Jaervis/cropmark-disambiguator-s2

Cropmark Disambiguator — Sentinel-2 Half

A 922-tile classification dataset (767 train + 155 eval) of Sentinel-2 spectral imagery over agricultural land in the UK (Historic England), Italy (Tavoliere), and Norway (Askeladden), labeled Yes/No for the presence of buried-archaeological-feature cropmark stress. Used to fine-tune Liquid AI's LFM2.5-VL-450M (LoRA on language model + multi-modal projector layers) for the Liquid AI × DPhi Space "AI in Space" hackathon (April–May 2026).

The original SFT corpus had 1,138 train + 157 eval rows, but ~360 rows referenced the same composite as a sibling row (curriculum repetition with consistent labels) and 11 rows referenced composites that were later moved to a rejected_composites/ quarantine after failing post-hoc quality gates; both classes are excluded from publication. The published 922 tiles are the unique, on-disk, label-consistent subset.

Why Sentinel-2 only?

During training, each tile here was paired side-by-side with a high-resolution Mapbox satellite image (left half = S2 spectral, right half = Mapbox RGB) so the model could combine spectral-stress detection with geometric context. We do not redistribute the Mapbox half. Mapbox's Terms of Service restrict the redistribution of "Map Assets" delivered by their Mapping APIs — including imagery returned by the Static Images and Static Tiles APIs — beyond the licensee's own application, and the Mapbox Maps APIs Caching documentation further constrains caching to end-user devices. We read these terms as incompatible with publishing Mapbox-derived imagery to a third-party dataset repository, and we err on the side of not redistributing.

What this means in practice:

• What you can reuse here: the Sentinel-2 spectral half (256×512 px crops, RGB-encoded NIR/Red/RedEdge1 composite) plus the Yes/No label per tile.
• What you cannot get from this dataset: the high-res RGB context, exact source coordinates, or the original full 512×512 composites used in training.
• If you want to reproduce the full multi-modal stack: the original collection pipeline used a SimSat Docker wrapper around Sentinel-2 STAC + Mapbox Static Images API to fetch tile pairs by (lon, lat, timestamp) from the source archaeological databases listed in "Source attribution" below. We do not publish a precomputed coordinate manifest with this dataset; reproducing the full multi-modal stack would require re-querying those source databases under their own terms and pairing each site with a fresh Mapbox tile fetched with your own Mapbox token.

The published Sentinel-2 half retains the spectral-stress signal — NDVI, NDRE, PSRI, NMDI, NBR, and red/green ratio responses — that drives our physics-grounded discrimination logic. This is the same signal validated by Gravanis & Agapiou 2026 ("Physically-based modelling for retrospective detection of archaeological proxies (cropmarks)", Scientific Reports).

Dataset structure

data/
  train/
    metadata.csv     # columns: file_name, label
    *.jpg            # 767 S2-only crops, 256×512 px, RGB JPEG q95
  eval/
    metadata.csv
    *.jpg            # 155 S2-only crops

Each row in metadata.csv:

column	type	values
file_name	string	basename of the JPG, e.g. d_tile_0167.jpg
label	string	"Yes" (cropmark stress likely) or "No" (no stress)

No coordinates, no source IDs, no temporal metadata are included. The original collection pipeline tracked these privately (see "Source attribution" below for how to recover them); they are not redistributed here.

Loading example:

from datasets import load_dataset
ds = load_dataset("Jaervis/cropmark-disambiguator-s2")
print(ds)
print(ds["train"][0])

Source attribution

Site coordinates were collected from three public archaeological databases:

• Historic England Aerial Investigation and Mapping (AIM) — open under the Open Government Licence v3.0. Used for cereal-land cropmark records and false-positive categories (drains, geological lineaments, cultivation marks, natural features).
• Riksantikvaren / Askeladden OGC API (Norwegian Directorate for Cultural Heritage) — public OGC service. Used for protected and dismissed (FJE/IKKEV) site records.
• Tavoliere prehistoric sites (Apulia, Italy) — coordinates compiled from published archaeological literature.

Sentinel-2 imagery was retrieved from a SimSat Docker wrapper around Sentinel-2 STAC services. Sentinel-2 data is © European Union, contains modified Copernicus Sentinel data 2025–2026, redistributed under the Copernicus Sentinel Data Terms and Conditions.

Quality gates applied during collection

Every source tile passed all four gates before entering the training set:

1. S2 not blank — file size > 5 KB, > 50% nonzero pixels.
2. S2 not cloudy — cloud_cover ≤ 30%, mean reflectance < 4000, NDVI > 0.1, < 20% saturated pixels.
3. High-resolution context tile available — used during collection only; not redistributed.
4. Agricultural land cover — green-edge density check; forest, water, urban, and rock tiles rejected (cropmarks form only in agricultural fields).

When in doubt, tiles were thrown out — the project rule is "50 clean examples beat 200 noisy ones."

Label epistemology

Labels are coarse and Phase-1 only. This dataset records a single token (Yes/No) per tile.

• A Yes label means the tile coordinates were drawn from an archaeological-cropmark record in one of the source databases above. It does not certify that an expert manually verified the specific Sentinel-2 image used here for cropmark visibility.
• A No label means the tile was drawn either from a confounder category (drainage, geological lineament, cultivation, natural feature) or from a random-offset negative ≥ 1 km from any known archaeological record.
• The label-noise floor is non-zero: source databases contain known false positives (~40–95% across studies), and our random-offset negatives may sit on undocumented archaeological sites.

The model trained on this dataset flags candidates for verification, not corrections of expert errors. Do not use the published model or this dataset for treasure-hunting or unauthorized site identification. Heritage protection laws apply in all three source jurisdictions.

Limitations

• Cereal-land bias — only agricultural cropland is represented. Pastoral, wooded, and urban landscapes are not in the training distribution.
• Geographic scope — UK, Norway, and Italy only. Cross-geography generalization to Spain, Denmark, France, etc. is not validated by this dataset alone (see project repo for separate cross-geography work).
• Classification-only — no bounding boxes, no captions, no multi-task data. Phase-1 training output.
• No temporal axis — single best-cloud-cover Sentinel-2 acquisition per tile. Multi-date temporal anomaly detection is part of our explanation pipeline but is not in this dataset.
• No coordinates published — reproducibility of the original collection requires re-querying the source APIs above with your own pipeline.

License

• Labels and metadata.csv: CDLA-Permissive-2.0 (see LICENSE).
• Sentinel-2 imagery: © European Union, modified Copernicus Sentinel data, Copernicus Sentinel Data Terms.
• Source attribution data (site categories) underlying the labels: see the per-source license links in "Source attribution."

Citation

If you use this dataset, please cite both the physics anchor and the dataset card:

@article{gravanis2026cropmarks,
  title   = {Physically-based modelling for retrospective detection of archaeological proxies (cropmarks)},
  author  = {Gravanis, E. and Agapiou, A.},
  journal = {Scientific Reports},
  year    = {2026},
  doi     = {10.1038/s41598-026-45441-0}
}

Training code

Fine-tuned with Liquid4All/leap-finetune (LoRA, VLM SFT) on Modal H100. Training configs (cropmark_classify_v4.yaml, cropmark_classify_phase1.yaml) live in the hackathon submission repository.