Overview

Trazo is a deep learning training dataset and suite of field boundary segmentation models developed to support sustainable supply chain monitoring across South America. Built on the Fields of The World (FTW) baseline dataset and model architecture, Trazo extends FTW to the varied and challenging agricultural landscapes of South America. Read about Trazo.

Use case: These annotations can be used directly alongside the Fields of The World benchmark to train or fine-tune field boundary segmentation models. The dataset is formatted for drop-in compatibility with the FTW training pipeline (256×256 pixel chips, 3-class labeling: interior, boundary, background. 2-class and instance are also available).

Folders are grouped by ecoregion, sampling strategy, or the provenance of a project partner with 90% allocated to training and 10% to validation. The test set folders are allocated 100% to testing. The distinction matters because test set folders were truly randomly sampled, making them the most statistically rigorous option for model evaluation. Other folders (e.g., ecoregion, Mundo Agropecuario, active learning samples) were selected based on poor model performance, a query strategy, or partner availability and are therefore better suited for training and validation than for unbiased testing. Though all folders can technically be used for any split. To reassign splits for your use case, regenerate the parquet files using the build_chips_parquet.py tooling.

This dataset was 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.

Training Data

Dataset Composition

Training data includes nearly 2,000 sample chips across South America.

Data sources include:

• 909 samples from across 17 of South America’s ecoregions targeting areas of poor baseline model performance and agricultural practices unique to South America (represented as the orange polygons on the map above, from 2020-2024. Each field is labeled with its year).

• ~500 samples in Paraguay based on data from Mundo Agropecuario (the green polygons on the map above).

• 400 samples from Mato Grosso, sampled from challenging contexts with active learning query strategies based on output of the first Trazo model (the blue polygons on the map above).

Example Annotations

Agro-Ecological Diversity

A core goal of the Trazo dataset is the representation of a diversity of agricultural practices. This figure shows examples of field boundaries across South America ecoregions; the Chaco, Chiquitania, Araucaria, Amazon, Cerrado, Caatinga, and others. You can tour and view the field boundary annotations.

Data Format

All annotations are formatted for direct compatibility with the FTW training pipeline:

• 256×256 pixel Sentinel-2 image pairs of RGB + NIR for two time windows – WindowA which is typically the planting season and WindowB which is typically the harvest season – as .tif files
• 3-class label mask (field interior, field boundary, background)
• The majority are organized by ecoregion, others are organized by sampling type, test set, or provenance from a project partner

The dataset is intended to be used in conjunction with the FTW benchmark dataset by appending South America chips to an existing FTW training split. This dataset and each of its directories can be used as a standalone South America fine-tuning set. Each directory of annotated fields includes an 'annotation_info.txt' which lists the year and annotation authors. It also indicates if the boundaries were annotated against Planet Labs imagery. Those that were not were annotated using Sentinel-2 basemaps alone. Those that were annotated with monthly Planet Labs basemaps are appropriate for field boundary detection with PlanetScope SR (3m).

All field boundary annotations completely cover all fields within their chips. When regenerating window_a and/or window_b, we recommend searching for new imagery within the extent of the existing tif files to prevent misalignment between the field annotations and the chip extent. Regenerating grids over these annotations will require these samples to be presence-only due to this misalignment. We therefore suggest using the bounding box of existing TIFs to pull new harvest and planting season imagery.

Dataset Directory Structure

Why Field Boundaries Matter

Agricultural field boundaries are a key resource for building sustainable supply chains, as boundaries link agricultural commodities directly to their sites of production. When combined with earth observation data that tracks land-use change, this plot-level information makes it possible to assess whether production is linked to recent deforestation. These insights strengthen monitoring and risk assessment and support compliance with sustainability commitments and emerging regulations, including the European Union Deforestation Regulation.

Acknowledgments

This dataset was produced in collaboration with WRI and Arizona State University, with support from the Land and Carbon Lab. This work was funded through a Walmart Foundation grant.

Citation, License and Contact

Creative Commons Attribution 4.0 International (CC BY 4.0)

Free to use, share, and adapt for any purpose, including commercial applications such as supply chain traceability, provided that Trazo / WRI is properly cited.

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

Corresponding author: Tristan Grupp - tristan.grupp@wri.org

Kerner et al. (2025) — Fields of The World: A Machine Learning Benchmark Dataset For Global Agricultural Field Boundary Segmentation AAAI 2025, vol. 39(27), pp. 28151–28159. DOI: 10.1609/aaai.v39i27.35034

Corresponding author: Hannah Kerner - hkerner@asu.edu