allenai/planteur-benchmark

PLANTEUR - Satellite Image Time Series Crop Type Segmentation Benchmark

A French-overseas-territories crop-type semantic + panoptic segmentation dataset, packaged in almost the exact same PASTIS file format so that existing PASTIS dataloaders work mostly unchanged with the Sentinel-2, Sentinel-1, and label arrays (Landsat-8, which PASTIS does not have, needs the loader in the GitHub repository). Code (data loaders for all sensors including Landsat, collate function, demo): https://github.com/allenai/planteur-benchmark

from huggingface_hub import snapshot_download
root = snapshot_download("allenai/planteur-benchmark", repo_type="dataset")  # ~258 GB: ~133 GB arrays and labels, ~21 GB monthly mosaics, ~104 GB embeddings

Territories: Guadeloupe, Martinique, Guyane, Réunion, Mayotte, France's five overseas departments and regions (départements et régions d'outre-mer, DROM), and a small number of patches from Saint-Martin, an overseas collectivity (collectivité d'outre-mer, COM) rather than a DROM.

Contents (2,750 patches, 128x128 px @ 10 m)

Monthly mosaics, embeddings and the experiment protocol files have their own sections below.

Imagery (one time series per patch; T varies by patch and sensor)

• DATA_S2/
  • S2_<id>.npy : Sentinel-2 L2A, (T_s2, 10, 128, 128) uint16.
  • Bands: B02, B03, B04, B05, B06, B07, B08, B8A, B11, B12.
• DATA_S1/
  • S1_<id>.npy : Sentinel-1 RTC, (T_s1, 2, 128, 128) float32, VV and VH.
  • Radiometrically terrain-corrected linear backscatter (not the 10*log10 dB transform of PASTIS-R). 0 marks a pixel with no valid observation on that date (0.3% of pixels, mostly swath edges).
• DATA_LS/
  • LS_<id>.npy : Landsat-8 Collection 2 Level-1, (T_ls, 10, 64, 64) uint16 (scaled digital numbers, not surface reflectance).
  • Bands: B1, B2, B3, B4, B5, B6, B7, B9, B10, B11. Same patch extent as the arrays above, at half their linear resolution.
• DATA_LS_PAN/
  • LSPAN_<id>.npy : Landsat panchromatic B8, (T_ls, 1, 128, 128) uint16, on the same 128x128 grid as S2/S1. T_ls matches DATA_LS for every patch.

Labels

• ANNOTATIONS/
  • TARGET_<id>.npy : (1, 128, 128) uint8 semantic labels, 8 classes + 255 void (see Classes).
• ANNOTATIONS_GEOTIFF/
  • TARGET_<id>.tif : the same labels as georeferenced GeoTIFFs (uint8, 128x128, nodata=255) carrying each patch's CRS and affine transform; pixel values are identical to the .npy. See Georeferencing.
• INSTANCE_ANNOTATIONS/ : panoptic targets, same convention and filenames as PASTIS.
  • INSTANCES_<id>.npy : per-pixel parcel/instance id.
  • HEATMAP_<id>.npy : object centerness heatmap.
  • ZONES_<id>.npy : pixel -> instance zones.
  • Instances are the parcels of the 7 crop classes (47,256 parcels; every parcel has a single class); void fields and Background have no instance. Instance IDs are global, so a parcel crossing a patch border keeps its ID in each patch (58,489 parcel–patch instances).

Metadata and splits

• metadata.geojson : per-patch ID_PATCH, dates-S2, dates-S1, dates-LS ({step: YYYYMMDD}), orbits-S1 ({step: "A" | "D" | "AD"}, the Sentinel-1 orbit direction of each dates-S1 step; "AD" = same-day ascending and descending passes averaged), Fold, split (train/val/test), Island, geometry (real patch polygon, EPSG:4326). Same-day acquisitions are mosaicked to a single observation, so dates are unique and chronological per patch and exactly match the array length T (PASTIS's one-observation-per-date rule).
• splits/
  • 443.json : the default train/val/test split as explicit ID lists (same assignment as Fold).
  • loio.json : leave-one-island-out split definitions (see Splits).
• NORM_S2_patch.json, NORM_S1_patch.json, NORM_LS_patch.json, NORM_LS_PAN_patch.json : per-band train mean/std, keyed Fold_1..Fold_5 (each holds the same train statistics) as the loader expects. S1 and Landsat statistics exclude nodata (0) pixels.

