Fields of the World — Global
The first global, wall-to-wall agricultural field-boundary dataset at 10 m resolution: ~1.6 billion field polygons per year — 1.63 billion for 2024 and 1.58 billion for 2025 (~3.2 billion polygon records across both years) — spanning 195 countries and territories, produced by applying the PRUE field-boundary segmentation model (a U-Net with an EfficientNet-B7 encoder, trained on the Fields of The World benchmark) to cloud-free Sentinel-2 mosaics. Published openly under CC-BY-4.0 by Taylor Geospatial and collaborators (Microsoft AI for Good, ASU, WashU in St. Louis, Oregon State, Clark).
Paper: Robinson et al. 2026, The first global agricultural field boundary map at 10 m resolution (arXiv:2605.11055).
Explore this catalog
- STAC catalog (root): https://data.source.coop/ftw/global-data/catalog.json
- Browse interactively: Portolan browser
- For AI agents:
llms.txt— a machine-readable description of the whole dataset. Point Claude Code or Gemini CLI at it and ask it to query the data, build interactive maps, or generate charts. Each collection also has its ownllms.txt. - Source repository: https://github.com/fieldsoftheworld/ftw-data-catalog
Storage: the public URL base https://data.source.coop/ftw/global-data/ is physically backed by
s3://us-west-2.opendata.source.coop/tge-labs/ftw-global-data/ (anonymous read).
Collections
Key facts
- 1.63 billion field polygons for 2024 and 1.58 billion for 2025 (~3.2 B across both years), over 195 countries and 574 admin partitions, at 10 m resolution. Each year is an independent global prediction, so the per-year count (~1.6 B) — not the both-years sum — is the number of distinct fields in a year.
- Every vector polygon carries a
confidencescore (0–100), sampled at the polygon's representative point from the 500 m PRUE confidence layer and rescaled (raw / 0.578178 × 100, the same scaling the FTW inference app uses). - Model: PRUE (U-Net / EfficientNet-B7), trained on the CC-BY subset of the Fields of The World benchmark (24 countries).
- A field here is a remote-sensing field unit (a connected component of predicted field-interior pixels), not a cadastral/legal parcel. This is not a land-tenure product.
- Outputs are fiboa/vecorel-compliant GeoParquet (vectors) and Cloud-Optimized GeoTIFFs / Zarr (rasters).
- Validation (paper): mean pixel-level recall 0.85 over 24 countries (14 > 0.90); confidence-model leave-one-country-out mean AUC 0.842.
Data products
All files are anonymous-read on Source Cooperative and work directly over HTTP — no account or API key.
Field boundary predictions (GeoParquet + PMTiles)
The PRUE model runs over the Sentinel-2 feature composites to produce per-pixel field probabilities;
vectors are derived by thresholding and polygonizing into fiboa/vecorel GeoParquet v1.1.0, then
partitioned one file per country (195 countries / 574 files, ~210 GB; the nine largest countries —
e.g. the US, India, China, Brazil — are split by admin subdivision) and enriched with a per-polygon
confidence column (0–100). Each polygon also carries metrics:area, metrics:perimeter, and a
determination:datetime (2024 or 2025). Query a country partition directly with DuckDB — only the
needed bytes are fetched:
For web maps, the collection also exposes global PMTiles (a 2024 and a 2025 layer) and per-country PMTiles, with MapLibre styles for plain green boundaries and for shading each field by its confidence.
Prediction confidence & quality (500 m)
Global 500 m rasters quantifying where the 10 m field predictions can be trusted (confidence, field/boundary density, entropy, cropland consensus, precision/recall). Read a window of the confidence COG — only the needed bytes are fetched:
Or load the whole collection lazily as an xarray stack:
Features — Sentinel-2 composites (COG & Zarr)
Planting- and harvest-season median composites over ~5–10 quality-masked Sentinel-2 scenes (the
model's inputs), as per-tile COGs (~22.7 k tiles/year) and a single global EPSG:4326 Zarr V3 mosaic
at 8.983119e-5° (~10 m at the equator):
Prediction probabilities (Zarr)
The raw PRUE softmax bands [non_field_background, field, field_boundaries] on the same grid as the
features (so they stack), from which the vectors are thresholded at 0.5 and polygonized.
Caveats
- The field-density
_filteredvariant's exact threshold/method is pending author confirmation. - The confidence layer is conservative outside the FTW training distribution (e.g. smallholder systems): real fields there may receive low confidence. Prefer the unfiltered density + continuous confidence over a hard threshold in such regions.
- Polygons are remote-sensing field units, not legal parcels; one parcel may map to many polygons or none.
License
CC-BY-4.0.
Cite
Robinson, C., Muhawenayo, G., Khanal, S., Fang, Z., Corley, I., Tárano, A. M., Estes, L., Marcus, J., Jacobs, N., Kerner, H., Becker-Reshef, I., & Lavista Ferres, J. M. (2026). The first global agricultural field boundary map at 10 m resolution. arXiv:2605.11055.
Contact
isaac.corley@taylorgeospatial.org
Published with Portolan.