Trazo: Agricultural Field Boundary Models for South America
Trazo is a suite of deep learning models, training data, and methods for automated agricultural field boundary delineation in South America, built on the
Fields of the World (FTW) baseline architecture. These four model checkpoints are tailored to the diverse agricultural landscapes of South America, with a focus on soy-producing ecoregions. They are released as open-source digital public goods under a
CC-BY-4.0 license.
Trazo is the Spanish word for brushstroke, from the verb trazar — to trace, to draw, to plot. By tracing fields by hand in satellite imagery, we choose the most powerful samples for creating robust, generalizable field boundary detection models. Trazo also echoes the English word "trace": to follow a path, uncover origins, and make hidden connections legible.
These models were produced as part of a collaboration with
World Resources Institute (WRI) and
Arizona State University with support from the
Land and Carbon Lab. This work was funded through a
Walmart Foundation grant.
Read and cite the
technical note.
trazo2
This model is Trazo2.
Trained on the full global FTW dataset combined with all 17-ecoregion South American data and partner data. Tests whether combining global coverage with regional data improves generalization. Uses the rebalanced class weights, higher learning rate (0.003), and up to 300 epochs with early stopping.
Intended Use
Trazo models are designed to:
- Delineate agricultural field boundaries from Sentinel-2 imagery across South America
- Support supply chain traceability and sustainability compliance (e.g., the EU Deforestation Regulation / EUDR)
- Enable deforestation risk assessment at the field level
- Serve as pretrained checkpoints for fine-tuning to new regions or crop systems
Architecture
All Trazo models share the same architecture inherited from the FTW baseline:
| Component | Detail |
|---|
| Architecture | U-Net |
| Backbone | EfficientNet-B3 |
| Input | Sentinel-2 chips, 256×256px, RGB + NIR, 2 seasonal time windows |
A key change across Trazo variants (SA, 2, 3) relative to the FTW baseline is a rebalanced loss weighting. The FTW baseline used class weights of 0.88 (boundary) / 0.04 (background) / 0.08 (interior), which caused systematic boundary over-prediction. Trazo models reduce the boundary weight to 0.75 (boundary) / 0.05 (background) / 0.20 (interior) — except the initial trazo checkpoint, which retains the closer-to-baseline weights.
Model Variants
Four checkpoints are provided. They share the same architecture but differ in training data composition, class weighting, learning rate, and training budget. Together they allow users to select the right trade-off between global generalization and South America–specific performance.
trazo1
The initial Trazo checkpoint. Trained from the FTW baseline and 559 new South American chips drawn from 10 of the 17 soy-producing ecoregions (the subset fully processed at the time of first training). Trained at a learning rate of 0.001 with the original FTW-style class weights.
- Achieves 77%+ pixel IoU on South American soy fields, vs. 48% for the FTW baseline
- Higher boundary class weight makes it more sensitive to boundaries
- Well-suited to large, uniform row-crop systems
trazoSA
Trained on South American ecoregion data and partner data from Mundo Agropecuario in Paraguay, plus Brazilian data from the FTW baseline — the full set of 909 ecoregion chips spanning all 17 ecoregions — without the broader global FTW training data. Isolates the effect of South America–specific training data and a highly consistent approach to a field boundary definition.
- Tends to achieve the highest pixel precision and object precision
- Isolates South American agriculture and a consistent approach to field boundaries definition
trazo2
Trained on the full global FTW dataset combined with all 17-ecoregion South American data and partner data. Tests whether combining global coverage with regional data improves generalization. Uses the rebalanced class weights, higher learning rate (0.003), and up to 300 epochs with early stopping.
trazo3
Extends trazo2 with an additional 400 challenging samples from Mato Grosso, Brazil, selected via an active learning pilot study. Challenging samples were identified using four query strategies:
- Land cover discrepancy between Trazo predictions and MapBiomas agricultural extent (interior + boundary classes)
- The same discrepancy using the interior class only
- Lowest model confidence chips
- Chips at low confidence agricultural areas
Adding these samples improved object-level recall in particular — recovering more complete field objects in complex, large-scale agricultural landscapes.