Documentation and preview

• documentation/
  • class_map.json : 26-class source id -> class id.
  • label_names.json : class names.
  • colormap.txt : visualization colors for classes 0-7.
• sample/
  • sample.parquet : a 25-patch preview (random, seed 0) for the dataset viewer: S2 true-colour and colourised label PNGs plus per-patch metadata (island, split, CRS, bounds, sequence lengths, classes present). Previews only; train from the arrays above.

Monthly mosaics (as used in the experiments)

The monthly inputs used in some of the paper's linear-probe and fine-tuning experiments and for the U-TAE / TSViT baselines, exported exactly as the models saw them. Month m (0-11) covers the 30 days from 2019-01-01 + 30·m days (so the year ends 2019-12-27). Each month is the lowest-cloud-cover scene mosaic of that period (Sentinel-2 and Landsat ranked by scene cloud cover; Sentinel-1 not ranked), so it is not reproducible from the DATA_* arrays alone. Lower-resolution bands are upsampled to the output grid with nearest neighbour, as the experiment loader did. A month with no scene is all zeros: Sentinel-2 and Sentinel-1 have all 12 months for every patch; Landsat (16-day revisit) is missing 1-3 months in 1,627 patches (3,117 months in total).

• DATA_S2_MONTHLY/S2_<id>.npy : (12, 12, 128, 128) uint16. Bands: B02,B03,B04,B05,B06,B07,B08,B8A,B11,B12 (the DATA_S2 order), then B01,B09.
• DATA_S1_MONTHLY/S1_<id>.npy : (12, 2, 128, 128) float32 (VV,VH), linear backscatter; 0 = no valid observation, as in DATA_S1.
• DATA_LS_MONTHLY/LS_<id>.npy : (12, 10, 64, 64) uint16, same bands and scaling as DATA_LS.
• DATA_LS_PAN_MONTHLY/LSPAN_<id>.npy : (12, 1, 128, 128) uint16, B8.
• MONTHLY/months.json : the 12 month periods, and per patch and sensor which months have a scene (present[<id>]["S2"|"S1"|"LS"|"LS_PAN"], 12 booleans).

The 10 m Sentinel-2 bands of a monthly mosaic equal one DATA_S2 observation from that month where the patch lies inside a single Sentinel-2 tile; near tile edges a month can combine two scenes. The 20 m and 60 m bands differ from DATA_S2, which stores them resampled differently.

Experiment protocol files

Small files that pin down the training sets of the paper's protocols, so they can be reproduced exactly:

• splits/fewshot_pixels.json : the few-shot training pixels, {"seed": 0, "X": {X: {patch id: [[row, col, class], ...]}}} for X = 10, 25, 100, 1000 labelled pixels per class (exactly X per class, all from train patches; rows/cols index the 128x128 grid). Train on these pixels only (everything else void); val and test keep their full labels. The same PLANTEUR pixels are used in the PASTIS-transfer runs; PASTIS itself is not redistributed here.
• ANNOTATIONS_BALANCED/TARGET_<id>.npy : labels for the class-balanced protocol, same format as ANNOTATIONS. In train patches, a subset of Background pixels is set to 255 (void) so that the remaining Background equals the total crop pixels (6.22M each); val and test patches are identical to ANNOTATIONS.
• NORM_S2_MONTHLY_patch.json : per-band train mean/std of the monthly mosaics (the 10 DATA_S2 bands of DATA_S2_MONTHLY), in the same Fold_1..Fold_5 format as NORM_S2_patch.json; this is what normalized U-TAE and TSViT on the monthly inputs.

Precomputed embeddings

The embedding products the paper linear-probes, exported per patch on the same 128x128, 10 m grid as the labels (pixel [r, c] of an embedding aligns with label pixel [r, c]). All are for 2019.

