Agricultural field boundaries produced by Trazo models for key geographies in South America for the 2023-2024 planting season. Post-processed.
# Trazo Field Boundaries — Early Adopters Program release (2025)
> **Superseded data.** These are the field boundaries reported in the Trazo technical note,
> representing the state of the work in late 2025. For current work use the main collections —
> [`trazo3-fields`](../trazo3-fields/) is recommended for most use cases.
EAP stands for Early Adopters Program: these datasets were shared with end users who gave
structured feedback on them. They are published for three reasons — the technical note
documents them, they carry that user feedback, and together they show how the data product
evolved. Organised by model, mirroring the main catalog.
## Collections
| Collection | Model | Files | Fields |
|---|---|---:|---:|
| [eap-trazo1](eap-trazo1/collection.json) | Trazo1 | 10 | 6,639,770 |
| [eap-trazo2](eap-trazo2/collection.json) | Trazo2 | 10 | 6,092,989 |
| [eap-trazosa](eap-trazosa/collection.json) | TrazoSA | 8 | 2,915,878 |
| [eap-trazo3](eap-trazo3/collection.json) | Trazo3-era | 2 | 2,721,782 |
`eap-trazosa` is the **only published output of the TrazoSA model** anywhere in this catalog.
TrazoSA excludes the global Fields of The World baseline entirely, training only on South
American data, and is the most conservative of the models — it reports the fewest boundaries
in every region it shares with the others.
## Schema
Every file across all four collections shares one schema, so the whole release globs:
```sql
SELECT count(*) FROM read_parquet('eap-release/*/*.parquet');
```
| Column | Meaning |
|---|---|
| `det_method` | detection method; constant `auto-imagery` |
| `id` | feature id within the file; not stable across models or releases |
| `area` | field area in **hectares** |
| `perimeter` | field perimeter in **metres** |
| `geometry` | MultiPolygon boundary, EPSG:4326 |
This is a narrower attribute set than the released data: there is no MapBiomas class
(`mbmode24`) and none of the Hansen forest-loss columns found in `trazo3-fields`.
## Comparing counts between models
Field counts are **not directly comparable across models**. Trazo1 and Trazo2 fragment fields
more finely than Trazo3, so they report more boundaries over identical ground. For example
Piauí: 133,991 fields under Trazo2 against 71,037 under Trazo1 in this release, and 83,177
under Trazo3 in the current data. None of these is wrong; they are different models.
## Provenance notes
- All files were reprojected to EPSG:4326 from their source CRS (UTM zones 19S–23S; the
Bolivia layer was SIRGAS 2000 / Brazil Polyconic).
- `eap-trazo1/brazil_rio_grande_do_sul_2024.parquet` is the "grassland left in" variant —
grassland was not filtered out of this region as it was elsewhere.
- `eap-trazo3` holds two superseded datasets, Bolivia (1,270,024) and Mato Grosso (1,451,758),
which are the versions the technical note's coverage table reports. Both have since been
reprocessed; see `trazo3-fields` for current versions.
- Bolivia's `area` and `perimeter` were recomputed in South America Albers Equal Area, because
its source carried only distorted polyconic values. All other files carry their source values.
## Caveats
Boundaries are model-derived and have not been field-verified. Object-level metrics on the
model cards use a punishing 50% IoU matching threshold that systematically depresses them —
treat the pixel metrics as the headline. Models are optimized for annual crops; pastures, tree
crops and agroforestry are not tailored for. Per the model cards, these boundaries are **not
intended to be used alone** to support deforestation- or conversion-free claims.
## References
- [Technical note (Grupp et al. 2026)](https://www.wri.org/research/field-boundaries-south-america)
- [Trazo on the ground — end-user feedback](https://wri.github.io/trazo/landing/map.html)
- [Trazo model cards](https://huggingface.co/collections/worldresourcesinstitute/trazo-models)