phisat2-s2-triplets / README.md
clementgilli's picture
Update README.md
f0fa7ef verified
|
Raw History Blame Contribute Delete
5.92 kB
metadata
license: apache-2.0
pretty_name: Φsat-2 / Sentinel-2 coregistered triplets
task_categories:
  - image-segmentation
  - image-to-image
tags:
  - remote-sensing
  - earth-observation
  - satellite-imagery
  - domain-adaptation
  - sentinel-2
  - phisat-2
size_categories:
  - 10K<n<100K

Φsat-2 / Sentinel-2 coregistered triplets

Coregistered pairs of real Φsat-2 acquisitions and temporally matched Sentinel-2 L1C images, with ESA WorldCover reference labels and per-patch metadata. Each real Φsat-2 patch also comes with a simulated Φsat-2 counterpart, which makes the dataset suited to cross-sensor domain adaptation from Sentinel-2 to Φsat-2 and to comparing simulated and real acquisitions of the same scenes.

Contents

  • 119 Φsat-2 tiles (October 2024 – May 2025) over 80 Sentinel-2 MGRS tiles; median time gap to the matched Sentinel-2 image 3 days (max 21).
  • 19,145 coregistered 256 × 256 patches. SuperPoint + LightGlue on the Φsat-2 panchromatic band vs. Sentinel-2 B04, per-patch refinement; patches with registration RMSE > 0.85 px rejected (median 0.51 px).
  • Per patch: one GeoTIFF per modality, 256 × 256 pixels on the Φsat-2 grid (EPSG:4326, ≈ 4.75 m pixels).
    Folder Content Bands (in file order) Type
    tiles/phisat2 Real Φsat-2, L1C B00 PAN (625 nm), B01 (490 nm), B02 (560 nm), B03 (665 nm), B04 (705 nm), B05 (740 nm), B06 (783 nm), B07 (842 nm) integer digital numbers
    tiles/sentinel2 Matched Sentinel-2 L1C, resampled to the Φsat-2 grid B01, B02, B03, B04, B05, B06, B07, B08, B8A, B09, B10, B11, B12 uint16, L1C digital numbers with the +1000 radiometric offset (reflectance = (DN − 1000) / 10000)
    tiles/simulated_phisat2 Simulated Φsat-2 (see below) B02, B03, B04, PAN, B08, B05, B06, B07 (Sentinel-2-equivalent band names) float32
    masks/worldcover ESA WorldCover 2021 v200, reprojected onto the patch 1 band, WorldCover class codes (10–100) uint8
    masks/phisat2_cloud Cloud mask of the Φsat-2 image (OmniCloudMask) 1 band: 0 clear, 1 thick cloud, 2 thin cloud, 3 cloud shadow; 255 no data uint8

The simulated images are the aligned Sentinel-2 images processed by the Φsat-2 end-to-end simulator of Longépé et al. (2024), which models the target sensor's spectral bands, spatial response (MTF) and noise (SNR). Every real Φsat-2 patch therefore has a simulated counterpart of the same scene.

The cloud masks are generated automatically and are of limited quality. WorldCover is missing for 27 tiles (see metadata/tiles.csv, column wc_exists).

Layout

data/<pair_id>.tar            one archive per tile
splits/train.csv, val.csv, test.csv, train_100.csv, train_1000.csv, train_10000.csv
metadata/tiles.csv            per-patch audit: location, WorldCover validity, cloud fraction
metadata/manifest.csv         files, patches, size and sha256 of each archive

Extracting the archives restores pairs/<pair_id>/lightglue_coregistration/{tiles,masks,audit}/.... Split files list tile_id = <pair_id>_<rRRRRRR_cCCCCCC>.tif.

Splits

Location-disjoint (Sentinel-2 MGRS tile): no location or tile appears in more than one split. Validation and test locations were drawn at random among those with WorldCover labels (most representative of 500 draws in WorldCover class distribution); they keep only patches with ≤ 5% pixels without a valid WorldCover class and ≤ 10% pixels flagged by the cloud mask. Train: 14,198 patches (84 tiles), val: 2,483 (16), test: 2,464 (19). The budget subsets are nested and drawn proportionally to each tile's size.

Usage

from huggingface_hub import snapshot_download
import tarfile, pathlib
 
root = pathlib.Path(snapshot_download("ESA-PhiLab-Edge/phisat2-s2-triplets", repo_type="dataset", local_dir="triplets"))
for t in sorted((root / "data").glob("*.tar")):
    with tarfile.open(t) as tf:
        tf.extractall(root)          # -> triplets/pairs/<pair_id>/...

License and attribution

  • Sentinel-2: contains modified Copernicus Sentinel data [2024–2025].
  • ESA WorldCover 10 m 2021 v200: © ESA WorldCover project 2021 / Contains modified Copernicus Sentinel data (2021) processed by ESA WorldCover consortium; licensed under CC BY 4.0 (citation below).

Citation

@misc{triplet_dataset,
  author       = {Del Prete, Roberto and Gilli, Cl{\'e}ment},
  title        = {{$\Phi$sat-2}/{Sentinel-2} Coregistered Triplets},
  howpublished = {Hugging Face dataset},
  year         = {2026},
  doi          = {10.57967/hf/11622}
}

The simulated Φsat-2 images were generated with:

@article{longepe2024simulation,
  author  = {Long{\'e}p{\'e}, Nicolas and Petrelli, Isabella and Kadunc, Nika Oman and Peressutti, Devis and Del Prete, Roberto and Casaburi, Mauro and Babkina, Irina and Vercruyssen, Nathan and Luis, Elisa Callejo and Elorza, {\'A}lvaro Mor{\'o}n and Marchese, Valentina and Kidron, Agne Paskeviciute and Melega, Nicola},
  title   = {Simulation of Multispectral and Hyperspectral {EO} Products for Onboard Machine Learning Application},
  journal = {IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing},
  year    = {2024},
  volume  = {17},
  pages   = {17651--17665},
  doi     = {10.1109/JSTARS.2024.3434437}
}

The WorldCover labels come from:

@misc{zanaga_2022_7254221,
  author    = {Zanaga, Daniele and Van De Kerchove, Ruben and Daems, Dirk and De Keersmaecker, Wanda and Brockmann, Carsten and Kirches, Grit and Wevers, Jan and Cartus, Oliver and Santoro, Maurizio and Fritz, Steffen and Lesiv, Myroslava and Herold, Martin and Tsendbazar, Nandin-Erdene and Xu, Panpan and Ramoino, Fabrizio and Arino, Olivier},
  title     = {{ESA WorldCover 10 m 2021 v200}},
  year      = {2022},
  publisher = {Zenodo},
  version   = {v200},
  doi       = {10.5281/zenodo.7254221}
}