Trazo1 Field Boundaries — Brazil (2023–2024)
Agricultural field boundaries for Mato Grosso do Sul and Paraná, Brazil, delineated from Sentinel-2 imagery by the Trazo1 model for the 2023–2024 planting season. Trazo1 uses heavy boundary weighting (0.88/0.04/0.08), the most sensitive model to soft boundaries and a good choice for large, uniform row-crop systems; it posts the highest SA Soy pixel IoU (77.09%) but is weaker outside soy. Files are named {model}{country}{region}_{year}; read them all with the glob trazo1-fields/*.parquet. Full-attribute GeoPackages for desktop GIS are in geopackages/.
Spatial Coverage
- Bounding Box: [-57.945727433991976, -27.20507186738115, -48.882768560452064, -18.081309522931576]
Temporal Coverage
- Start: 2023-07-01T00:00:00Z
- End: 2024-05-31T23:59:59Z
Schema
Files
Quick Start
STAC Metadata
- root:
../catalog.json - parent:
../catalog.json
Source
https://huggingface.co/worldresourcesinstitute/Trazo1
Processing Notes
Most sensitive to soft boundaries — those defined by two abutting field interiors rather than a hard feature such as a road. May be the best choice in large-scale row cropping systems where such boundaries are subtle. Covers the Brazilian states of Mato Grosso do Sul and Paraná.
U-Net with EfficientNet-B3 backbone over Sentinel-2 chips (256×256 px, RGB + NIR, two seasonal windows). Trained on the FTW baseline plus 559 new South American chips drawn from 10 of the 17 soy-producing ecoregions — the narrowest training set of the suite, which shows in its weaker performance outside soy.
Trazo1 is distinguished by heavy boundary weighting in its loss: class weights 0.88 boundary / 0.04 background / 0.08 interior (LR 0.001, 100 epochs), against the rebalanced 0.75 / 0.05 / 0.20 used by Trazo2 and TrazoSA. That is what makes it the most boundary-sensitive model of the suite.
Trazo1 posts the highest SA Soy Pixel IoU of any Trazo model (77.09%, against 47.69% for the FTW baseline). Trazo1 model card
Bulk access. Files are partitioned by country; read them all with the glob
trazo1-fields/*/*.parquet.
Each file carries the mbmode24 column: the mode of MapBiomas' 2024 agriculture
classification (see Table A1 of the technical note).
Citation
Grupp et al. (2026). Field Boundaries of South America. World Resources Institute & Arizona State University.
Attribution
Trazo / World Resources Institute
License
Contact
Tristan Grupp — World Resources Institute tristan.grupp@wri.org
Known Issues
Trazo1 is strongest on soy (77.09% Pixel IoU on SA Soy) but weaker outside it — 57.76% on
Chiquitania, below Trazo2 (66.61%) and Trazo3 (68.77%). Outside soy-dominated row cropping,
prefer trazo3-fields; Trazo2 is the best evaluated and characterized model.
Object-level metrics on the model card (object precision 26.15%, object recall 22.72% on SA Soy) use a punishing 50% IoU object-matching threshold, which systematically depresses them. Treat the pixel metrics as the headline figure.
Trazo1 exposes the same attribute columns as Trazo2 — the models themselves are distinct, with different training data and class weighting. Trazo3 adds Hansen forest-loss columns.
Optimized for annual crops (soy, corn); pasture and tree crop performance is untested. Agroforestry and other matrixed contexts may be undetectable at Sentinel-2 resolution. Boundaries are model-derived and have not been field-verified.
Generated by Portolan from STAC metadata and .portolan/metadata.yaml