falafel-hockey/sentinel2-lejepa-global-diverse-256

Sentinel-2 LeJEPA Preset-Biased (Small)

A small, preset-biased Sentinel-2 L2A chip dataset curated for self-supervised pretraining of a LeJEPA ResNet-18 encoder. Built as a reproducibility artifact for the Sentinel Change Explorer proof-of-concept foundation-model change-detection feature.

This is a proof of concept, not a general-purpose EO pretraining corpus. It is intentionally tiny (~thousands of chips) and biased toward the five demo AOIs the Sentinel Change Explorer app highlights. Use it to reproduce that specific PoC, not as a substitute for SSL4EO-S12, Clay, or Prithvi.

Dataset snapshot

Sampling methodology

Chips are drawn from two sources in roughly a 70/30 mix:

1. 1.Preset AOIs (~70%). For each of the 5 demo presets in the Sentinel Change Explorer app, the builder expands the tight demo bbox into a 10 km square centered on the preset's centroid, searches STAC (Element84 Earth Search v1) for Sentinel-2 L2A scenes in both before_range and after_range, loads the 5 reflectance bands + SCL via the same src.sentinel.load_bands the app uses, and tile-crops into non-overlapping 128x128 chips.
2. 2.Global diversity points (~30%). A hand-curated list of 30 globally diverse points (deserts, forests, croplands, urban cores, coasts, ice, wetlands) across every inhabited continent, each sampled at 2-3 dates spread across seasons. Same fetch-and-tile flow with a 5.12 km AOI.

Rejection filters

Every candidate chip is tested against two filters and dropped if it fails either:

• —Cloud/shadow fraction > 25%, computed from the Sentinel-2 Scene Classification Layer (SCL classes 3, 8, 9, 10).
• —Fill fraction > 10%, defined as pixels where all 5 reflectance bands equal zero (true no-data, not just a single dark band).

Preset AOIs

Global diversity points

• —sahara_algeria — (2.00, 25.00)
• —gobi_mongolia — (104.00, 43.50)
• —atacama_chile — (-69.30, -23.80)
• —namib_namibia — (15.00, -23.50)
• —simpson_australia — (137.50, -25.50)
• —amazon_brazil — (-60.00, -3.50)
• —congo_drc — (21.00, -1.00)
• —boreal_canada — (-95.00, 54.00)
• —siberia_taiga — (105.00, 62.00)
• —pnw_usa — (-123.50, 47.50)
• —iowa_corn_belt — (-93.50, 42.00)
• —pampas_argentina — (-62.00, -35.00)
• —po_valley_italy — (10.50, 45.00)
• —punjab_india — (75.50, 30.70)
• —tokyo_japan — (139.75, 35.70)
• —nyc_usa — (-73.95, 40.75)
• —lagos_nigeria — (3.40, 6.50)
• —sao_paulo_brazil — (-46.63, -23.55)
• —cairo_egypt — (31.25, 30.05)
• —shanghai_china — (121.47, 31.23)
• —chesapeake_bay — (-76.20, 38.50)
• —dutch_coast — (4.50, 52.50)
• —normandy_france — (-0.50, 49.30)
• —greenland_glacier — (-49.70, 69.20)
• —alps_switzerland — (8.00, 46.50)
• —andes_peru — (-72.00, -13.50)
• —himalaya_nepal — (86.50, 27.80)
• —everglades_usa — (-80.80, 25.80)
• —pantanal_brazil — (-56.00, -17.50)
• —okavango_botswana — (22.80, -19.30)

Schema

Each row is:

{
    "bands": Array3D(shape=(5, 128, 128), dtype=uint16),
    "bbox": Sequence(float32, length=4),        # (west, south, east, north) WGS84
    "acquisition_date": Value(string),          # ISO date of the source scene
    "scene_id": Value(string),                  # STAC item id
    "source": ClassLabel(names=["preset", "global"]),
    "preset_name": Value(string),               # "" for global chips
}

Normalization stats

Per-band mean and standard deviation computed over the training split (uint16 reflectance, before any scaling):

These are also shipped as norm_stats.json in the dataset bundle. The matching LeJEPA model repo embeds a copy so inference doesn't need to pull the dataset.

Usage

from datasets import load_dataset

ds = load_dataset("falafel-hockey/sentinel2-lejepa-global-diverse-256")
print(ds)
# DatasetDict with "train" and "validation" splits

sample = ds["train"][0]
print(sample["bands"].shape)   # (5, 128, 128)
print(sample["source"])        # 0 = preset, 1 = global

The companion pretrained LeJEPA ResNet-18 (5-band) is published separately and consumes these chips at native resolution without further resizing.

Limitations

• —Tiny scale. Thousands of chips, not millions. A real SSL corpus for remote sensing is 2-3 orders of magnitude larger. Expect the resulting features to overfit to the sampled AOIs and date windows.
• —Preset bias by design. 70% of chips come from 5 specific locations chosen because they are the demo AOIs in the companion app. This is intentional for the PoC but makes the features a poor fit for general-purpose EO tasks.
• —Single sensor, single level. Sentinel-2 L2A only. No Sentinel-1, no Landsat, no other modalities.
• —5 bands only. B02, B03, B04, B08, B11. The red-edge, cirrus, and SWIR22 bands are intentionally excluded to keep the model compact for M1 inference.
• —No deduplication across dates. Chips from the same AOI across different acquisition dates are both kept. This is a feature for temporal-invariance pretraining, but means chips are not i.i.d.

License and attribution

• —Chips are released under CC-BY-SA-4.0, matching Copernicus Sentinel data's terms for derived products.
• —Contains modified Copernicus Sentinel data [2023-2026], ESA. Source imagery: Sentinel-2 L2A via Element84 Earth Search v1.

Citation

@misc{sentinel2_lejepa_preset_biased_small,
  title  = {Sentinel-2 LeJEPA Preset-Biased (Small)},
  author = {Wheelis, Alex},
  year   = {2026},
  url    = {https://huggingface.co/datasets/falafel-hockey/sentinel2-lejepa-global-diverse-256}
}

@misc{balestriero2025lejepa,
  title  = {LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics},
  author = {Balestriero, Randall and LeCun, Yann},
  year   = {2025},
  eprint = {2511.08544},
  archivePrefix = {arXiv}
}