The FTW benchmark data includes all the chips and associated imagery to train field boundary models. This repo is starting as a demonstration of how to organize it, and will evolve to be alpha and beta releases of the 2.0 dataset. This catalog is a 'git-backed' Portolan catalog, backed by https://github.com/fieldsoftheworld/benchmark-data-catalog
Benchmark chips for Luxembourg cut from the harmonized field boundary collection for Luxembourg (edition 2026).
![]()
Open this collection in the data browser, read the agent guide, or query items.parquet directly.
This collection holds 775 chips and 87,997 field polygons. Each chip is a square STAC item carrying the label masks rasterized from the field boundaries that fall inside it.
Chips per benchmark split:
How much of a chip is mapped field, as a distribution over the chips (field_coverage_pct):
Every chip carries these label rasters as STAC assets: instance_mask, semantic_2class_mask, semantic_3class_mask, decode_boundary_mask, decode_distance_mask.
680 chips have Sentinel-2 scenes selected for them, one in the planting window and one in the harvest window of the crop calendar. Planting imagery was acquired between 2025-03-22T10:47:28.853000Z and 2025-05-11T10:47:38.751000Z, averaging 0.1% cloud cover (worst 1.9%). Harvest imagery was acquired between 2025-07-25T10:47:07.706000Z and 2025-09-30T10:47:21.386000Z, averaging 0.2% cloud cover (worst 2.0%).
Ready-made map styles ship with the collection:
block3x3 strategy, random seed 42id, which carries the grid cell each chip was cut from.items.parquet mirrors every chip item, so the whole collection can be queried without walking the catalog:
See AGENTS.md for the schema, field notes and more queries.
Fields of the World 1.0 published luxembourg with 808 chips (643 train / 81 val / 84 test) cut from 29018 parcels declared for 2022, under CC0-1.0. This collection has 775 chips (601/79/95) cut from 87997 fields declared for 2026.