FTW started as a benchmark dataset and has grown into a full stack for agricultural field boundary detection — datasets, pretrained models, a CLI, a QGIS plugin, a browser-based inference app, and an open API. Everything lives in the fieldsoftheworld GitHub organization.

The FTW Benchmark Dataset is a global collection of labeled agricultural field boundaries built for open benchmarking. It harmonizes a range of open sources into 1.6 million parcels and 70,000+ labeled samples spanning 24 countries across 4 continents.

Explore and download the dataset from Source Cooperative.

What's included in the FTW Benchmark Dataset?

• Publicly available, labeled field boundaries from around the world
• Uniform data formatting and metadata schema
• A maintained subset of training/validation/test data
• Tools to evaluate model performance on shared benchmarks
• Contribution guidelines for submitting new datasets

Reference segmentation models trained on the FTW Benchmark Dataset. They serve as both starting points for fine-tuning and as performance baselines for new methods. Four pretrained checkpoints are available — two-class and three-class variants, each in CCBY (permissive-license data only) and FULL (all FTW data) flavors.

Open-source tools that cover the full workflow — from dataset conversion, to inference on new imagery, to visualization in QGIS.

Browse the global FTW dataset — 3.17B field polygons across 241 countries — directly in your browser. Pan the map, zoom into any region, and download field boundaries as GeoParquet or GeoJSON for your own workflows. File an issue if you spot something off or want to request a feature.

You can also run inference on demand — draw an area of interest, pick a model, and get vector field boundaries back without writing any code. Powered by the open-source ftw-inference-api, and under active development.

Large-scale outputs of model-inferred field boundaries, such as country-wide field maps. No full-country release has shipped yet, but work is underway in Kenya, Vietnam, Laos, Colombia, and other regions.

These products will be versioned and openly published to support external evaluation, planning, and downstream applications. As model quality improves we aim to scale toward a complete global field boundary map.

Field Boundaries for Agriculture (fiboa)

The fiboa project publishes field boundary datasets in a unified open specification, curated from government open data and hand-labeling campaigns. Fiboa datasets live at source.coop/fiboa, and the data in the Source Cooperative kerner-lab organization is already in fiboa format.

Going forward, FTW benchmark datasets may draw from fiboa sources before subsetting and harmonization — fiboa acts as the upstream collection layer, FTW as the curated training and benchmarking layer on top.