Fields of the World — Global Data (2nd Edition)

2nd Edition of the Fields of the World (FTW) global field-boundary predictions for 2017 to 2025: 1,242,648,258 predicted field polygons as cloud-native GeoParquet, and 67,197 field/boundary-probability tiles (25.8 TB) as Cloud-Optimized GeoTIFFs. Both derive from the TGE Labs Sentinel-2 quarterly cloudless mosaics, and both are part of Fields of the World.

Examples

1st and 2nd edition

Sorriso, Brazil, 2025 Q3. Large fields, same imagery in both panels.

Normandy, France, 2025 Q3. Irregular fields and hedgerows, same imagery in both panels.

2018 and 2025

Matopiba, Brazil. New rectangular field blocks.

West Bahia, Brazil. Rectangular fields replaced by center pivots.

Imagery is Copernicus Sentinel data 2017–2025 from the TGE Labs quarterly mosaics, outlines are FTW predictions, and these are selected examples, not an accuracy assessment.

Three ways in

• The interactive map lets you pan to a region, switch years, and compare the predictions with the imagery they came from.
• The Portolan browser walks the STAC tree and shows what each year contains.
• Source Cooperative hosts the files for download, and you can also read them in place over HTTP range requests from https://data.source.coop/ftw/global-data-2e/.

AGENTS.md beside this file is the guide for agents. Read Limitations before drawing conclusions from any of these numbers.

Data license: CC-BY-4.0

Collections

One collection per year in each tree, with both products on the same footprint.

The vector product is one GeoParquet file per UTM zone per year, at vector/{year}/zone=NN/utm{NN}.parquet, with a PMTiles archive per year for maps. The vector README has a runnable query.

The raster product is one COG per tile at 2.5 m, at raster/{year}/…/{tile}.tif, with field (band 1) and boundary (band 2) probabilities. All years share one tile grid, so a tile can be compared across years. Find tiles in the raster index. The raster README has the details.

The vectors are polygons derived from the same probabilities as the rasters, so the two are not independent measurements.

How this was made

The FTW unet_balanced_fp32.onnx model ran on the Sentinel-2 quarterly mosaics (four quarters, 10 m) and produced field and boundary probabilities on a 2.5 m grid, which are the rasters. Nodata gaps in the mosaics were filled from a 31-pixel window before inference, with TF32 off and no test-time augmentation. BoundaryVote (nbg-pb-h0.01-t0.5+R25+F10+G2+A900+q1) turned the probabilities into polygons, and parcels over 5 km², parcels that are mostly inland water and parcels on the sea were removed. The full pipeline is documented in pipeline/README.md.

Limitations

These are model predictions, not a survey. A field here is a remote-sensing field unit, a connected component of predicted field-interior pixels, and not a cadastral or legal parcel. This is not a land-tenure product, and one legal parcel may map to many polygons or to none. Counts, areas and perimeters are predicted quantities that carry the model's errors. See Fields of the World for the project and its definitions.

• The score column is the model's mean field probability inside the parcel (0 to 100), not a calibrated probability, so a score of 80 is not an 80% chance that the parcel is real.
• Predictions are weaker outside the training distribution, for example in smallholder systems. Prefer the continuous score over a hard threshold there.
• Coverage is cropland-gated, so it is not global. Only MGRS tiles with at least 1% cropland were processed, so an empty region means the pipeline never ran there. It does not mean the pipeline found no fields.
• Inside a processed tile, only the water, sea and size rules described above remove parcels, so scrub and built-up ground can still carry predicted parcels.
• Where the Sentinel-2 mosaics have large nodata gaps, mainly the sparse 2017–2019 mosaics, detections are probably under-reported even after the gaps were filled. Parcel counts for 2017, 2018 and 2019 are about 9–12% below 2025, and part of that gap reflects the missing imagery.
• Each year is an independent prediction. A parcel id carries no meaning across years, so compare years in the vectors with a spatial join. The rasters share one grid and compare per pixel.

Coordinate systems

Vector GeoParquet is WGS 84 lon/lat (EPSG:4326) in every zone file, so ST_Area on geometry returns square degrees, so read metrics:area (m²) instead. Each COG is in its own tile's UTM zone, so a mosaic across zones needs a warp. PMTiles are Web Mercator (EPSG:3857).

Reading an area across tiles

See the raster README for a short rasterio example that mosaics tiles for a lon/lat box.

Fixing this metadata

catalog/ in fieldsoftheworld/ftw-global-data-catalog is this catalog, and the same repository holds the processing pipeline. Report problems in the issue tracker.