Agricultural field boundaries produced by Trazo models for key geographies in South America for the 2023-2024 planting season. Post-processed.
# Trazo Field Boundaries of South America
> 10.9 million agricultural field boundaries across South America for the 2023–2024 planting
> season, delineated from Sentinel-2 imagery by the Trazo model suite. Published by the World
> Resources Institute and Arizona State University under CC-BY-4.0.
This is a Portolan catalog (a STAC profile) of cloud-native GeoParquet. Every file is
GeoParquet 2.0 with native Parquet GEOMETRY logical type, ZSTD compression, Hilbert spatial
ordering, bbox covering metadata, and ~25,000-row row groups, so it supports HTTP range
requests, row-group skipping, and spatial filter pushdown. Query it directly over the network
with DuckDB — no download required.
## Collections
Three collections, one per model. The models share a U-Net / EfficientNet-B3 architecture
inherited from Fields of The World, but they are **distinct models** with different training
data and class weighting.
- [trazo3-fields](trazo3-fields/collection.json): **Recommended for most use cases.** Argentina
(Chaco, Córdoba), Bolivia, nine Brazilian states, Paraguay (Chaco, non-Chaco). 14 files,
6.68M fields. Broadest coverage; 79.17% pixel IoU on the Mato Grosso test set. Includes
Hansen forest-loss attributes and a PMTiles tileset for web mapping.
- [trazo2-fields](trazo2-fields/collection.json): Five Brazilian states (Goiás, Pará, Piauí,
São Paulo, Tocantins) and Paraguay. 7 files. The best evaluated and characterized model.
- [trazo1-fields](trazo1-fields/collection.json): Mato Grosso do Sul and Paraná, Brazil.
2 files. Heaviest boundary weighting (0.88), so most sensitive to soft boundaries; highest
SA Soy pixel IoU (77.09%) but weakest outside soy.
## File organization
Files sit directly in each collection, named `{country}_{region}_{year}.parquet` — for example
`trazo3-fields/brazil_mato_grosso_2024.parquet`. The country prefix groups regions when sorted.
Read a whole collection with a glob:
```sql
SELECT count(*) FROM read_parquet('trazo3-fields/*.parquet');
```
All files within a collection share an identical schema and column order, so globbing works
under strict schema matching.
## Key columns
- `geometry` — MultiPolygon field boundary, EPSG:4326
- `Id` — automatically generated feature id; no external meaning, not stable across releases
- `mbmode24` — area-weighted majority (mode) MapBiomas agriculture class for 2024; class codes
are in Table A1 of the technical note
- `mbcov_area_2024`, `mbvalid_area_2024` — QA columns (MapBiomas raster coverage/validity, m²)
- trazo3 only: `hansen_covered_area`, `hansen_loss_area`, `mode_year`, `firstyear`,
`firstyearmajority`, and `deforestarea0104` / `0509` / `1014` / `1520` / `2124` — Hansen
forest-loss areas by period, **in square metres**. `firstyear` is the earliest year
accounting for ≥33% of loss area; `firstyearmajority` the earliest year exceeding 50%.
A Hansen `lossyear` of 0 means "no detected loss", not "cleared before 2001".
## Important caveats
- Boundaries are model-derived and have **not** been field-verified.
- Object-level metrics reported on the model cards use a punishing 50% IoU object-matching
threshold, which systematically depresses them; treat the pixel metrics as the headline.
- Models are optimized for annual crops (soy, corn, other row crops). Pastures, tree crops and
agroforestry are not tailored for and may be undetectable at Sentinel-2 resolution.
- Per the model cards, these boundaries are **not intended to be used alone** to support
voluntary deforestation- or conversion-free claims.
## References
- [Technical note (Grupp et al. 2026)](https://www.wri.org/research/field-boundaries-south-america)
- [Trazo model cards](https://huggingface.co/collections/worldresourcesinstitute/trazo-models)
- [Trazo training data](https://source.coop/wri-data-lab/trazoannotations)
- [Fields of The World](https://fieldsofthe.world/)