Sergey395/archaeology-external-expds

Archaeology: external datasets for auxiliary training

Описание на русском · Отчёт о новых источниках · Training index · Completion report

Selected archaeological imagery and geophysics, preserved as individual files with source metadata and derived candidate annotations. This collection is private. It does not include the competition's private training data or credentials.

Source	Prepared candidate annotations	Use
Boviolles	263 polygons	Magnetic anomaly pretraining
Vieil-Évreux	395 polygons	Electrical resistivity map pretraining
Chactún	7624 raster connected components across 2094 tiles	LiDAR building segmentation pretraining
UCL Remote Sensing DL	446 polygons across 154 images	Ruined structures in aerial/oblique imagery
Grotte di Castro	No target-ready annotations	Raw ERT/CCR; processing and annotation needed
Harlow Hill	No target-ready annotations	Research-only magnetometry/ERT originals

The prepared subset has 2250 image/label pairs and 8728 candidate polygon records/components. These are not 8728 independently verified archaeological objects. Published labels were converted, screened and geometrically repaired where indicated; no new expert archaeological ground truth was created. Improvement on the target models has not been measured.

Layout

• candidate_training/: TIFF + GeoJSON in modality/class naming conventions resembling the competition dataset.
• originals/: individual original files; archive provenance and completion manifests are under provenance/.
• reference_labels/: original geophysics classes, including excluded features.
• training_index.jsonl: image/label pairs, modality, coordinate space and split group.
• auxiliary_training/: separate UCL mound and Chactún platform image/label pairs and indexes; these labels are not target competition classes.
• supplementary_sources/: Yambol mound points and Sławków ERT/EM measurements; these are not segmentation training pairs.
• qa/: visual inspection previews.
• tools/: reproducible conversion/import tools; no credentials are embedded.

Coordinate compatibility: Chactún and UCL annotations use image pixel coordinates, with no invented geographic CRS. Their TIFFs do not provide geographic positioning. Geophysics preserves source projected coordinates. Boviolles' declared CRS is suspect despite raster/vector alignment; do not reproject or geographically fuse it until verified. A loader must honor the index's coordinate_space; this is not a drop-in geographically referenced competition submission.

Label caveats: source masks can merge touching buildings and split structures at tile boundaries. Chactún vectors were checked by exact rasterization back to the original masks. UCL class IDs differ between source train/test; mapping uses class names. Mounds were not automatically mapped to burial kurgans. Magnetic dipoles were not automatically converted into FindsPoints. Unlabelled geophysics pixels are not verified negative examples.

Attribution and licenses

• Boviolles / Vieil-Évreux: Lieven Verdonck (2025), Zenodo v2, DOI 10.5281/zenodo.15422270, CC BY 4.0. Changes: grayscale RGB rendering, GeoJSON conversion, geometry repair, candidate filtering and provenance fields.
• Chactún: Kokalj, Džeroski, Šprajc, Štajdohar, Draksler and Somrak, source dataset, data descriptor, CC BY 4.0. Changes: building masks polygonized in pixel coordinates; candidate class mapping and TIFF aliases.
• UCL: Remote Sensing DL Data, CC0 according to the source record. Full author metadata retained in provenance/. Changes: usable ruin polygons selected, invalid geometry screened/repaired, RGB TIFF rendering preserving dimensions. Original author markings preserved.
• Grotte di Castro: source dataset, CC BY 4.0. Author metadata retained in provenance. Selected ERT/CCR files are unmodified.
• Harlow Hill: Sam Bithell and Stefan Nielsen, Durham dataset, CC BY-NC-SA 4.0. Noncommercial/ShareAlike restrictions apply; not automatically eligible for commercial model training.

Licenses apply per source, not as one replacement license for the entire repository. Authors do not endorse these derived labels. Archaeoscape is excluded due to its no-redistribution terms.

The 2026-10-16 additions retain their own source licenses: UCL CC0; Chactún CC BY 4.0; Yambol CC BY-NC-SA 4.0; Sławków CC BY 4.0. See the linked Russian report and source metadata for conversion details. UCL and Chactún auxiliary polygons are in pixel coordinates with no geographic CRS; Yambol and georeferenced Sławków points are in EPSG:3857. Sławków's local survey coordinates remain local and have no assigned EPSG code.

Keep sites together when splitting train/validation. Preserve UCL source split metadata but do not assume geographic independence. Use external data for auxiliary pretraining followed by target-data fine-tuning; verify usefulness with the competition metric.