- Best overall performer for Mato Grosso
- Recommended when object recall is the key metric
- Trained on median composite imagery and therefore highly applicable for inference with median() windowA and windowB
Model Comparison
All three later variants (trazoSA, trazo2, trazo3) were trained with identical hyperparameters and differ only in training data composition, allowing a comparison of data contribution.
| Model | Class weights (Bd / BG / I) | Epochs / LR | Training data |
|---|
| FTW | 0.88 / 0.04 / 0.08 | 100 / default | Global FTW baseline |
trazo | 0.88 / 0.04 / 0.08 | 100 / 0.001 | FTW + 10-ecoregion SA (559 chips) |
trazoSA | 0.75 / 0.05 / 0.20 | 300 (ES) / 0.003 | 17-ecoregion SA + partner data |
trazo2 | 0.75 / 0.05 / 0.20 | 300 (ES) / 0.003 | Full FTW + 17-ecoregion SA + partner data |
trazo3 | 0.75 / 0.05 / 0.20 | 300 (ES) / 0.003 | Full FTW + 17-ecoregion SA + partner data + 400 AL samples |
Bd = boundary, BG = background, I = interior. ES = early stopping.
Training data composition
| Training Data | FTW | trazo | trazoSA | trazo2 | trazo3 |
|---|
| FTW global baseline | ✓ | ✓ | — | ✓ | ✓ |
| 10-ecoregion SA chips | — | ✓ | — | — | — |
| 17-ecoregion SA chips | — | — | ✓ | ✓ | ✓ |
| Mundo Agropecuario (Paraguay) | — | — | ✓ | ✓ | ✓ |
| Active learning samples (Mato Grosso) | — | — | — | — | ✓ |
Full training data, annotation protocols, and ecoregion coverage are documented on the
technical note.
Recommended Usage
| Use case | Recommended model |
|---|
| South American-specific agricultural systems | trazoSA |
| Mato Grosso or complex landscapes, median() composite input imagery | trazo3 |
| Large uniform row-crop systems, boundary-sensitive tasks | trazo |
| Fine-tuning starting point for a new region | Any checkpoint + full fine-tuning, then MagMax merge |
Users working in geographies not covered by this release, or needing boundaries updated to a different season or year, can apply any of the four checkpoints directly using the
inference pipeline. For regions with distinct field typologies, fine-tuning on locally collected annotations will substantially improve performance. Even a modest set of high-quality samples (on the order of 50 annotated chips) is sufficient to improve pixel IoU from the FTW baseline of ~42% to above 80% with full fine-tuning.
When
fine-tuning while preserving cross-region generalization, use MagMax after full fine-tuning by merging with the base checkpoint. Full fine-tuning on a geographically constrained sample risks catastrophic forgetting; MagMax retains and in some cases improves cross-region performance.
Limitations
- Models are strongest for annual crops (soy, corn, other row crops). Detection of pastures, tree crops, plantations, and mosaicked systems is possible but not yet tailored nor rigorously tested for accuracy.
- Performance may degrade in regions, seasons, or cropping systems underrepresented in training data.
- Agroforestry and other matrixed agricultural contexts may be undetectable at Sentinel-2 resolution — outputs should be used with caution in these settings.
- Field boundaries are not intended to be used alone to support voluntary deforestation-and-conversion free (DCF) claims. In line with best practice from the Accountability Framework Initiative (AFi), DCF claims should be made at the farm or property level.
- It is the responsibility of supply chain actors to verify and assess AI-generated field boundaries for their specific use cases.
