FTW Global — Field Boundary Predictions (GeoParquet)
Global agricultural field-boundary polygons predicted by the PRUE model (ftw-baselines), published as cloud-native GeoParquet partitioned by country with per-country PMTiles for web visualization. Part of Fields of the World.
What's here
- GeoParquet, partitioned by country under
alpha/results-by-admin-conf/admin:country_code=<CC>/. Query the whole dataset with the Portolan glob asset:alpha/results-by-admin-conf/admin:country_code=*/*.parquet. - One STAC item per parquet partition (id = the parquet filename). Large countries are split by admin subdivision and grouped under a per-country sub-catalog.
- Per-country PMTiles alongside each parquet (
<name>.pmtiles). - Two collection-level PMTiles for the global web view:
ftw-global-fields-2025.pmtiles— default display (2025).2024_with_confidence.pmtiles— 2024.
- MapLibre styles in
styles/that color fields by confidence (red→green, 0–100) with the recommended>= 69reliability filter and a legend.
How this was made — the model & training data
These boundaries are predictions from the PRUE model (Muhawenayo et al. 2026, PRUE: A Practical Recipe for Field Boundary Segmentation at Scale) — a U-Net semantic-segmentation model with composite loss functions and targeted data augmentations. In a benchmark of 18 models it reached 76% IoU and 47% object-F1 on the FTW benchmark (a +6% / +9% gain over the prior baseline), outperforming instance-segmentation and geospatial-foundation-model approaches, and is more robust to changes in illumination, spatial scale, and geographic location.
PRUE was trained and evaluated on the Fields of the World (FTW) benchmark — get the data on Source Cooperative: https://source.coop/kerner-lab/fields-of-the-world (Kerner et al. 2025, AAAI, arXiv:2409.16252). The benchmark is 70,462 samples across 24 countries on four continents (Europe, Africa, Asia, South America), pairing multi-date multispectral Sentinel-2 imagery with instance- and semantic-segmentation field masks; models pre-trained on FTW generalize better (zero-shot and fine-tuned) to held-out countries.
To build this global layer, PRUE is run over global Sentinel-2 planting/harvest median
composites (the features collection / prediction Zarr), and the softmax outputs for
[non_field_background, field, field_boundaries] are thresholded at 0.5 and polygonized
into these vectors; the confidence raster summarizes per-cell reliability. The broader
FTW ecosystem and downstream recipes (crop-type mapping, forest-loss attribution) are in
the FTW field guide (Corley et al. 2026), which
reports median predicted field sizes of 0.06 ha (Rwanda) to 0.28 ha (Switzerland)
across five countries / 4.76M km².
The confidence column
Each polygon carries a confidence value on a 0–100 scale. It is derived,
not part of the raw model vector output:
- Sampled at each field's point-on-surface from the 500 m PRUE confidence
COG (the
predictions/confidencecollection). - Rescaled
raw / 0.578178 * 100and clamped to 100 — i.e.0.578178is treated as 100%, matching the FTW inference app legend (raw0.404→70,0.463→80,0.521→90,0.578→100). Cells with no data become null. - This centroid sample is the closest single value to the app's
confidence_mean(for the common case of fields smaller than a 500 m cell they are identical). - Recommended default filter:
confidence >= 69(raw 0.4). Confidence reflects cell-level model reliability, not individual-polygon geometric accuracy.
It is computed by the scripts in this repo:
add_confidence.py,
process_partition.sh,
run_rails.sh,
make_pmtiles.py.
Schema
Field definitions use the wording of the fiboa core specification
and the vecorel geometry-metrics
and administrative-division
extensions (also machine-readable in the collection's table:columns):
The per-country PMTiles split fields into two layers by prediction year — 2024 and
2025 — derived from determination:datetime.
Using the data
License
CC-BY-4.0. Produced by the Taylor Geospatial Institute and the Microsoft AI for Good Research Lab.