Fields of the World Benchmark Data

Training and evaluation data for models that delineate agricultural field boundaries. Each chip is a fixed 2 km cell of a shared, MGRS-derived grid, carrying label masks rasterized from official field boundary declarations, a train / validation / test assignment, and — in most collections — paired Sentinel-2 imagery. One collection per country or region.

This is version 2 of the benchmark introduced in Fields of The World (AAAI 2025). The boundaries come from the harmonized field boundary catalog, which republishes each country's own parcel declarations in one schema; the chips, masks and imagery here are cut from it with ftw-dataset-tools (ftwd) from a recipe committed in the source repository, so any collection can be rebuilt from the harmonized edition it names.

Browse the whole catalog · agent guide · the paper · version 1.0

Collections

3 collections · 17,120 chips · 1,553,660 field polygons

What each collection carries

stored means a four-band GeoTIFF clipped to the chip and published with it; linked means the chip's season item points at the whole Sentinel-2 scene on the source STAC API, for a reader to window. A chip with no scene still carries its masks and its split — pair it with imagery of your own, on the footprint in items.parquet.

Quick start

Everything here runs against the published files over HTTPS: no download, no credentials, no catalog walk.

Each collection publishes items.parquet, a stac-geoparquet mirror of every STAC item it holds, carrying a public URL for every asset. One query is therefore enough to build a training list — chip id, split, and the URLs of the rasters themselves:

Swap lu for any collection id in the table above. assets is a struct, so DESCRIBE SELECT assets.* FROM read_parquet(...) lists what a given collection's chips actually carry: where imagery is linked rather than stored, ask for assets.planting_visual.href and window the scene yourself.

To look before you query, open the data browser. It renders the collections, their chip footprints and the map styles each one ships; the Browse column above opens a single collection in it.

What a chip carries

Every chip is a STAC item on the FTW grid, so a cell id is the same square of ground in every collection. Each chip carries:

• Instance masks — one integer label per field polygon, for instance segmentation.
• Semantic masks — 2-class (field / not field) and 3-class (interior / boundary / not field), the two semantic label forms the FTW paper benchmarks.
• DECODE masks — a boundary mask and a normalized distance-to-boundary map, after Waldner et al. (2021). A distance surface lets a watershed separate touching fields that a binary extent mask merges into one. New in 2.0.
• A split assignment — train, val or test, drawn in 3×3 spatial blocks (80 / 10 / 10) so that neighbouring chips do not straddle two splits and a model cannot learn a test chip from its training neighbour.
• Field coverage, the share of the chip's area that is mapped field, and — where the harmonized source carries EuroCrops HCAT codes — the crop composition of the fields inside it.
• Sentinel-2 scenes, in the collections that have them: one from the planting window and one from the harvest window of the region's crop calendar, published as child items alongside the chip. Which collections carry imagery, and in what shape, is in the coverage table above.

What a collection carries

Alongside the chip items, each collection publishes the tables the chips were cut from and a mirror of the items themselves:

• <id>_fields.parquet, the field polygons used, and <id>_boundary_lines.parquet, their boundaries as lines.
• <id>_chips.parquet: every chip with its footprint, split assignment and field coverage.
• items.parquet: the stac-geoparquet mirror the quick start queries.
• chips.pmtiles and fields.pmtiles, with ready-made map styles under styles/, for looking at a collection before training on it.
• README.md, AGENTS.md, llms.txt and a thumbnail. A collection's README carries the numbers measured for that collection — split counts, the field-coverage distribution, imagery dates and cloud cover; its agent guide documents every column and ships worked queries with their results.

Limitations

Worth reading before publishing a number from this benchmark.

• Imagery is not uniform. See the coverage table above. A labels-only collection is still usable — every chip's footprint and grid cell are in items.parquet — but it is not a drop-in substitute for one with scenes.
• Labels are declarations, not observations. Masks come from parcels declared for one year, and the declaration year differs between collections. Parcels subdivided, merged or reshaped after that declaration are not reflected.
• Empty means unmapped, not fieldless. Area inside a chip with no polygon may be a field nobody declared. Chips on the edge of a source dataset may be only partly covered by it.
• Field coverage is skewed. Filter on ftw:field_coverage_pct when sparse chips would distort an evaluation; each collection's README gives its own distribution.
• Respect the splits. They are spatially blocked. Resampling chips at random leaks information between train and test.
• Cloud-free is not guaranteed. Scenes are selected against a crop calendar within a cloud threshold, so acquisition dates and residual cloud vary between chips — check ftw:planting_cloud_cover and ftw:harvest_cloud_cover.

Compared with Fields of the World 1.0

Version 1.0 remains published, unchanged, at kerner-lab/fields-of-the-world: 24 countries, some 70,000 chips and 1.6 million field polygons, one zipped folder per country. Version 2 is a rebuild rather than a re-release. It cuts every collection from the harmonized field boundary catalog instead of a one-off per-country conversion, so a collection can be regenerated when its source publishes a new edition; it adds the DECODE boundary and distance masks; and it publishes the chips as STAC with a queryable item mirror rather than folders of files. Each collection's own README states, in measured numbers, how it compares with the 1.0 folder it supersedes.

Version 2 is still being built out and does not yet cover everything 1.0 does — the collections table above is the whole of it. Use 1.0 for anything version 2 has not published yet.

Versioning

This edition is 2.0.0-alpha.1. While it is in alpha, collections are updated in place: a rebuild replaces the published files at the same URLs, and the updated stamp in a collection's collection.json says when that last happened. Treat any collection here as a moving target until 2.0.0 is released.

License

There is no single license for this catalog. Each collection carries the license of the field boundary data it was cut from — the License column above, and the collection's own page — and the masks derived from those boundaries are published under the same terms as their source. Check the collection you intend to use before redistributing it or the models trained on it. Sentinel-2 imagery is Copernicus data, free and open.

Citation

If you use this data in published work, please cite the benchmark paper and name the release you used.

For the data itself: Fields of the World Benchmark Data, version 2.0.0-alpha.1, Source Cooperative, accessed YYYY-MM-DD. Collections are updated in place while in alpha, so record the date you took them, and the updated stamp of each collection you used.

Contact

Open an issue for a problem with the data or a request for a country. Agents should start from the agent guide, the machine-readable tour of the same material; llms.txt is the short index.