Fields of The World · PlanetScope companion

Fields of the Planet

Field boundary mapping beyond 10 meters

Isaac Corley, Caleb Robinson, Jennifer Marcus, Hannah Kerner

Taylor Geospatial · 2026

Each row: a 10 m Sentinel-2 patch, the matched 3 m PlanetScope image, and the ground-truth FTW field boundaries.
Fields of the Planet pairs each Fields of The World Sentinel-2 patch with a co-registered 3 m PlanetScope image and the original field-boundary labels. Left to right: the FTW Sentinel-2 patch at 10 m, the matched PlanetScope surface-reflectance image at 3 m (4-band; RGB shown), and the ground-truth FTW vector polygons.

Abstract

A controlled test of resolution for field delineation

Field-boundary maps support crop monitoring, irrigation planning, and yield estimation, but many smallholder parcels span only a few 10 m Sentinel-2 pixels. Roughly 510 million farms under 2 ha account for about 84% of the world's farms by count, and a 10 m grid resolves their boundaries worst.

Fields of the Planet (FTP) is a 3 m PlanetScope companion to Fields of The World (FTW). It keeps the same polygons, seasonal windows, and train/test splits, and pairs 133,168 patch-window targets across 24 countries with co-registered PlanetScope surface reflectance. This isolates the effect of spatial resolution, holding the labels and protocol fixed and changing only the imagery.

We evaluate delineation as parcel recovery. Predictions are vectorized and scored as polygon objects, using panoptic quality (PQ), object F1 over the COCO IoU grid, and meter-scale matched-boundary error. Pixel overlap alone understates the benefit of higher resolution, so we do not rely on it.

Motivation

The 10 m grid loses small fields before training begins

The limit appears in the labels themselves. Rasterizing FTW's own polygons onto a 10 m grid merges adjacent parcels: only 7.5% of sub-0.5 ha fields remain separable as distinct polygons at 10 m, versus 88.3% at 3 m. A model trained on 10 m targets cannot recover fields the grid has already merged.

Qualitative comparison showing PlanetScope resolving dense smallholder parcels that Sentinel-2 merges.
PlanetScope resolves dense smallholder parcels that Sentinel-2 merges. Held-out examples show the 3 m image, the 10 m image, ground truth, and model outputs as mask / vectorized-polygon pairs.

Results

3 m imagery improves recovery and boundary localization

We train matched U-Net models with an EfficientNet-B3 encoder on each sensor. Replacing 10 m Sentinel-2 with 3 m PlanetScope improves every polygon-level metric, with the largest relative gain on the smallholder fields a 10 m grid cannot resolve.

Macro-averaged over the ten dense-label held-out countries. Better value in bold.
Metric Sentinel-2 (10 m) PlanetScope (3 m)
Panoptic quality (PQ)21.035.5
PQ on fields < 0.5 ha5.815.7
Object F1 at IoU 0.528.946.2
Matched-boundary error (m)18.67.4

Matched boundaries are also localized more than twice as accurately. A finer output grid alone (upsampled Sentinel-2) does not close the gap, because it cannot recover edges the input never resolved.

Held-out patches with the largest FTW to FTP polygon-PQ improvement.
Where 3 m helps most. Held-out patches with the largest FTW→FTP polygon-PQ improvement. In dense smallholder scenes the Sentinel-2 model merges neighboring parcels or leaves gaps, while the PlanetScope model recovers more of the parcel structure.

Dataset

What ships in the release

Coverage
66,584 patches · 133,168 image-window pairs · 24 countries / 25 labeled regions
Imagery
PlanetScope ortho_analytic_4b_sr — 4 bands (B/G/R/NIR), ~3 m GSD, native UTM, uint16
Windows
Two seasonal acquisitions per patch (planting and harvest)
Labels
3-class raster (background / interior / boundary) plus the original FTW vector polygons as GeoParquet
Format
One WebDataset tar per region (25 shards, ~94 GiB) and a GeoParquet index
License
CC-BY-NC 4.0 (inherits Planet's non-commercial terms); labels carry their per-region FTW licenses

Citation

BibTeX

@misc{corley2026fieldsplanetfieldboundary,
  title         = {Fields of the Planet: Field Boundary Mapping Beyond 10m},
  author        = {Isaac Corley and Caleb Robinson and Jennifer Marcus and Hannah Kerner},
  year          = {2026},
  eprint        = {2607.04449},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2607.04449}
}