South American Field Boundaries

Agricultural field boundaries produced by Trazo models for key geographies in South America for the 2023-2024 planting season. Post-processed.

Contents

root

trazo3-fields

Details

Visibility: Public
Owner: WRI Data Lab
Created: 22 Jun 2026
Last Updated: 17 Jul 2026

Analytics

28 days

Downloads

11,698

Data served

909.5 GB

Countries

28-day downloads

9 Jul 2026

5 Aug 2026

Analytics

README

Trazo3 Field Boundaries — South America (2023–2024)

Agricultural field boundaries across South America — Argentina (Chaco, Córdoba), Bolivia, nine Brazilian states, and Paraguay (Chaco and non-Chaco) — delineated from Sentinel-2 imagery by the Trazo3 model for the 2023–2024 planting season. This is the recommended collection for most use cases: broadest coverage and strongest overall performance (79.17% pixel IoU on the Mato Grosso test set). Includes Hansen forest-loss attributes per field. Files are named {model}{country}{region}_{year}; read them all with the glob trazo3-fields/*.parquet. A PMTiles tileset, default style and thumbnail are provided; full-attribute GeoPackages for desktop GIS are in geopackages/.

Spatial Coverage

Bounding Box: [-68.994184, -35.321848, -41.264648, 0.878872]

Temporal Coverage

Start: 2023-07-01T00:00:00Z
End: 2024-07-31T23:59:59Z

Schema

Column	Type	Description
Id	int32
hansen_covered_area	double
hansen_loss_area	double
mode_year	double
firstyear	double
firstyearmajority	double
deforestarea0104	double
deforestarea0509	double
deforestarea1014	double
deforestarea1520	double
deforestarea2124	double
mbmode24	double
mbcov_area_2024

Files

File	Size	Checksum
./README.md	-	-
./trazo3_bolivia_2024.parquet	-	-
./trazo3_brazil_mato_grosso_do_sul_2024.parquet	-	-
./trazo3_brazil_mato_grosso_2024.parquet	-	-
./trazo3_brazil_rio_grande_do_sul_2024.parquet	-	-
./trazo3_paraguay_non_chaco_2024.parquet	-	-
./trazo3.pmtiles	-	-
./llms.txt	-	-
./thumbnail.png	-	-
./trazo3_argentina_chaco_2024.parquet	-	-
./trazo3_argentina_cordoba_2024.parquet	-	-
./trazo3_brazil_goias_2024.parquet	-

Quick Start

STAC Metadata

root: ../catalog.json
parent: ../catalog.json

Source

https://huggingface.co/worldresourcesinstitute/Trazo3

Processing Notes

Recommended for most use cases. Trazo3 is the most recent and broadest-coverage model in the Trazo suite, covering Argentina (Chaco, Córdoba), Bolivia, nine Brazilian states, and Paraguay (Chaco and non-Chaco).

U-Net with EfficientNet-B3 backbone over Sentinel-2 chips (256×256 px, RGB + NIR, two seasonal windows). Trained on the full global FTW dataset, all 17 South American soy ecoregions, partner data from Paraguay, and 400 challenging Mato Grosso samples selected by active learning — which is what lifts its Mato Grosso score above the other models.

Test set	Pixel IoU	Pixel precision	Pixel recall
Mato Grosso (2,462 fields)	79.17%	86.43%	90.41%
SA Soy (1,400 fields)	74.96%	87.55%	83.90%
Chiquitania (3,961 fields)	68.77%	81.68%	81.31%

Trazo3 model card

Bulk access. Files are partitioned by country; read them all with the glob trazo3-fields/*/*.parquet.

Columns include mbmode24 (MapBiomas 2024 agriculture classification mode) and Hansen forest-loss areas by period (deforestarea0104, deforestarea0509, deforestarea1014, deforestarea1520, deforestarea2124), in square metres.

Citation

Grupp et al. (2026). Field Boundaries of South America. World Resources Institute & Arizona State University.

Attribution

Trazo / World Resources Institute

License

CC-BY-4.0

Contact

Tristan Grupp — World Resources Institute tristan.grupp@wri.org

Known Issues

Trazo3 leads on Mato Grosso (79.17% Pixel IoU) and is strong on Chiquitania. On the SA Soy test set the four models sit within about three points of each other (Trazo3 74.96%, Trazo1 77.09%) — a minor difference, but where soft boundaries matter (those defined by two abutting field interiors rather than a hard feature such as a road) trazo1-fields may be more sensitive, and trazo2-fields is the best evaluated and characterized.

Object-level metrics on the model card (object precision 26.90%, object recall 24.75% on Mato Grosso) use a punishing 50% IoU object-matching threshold, which systematically depresses them. Treat the pixel metrics as the headline figure.

Optimized for annual crops; pastures and tree crops are possible but not tailored for. Not intended to be used alone to support deforestation- or conversion-free claims. Boundaries are model-derived and have not been field-verified.

The Argentina/Córdoba file was reconstructed from the published GeoPackage during the 2026-07 migration; it had no GeoParquet counterpart in the original release.

Generated by Portolan from STAC metadata and .portolan/metadata.yaml