Agricultural field boundaries produced by Trazo models for key geographies in South America for the 2023-2024 planting season. Post-processed.
Every number in this file is measured from the Parquet footers or quoted from the source it names. Nothing here is estimated.
Agricultural field boundaries across South America for the 2023–2024 planting season, delineated from Sentinel-2 imagery by the Trazo model suite. Four children:
Current collections combined: 17,923,469 boundaries in 23 files. Superseded eap-release: 18,370,419 in 30 files. Whole catalog: 36,293,888 in 53 files.
Start with trazo3-fields. It has the broadest coverage and the strongest overall scores.
Two access paths, and they are not interchangeable.
One file, no credentials, plain HTTPS. Every asset href in this catalog is
a relative path under https://data.source.coop/wri-data-lab/trazofields. A reader resolves it and reads the file
directly.
A whole collection, still no credentials, but s3://. A glob has to be
expanded, expanding it needs a directory listing, and plain HTTPS gives no
listing. DuckDB says so directly: Consider SET allow_asterisks_in_http_paths = true;. Turning that on does not help, because the pattern is then sent
literally and returns 404. Use the bucket scheme instead.
Source Cooperative is not AWS, so the endpoint is explicit, and addressing is path-style.
No credential is set in that secret. The product is public and reads anonymously.
There is no shared feature identifier across collections. Id is unique inside one file only, and the same ground appears in several collections as different geometry from a different model. Join spatially, not on a key.
trazo1-fields and trazo2-fields share one column set. trazo3-fields adds MapBiomas and Hansen forest-loss columns. eap-release carries a narrower set with no MapBiomas and no Hansen columns at all.
Counts are not comparable between models. Trazo1 and Trazo2 fragment fields more finely than Trazo3, so they report more boundaries over the same ground. Summing two collections double-counts the ground they share. Query one collection at a time.
The 10,968,017 figure in the technical note is not the size of anything you
can download. It is a one-model-per-region composite, and every cell of it
comes from eap-release, the superseded 2025 release. The three current
collections hold 17,923,469 boundaries between them.
Files sit directly in the collection directory. There is no per-country
subdirectory. The glob is <collection>/*.parquet, not <collection>/*/*.parquet.
There is no bbox covering column. Spatial pruning uses the native Parquet
GeospatialStatistics that GeoParquet 2.0.0 carries. A reader that looks only
for a bbox covering struct falls back to a full scan.
The GeoPackages are a convenience copy, not a second source. They carry all source columns for desktop GIS. They are large, they are not cloud-native, and they are not range-readable in a useful way.
DuckDB ST_Length_Spheroid returns silently wrong finite values, and
ST_Union_Agg segfaults on geometry sets this size. Neither is used here, and
neither should be trusted on this data.
The World Resources Institute produces these boundaries and publishes them. This catalog is the official publication, not a mirror. Source Cooperative stores and serves the bytes under the wri-data-lab account. Arizona State University co-produced the models. The Land and Carbon Lab and the Walmart Foundation funded the work.
Licence: CC-BY-4.0. Cite Grupp et al. (2026), Field Boundaries of South America, World Resources Institute and Arizona State University.