Citation
If you use these models, please cite the WRI Trazo Technical Note and the FTW benchmark:
1@techreport{grupp2026trazo,
2 title={Field Boundaries of South America},
3 author={Grupp, Tristan and Elmustafa, Amna and Zvonkov, Ivan and Tarano, Ana M. and
4 Jupiter, Jaelah and Esipova, Elena and Blevins, Hunter and Nguyen, Catlynh and
5 Tenorio Fenton, Sofia and Winchester, Caroline and Stolle, Fred and
6 Rognstad, Rhiannon L. and Kerner, Hannah},
7 institution={World Resources Institute},
8 type={Technical Note},
9 year={2026},
10 month={June},
11 day={1},
12 doi={10.46830/writn.25.00033}
13 url={https://doi.org/10.46830/writn.25.00033
14}
1@article{kerner2025ftw,
2 title={Fields of The World: a machine learning benchmark dataset for global agricultural field boundary segmentation},
3 author={Kerner, Hannah and Chaudhari, Snehal and Ghosh, Aninda and Robinson, Caleb and Ahmad, Afri and Choi, Edward and Jacobs, Nathan and Holmes, Chris and Mohr, Matthias and Dodhia, Rahul and others},
4 journal={Proceedings of the AAAI Conference on Artificial Intelligence},
5 volume={39},
6 number={27},
7 pages={28151--28159},
8 year={2025},
9 doi={10.1609/aaai.v39i27.35034}
10}
License
Released under
CC-BY-4.0. You are free to use, share, and adapt this work for any purpose, including commercially, provided that Trazo is cited.
Links
Performance
Results across three South American test sets. All metrics are percentages.
SA Soy test set (1,400 fields, 30 chips across soy growing areas of South America)
| Metric | FTW | trazo | trazoSA | trazo2 | trazo3 |
|---|
| Pixel IoU | 47.69 | 77.09 | 76.91 | 76.05 | 74.96 |
| Pixel Precision | 71.39 | 92.24 | 93.05 | 91.01 | 87.55 |
| Pixel Recall | 58.95 | 82.44 | 81.59 | 82.22 | 83.90 |
| Object Precision | 5.50 | 26.15 | 38.61 | 34.83 | 30.82 |
| Object Recall | 9.02 | 22.72 | 24.37 | 28.37 | 34.59 |
Mato Grosso test set (2,462 fields, 100 chips)
| Metric | FTW | trazo | trazoSA | trazo2 | trazo3 |
|---|
| Pixel IoU | 38.98 | 66.57 | 75.10 | 76.27 | 79.17 |
| Pixel Precision | 52.17 | 71.26 | 87.35 | 87.19 | 86.43 |
| Pixel Recall | 60.65 | 91.00 | 84.26 | 85.90 | 90.41 |
| Object Precision | 2.23 | 19.39 | 30.33 | 25.37 | 26.90 |
| Object Recall | 6.58 | 14.14 | 16.20 | 16.37 | 24.75 |
Chiquitania test set (3,961 fields, 100 chips)
| Metric | FTW | trazo | trazoSA | trazo2 | trazo3 |
|---|
| Pixel IoU | 27.27 | 57.76 | 69.59 | 66.61 | 68.77 |
| Pixel Precision | 36.42 | 67.76 | 88.40 | 85.56 | 81.68 |
| Pixel Recall | 52.05 | 79.65 | 76.59 | 75.04 | 81.31 |
| Object Precision | 3.88 | 25.62 | 33.65 | 32.00 | 28.00 |
| Object Recall | 8.82 | 25.92 | 24.57 | 27.70 | 35.73 |
Bold = best value per metric within each test set. Pixel-level metrics measure per-pixel segmentation agreement. Object-level metrics require correct detection, separation, and delineation of complete field instances.
Note on interpreting metrics: Object precision and recall are systematically lower than pixel-level metrics; a single merged boundary or small geometric error can cause a well-segmented field to be counted as an object-level miss. For area estimation or land cover analysis, pixel IoU is most informative. For applications where parcel identity matters (e.g., supply chain traceability), object recall is key when combined with pixel IoU.
