | string | ID of the original Sentinel-2 product | Rows are stored across 20 `data-*.parquet` shards (~112k rows each), keyed to the source fragment by `(grid_cell, product_id, timestamp)`, so embeddings can be joined back to imagery, geometry, and other Major TOM expansions of [Core-S2L2A](https://huggingface.co/datasets/Major-TOM/Core-S2L2A). ## Input Data * Sentinel-2 (Level 2A) surface-reflectance fragments from [**Major TOM Core-S2L2A**](https://huggingface.co/datasets/Major-TOM/Core-S2L2A), normalised per band with the standard Major TOM `NORM_s2l2a` statistics. * Each fragment is encoded at its **native 10 m grid** (image size **1068 × 1068** px ≈ a 10.7 km tile), so one embedding summarises the whole fragment — no tiling, no overlap. * Input patch size: **120 m** (≈ an 89 × 89 patch grid); the encoder pools this into a single global descriptor per fragment. The embeddings build on the [**Major TOM**](https://huggingface.co/Major-TOM) Core-S2L2A dataset. ## Citation If you use this dataset, please cite UniverSat: ```bibtex @article{perron2026universat, title = {UniverSat: Resolution- and Modality-Agnostic Transformers for Earth Observation}, author = {Perron, Yohann and Astruc, Guillaume and Gonthier, Nicolas and Mallet, Clement and Landrieu, Loic}, journal = {arXiv preprint arXiv:2606.23503}, year = {2026} } ```