taylor-geospatial/ftw-planet

Fields of the Planet (FTP)

Paired PlanetScope SR scenes, two seasonal windows per patch (planting and harvest), co-registered with Fields of The World field-boundary labels, across 24 countries and 25 labeled regions.

• 66,584 patches across 24 countries and 25 labeled regions, drawn from 70,484 labeled FTW patches
• 52,235 patches with both windows passing UDM2 usability (usable_pair = True)
• Imagery: PlanetScope ortho_analytic_4b_sr, 4 bands (B/G/R/NIR), 3 m GSD, native UTM, uint16 (reflectance = DN / 10000)
• Labels: 3 classes — 0 background, 1 field interior, 2 field boundary; uint8 with NBITS=2; boundaries rasterized with all_touched=True to match the FTW originals.

Results

Polygon-level results macro-averaged over the ten dense-label held-out countries dominated by smallholder fields (paper Table 1).

Method	Sensor	Backbone	PQ	SQ	RQ@.5	F1[.5:.95]	|ΔN|/N ↓	Bd. err mean (m) ↓	Bd. err p95 (m) ↓	Pixel IoU †	PQ small ‡	PQ med ‡	PQ large ‡
DelineateAnything *	PlanetScope	YOLO11x	9.5	73.3	12.7	7.0	0.75	13.7	37.8	51.1	1.7	7.1	16.3
DelineateAnything-S *	PlanetScope	YOLO11n	3.5	70.8	4.8	2.5	0.82	13.2	34.2	40.7	0.8	2.8	6.7
DelineateAnything v2 *	PlanetScope	YOLO11x	7.5	75.0	10.0	5.6	0.82	9.4	25.5	27.1	4.4	11.4	14.8
FTW-PRUE+	Sentinel-2	EfficientNet-B3	21.0	71.4	28.9	14.6	0.33	18.6	54.7	61.8	5.8	25.3	33.8
FTW-PRUE+	Sentinel-2	EfficientNet-B7	24.2	71.0	32.8	17.2	0.35	14.4	43.4	63.6	7.5	28.4	37.7
FTP-PRUE+	PlanetScope	EfficientNet-B3	35.5	75.7	46.2	27.1	0.33	7.4	22.8	68.8	15.7	39.2	52.0
FTP-PRUE+	PlanetScope	EfficientNet-B7	35.4	74.4	46.1	27.0	0.30	7.4	22.8	74.2	15.6	40.6	50.9

Bold marks the best value per column. * Released models evaluated without training on FTW or FTP, each at its best swept inference resolution and confidence setting. † Pixel IoU is not comparable across sensors due to differences in resolution. ‡ PQ for small (<0.5 ha), medium (0.5-2 ha), and large (>2 ha) ground-truth fields.

Layout

ftw-planet/
├── README.md
├── assets/                 # figures used in this card
├── index.parquet           # GeoParquet 1.1, one row per patch
└── dataset/
    ├── austria.tar
    ├── ...
    └── vietnam.tar         # 25 region shards, ~96 GiB total

Each tar is a WebDataset shard with five files per patch_id:

<pid>.window_a.tif        PlanetScope SR, planting window
<pid>.window_b.tif        PlanetScope SR, harvest window
<pid>.label.tif           3-class label
<pid>.polygons.parquet    true FTW field polygons, clipped to the patch
<pid>.json                metadata (mirrors the index row)

<pid>.polygons.parquet holds the original FTW vector field boundaries, reprojected to the patch's UTM grid and clipped to its bounds. This is the same vector source the .label.tif raster is burned from, so you can score polygon-level metrics against true geometry instead of 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.

Downloading

from huggingface_hub import hf_hub_download, snapshot_download

repo_id = "<this-repo-id>"  # e.g. "<org>/ftw-planet"

# one country shard
path = hf_hub_download(repo_id, "dataset/rwanda.tar", repo_type="dataset")

# the whole dataset
snapshot_download(repo_id, repo_type="dataset", local_dir="ftw-planet")

Reading the index

import geopandas as gpd
from huggingface_hub import hf_hub_download

repo_id = "<this-repo-id>"
idx = hf_hub_download(repo_id, "index.parquet", repo_type="dataset")
gdf = gpd.read_parquet(idx)
clean = gdf[gdf.usable_pair & (gdf.cloud_cover_a < 0.05) & (gdf.cloud_cover_b < 0.05)]

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:

INSTALL spatial; LOAD spatial; INSTALL httpfs; LOAD httpfs;

SELECT patch_id, country
FROM 'index.parquet'
WHERE bbox.xmin > -10 AND bbox.xmax < 25
  AND bbox.ymin > 35  AND bbox.ymax < 60
  AND usable_pair;

Index columns

Identity / geometry:

column	type	notes
patch_id	str	unique within country
country	str	one of 24 slugs
geometry	polygon	EPSG:4326 patch footprint
crs	str	native UTM CRS of the tifs (e.g. EPSG:32636)
bounds_4326	float[4]	[minx, miny, maxx, maxy] convenience field

Paths (relative to the tar / planet root):

column	example
image_a_path	rwanda/window_a/1592589.tif
image_b_path	rwanda/window_b/1592589.tif
label_path	rwanda/labels/1592589.tif

Scene provenance, per window suffix _a / _b:

column	notes
item_id_{a,b}	PlanetScope item ID
scene_date_{a,b}	UTC acquisition timestamp
cloud_cover_{a,b}	scene-level fraction in [0,1]
coverage_{a,b}	AOI coverage of the source scene
source_{a,b}	source product / pipeline tag

Per-patch UDM2 statistics (fraction of pixels in the patch), per window:

column	meaning
udm2_clear_{a,b}	clear sky
udm2_cloud_{a,b}	cloud
udm2_shadow_{a,b}	cloud shadow
udm2_light_haze_{a,b}	light haze
udm2_heavy_haze_{a,b}	heavy haze
udm2_snow_{a,b}	snow / ice
udm2_unusable_{a,b}	UDM2 unusable mask
udm2_confidence_mean_{a,b}	mean UDM2 confidence band
udm2_usable_flag_{a,b}	bool — derived per-patch quality

FTW season metadata:

column	notes
ftw_target_date_{a,b}	target acquisition date for each window
ftw_season_start	growing-season start (per FTW)
ftw_season_end	growing-season end (per FTW)

Quality:

column	type	notes
usable_pair	bool	both windows pass UDM2 usability — the primary training subset

Licensing

Fields of the Planet is made available under CC-BY-NC-4.0 (non-commercial), subject to the licensing terms of the underlying data sources.

PlanetScope imagery is © Planet Labs PBC and was obtained directly from the Planet archive under a research license for academic and nonprofit use. Use and redistribution of Planet imagery remain subject to the applicable Planet license terms. See the Planet Licensing Information Center for additional information.