PlanetScope SR (3m) variant of the FTW Benchmark dataset
Paired PlanetScope SR scenes (two windows per AOI — early- and peak-season) co-registered with Fields of The World v2 field-boundary labels, across 25 countries.
usable_pair = True)ortho_analytic_4b_sr, 4 bands (B/G/R/NIR), 3 m GSD,
native UTM, uint16 (reflectance = DN / 10000)uint8 with NBITS=2; boundaries rasterized with all_touched=True
to match the FTW originals.Each tar is a WebDataset shard with five files per patch_id:
<pid>.polygons.parquet holds the original FTW vector field boundaries
reprojected to the patch's UTM grid and clipped to its bounds — the same
vector source the .label.tif raster is burned from, so you can score
polygon-level metrics against true geometry rather than connected
components of the mask. Columns: id, geometry, area_ha (true
planimetric area in hectares), plus any of crop_id / crop_name /
area / perimeter present in the source. Patches with no fields carry
an empty (0-row) GeoParquet, so every sample has the file.
Tars are uncompressed; the TIFFs inside are ZSTD-22. They stream as
WebDataset shards and also extract cleanly with tar -xf <country>.tar.
The index is GeoParquet 1.1 with a bbox covering struct and is
Hilbert-sorted into 14 row groups, so spatial queries from DuckDB /
duckdb-wasm can prune row groups by bbox without parsing WKB:
Identity / geometry:
Paths (relative to the tar / planet root):
Scene provenance, per window suffix _a / _b:
Per-patch UDM2 statistics (fraction of pixels in the patch), per window:
FTW season metadata:
Quality:
This dataset is released under CC-BY-NC-4.0 (non-commercial).
Imagery is © Planet Labs PBC. The included AOIs were exported under the
NICFI / research program — refer to those terms for redistribution.
FTW v2 polygons are CC-BY-4.0; see fieldsoftheworld/ftw-baselines for
source terms.