• EMB_AEF/AEF_<id>.npy : (64, 128, 128) float32, AlphaEarth Foundations Satellite Embedding V1, 2019 annual (source); -1 = no data (no such pixels in this release).
• EMB_TESSERA_V1/TESSERA_<id>.npy : (128, 128, 128) float32, Tessera v1; no data is an all-zero vector (149 patches, 7 of them entirely: 291, 934, 989, 1015, 1038, 1058, 1980) or NaN (6 patches: 1394, 1395, 1812, 2262, 2273, 2638). Absent for 6 patches (1260-1265) where Tessera v1 has no coverage; see EMB_MISSING.json.
• EMB_TESSERA_V1_1/TESSERA_<id>.npy : (128, 128, 128) float32, Tessera v1.1; NaN = no data (4 patches: 1395, 1639, 2262, 2417).
• EMB_TESSERA_V2/TESSERA_<id>.npy : (128, 128, 128) float32, Tessera v2, generated with the released v2 inference code (source; no precomputed v2 product exists for 2019).
• EMB_OLMOEARTH_V1_3/OE13_<id>.npy : (128, 128, 128) float32, OlmoEarth v1.3, the 128-d projected read-out (the paper's 64-d read-out is the first 64 channels); patch size 2, latent patch size 1, 16 px windows, values int8 round-tripped exactly as the probes saw them.
• EMB_MISSING.json : {folder: [patch ids]} for embeddings with no source data.

Licenses: the AlphaEarth Foundations embeddings are CC-BY 4.0 — "The AlphaEarth Foundations Satellite Embedding dataset is produced by Google and Google DeepMind." Tessera embeddings are CC0.

Georeferencing

ANNOTATIONS_GEOTIFF/ gives the exact georeference of every label raster directly, rather than leaving it to be reconstructed — a common difficulty with PASTIS. The imagery arrays are on the same grid, so a patch's label GeoTIFF also georeferences its DATA_* arrays.

The patches span five UTM zones, so there is no single CRS for the dataset:

CRS	Patches	Territories
EPSG:32620	1291	Guadeloupe (735), Martinique (550), Saint-Martin (6)
EPSG:32621	59	Guyane
EPSG:32622	447	Guyane
EPSG:32738	140	Mayotte
EPSG:32740	813	Réunion

ANNOTATIONS_GEOTIFF/_crs_summary.json records these counts. The geometry field in metadata.geojson gives the same footprints in EPSG:4326.

import rasterio
with rasterio.open("ANNOTATIONS_GEOTIFF/TARGET_0.tif") as src:
    labels = src.read(1)          # (128,128) uint8, 255 = void
    crs, transform = src.crs, src.transform

Classes (8 + void)

0   Background
1   Meadow
2   Fruits, vegetables, flowers
3   Orchard
4   Sugarcane
5   Banana
6   Pineapple
7   Tropical tuber
255 Void  <-- masked out of loss and metrics (nodata_value=255)

Agricultural classes outside the 8-class taxonomy (Corn, Grapevine, Beet, Potatoes, Leguminous fodder, Soybeans, Vanilla, Ylang-ylang, Coffee/Cacao, plus nine temperate classes that never occur in the DROM windows) are mapped to 255 (void) rather than folded into Background: Background(0) (land outside the RPG plus RPG parcels whose crop code has no source class) stays a real, trained class, and void marks data that must not contribute to loss or metrics. This follows the PASTIS convention. See documentation/class_map.json for the full 26-class -> class-id mapping.

Void covers 92,145 px (0.20% of all pixels) in 385 patches; void fields are not parcel instances.

Metrics. Exclude void (255) from both loss and metrics. For semantic segmentation, report mean IoU over all 8 classes: a class absent from the labels of the evaluated patches (e.g. a crop that does not occur on a held-out island) counts as IoU 0 in the mean.

Representing void in a trainer. Void is 255 on disk, which metric code can ignore directly, since 255 never appears in the class list. But a trainer that allocates one output channel per class cannot represent 255 — e.g. UTAE-PaPs train_panoptic.py, whose --out_conv "[32, N]" sizes the head by num_classes. For those, remap 255 -> 8 and use num_classes 9, void_label 8. The eight trained class ids are unchanged either way.

Splits

Default (spatial-block, territory-stratified) — Fold in metadata.geojson: 1=train (1896), 4=val (411), 5=test (443). Whole 0.1° (~11 km) blocks go to one split. The 70/15/15 target is applied per territory, but blocks are indivisible, so the realised ratios vary, most for Mayotte (10 blocks):

Territory	train	val	test
Guadeloupe	526 (72%)	140 (19%)	69 (9%)
Martinique	426 (77%)	66 (12%)	58 (11%)
Guyane	359 (71%)	67 (13%)	80 (16%)
Réunion	524 (64%)	117 (14%)	172 (21%)
Mayotte	55 (39%)	21 (15%)	64 (46%)
Saint-Martin	6	0	0

Parcels crossing a block edge can still reach two splits: 592 of the 47,256 parcels have pixels in patches of more than one split (321 train/test, 240 train/val, 30 val/test, 1 in all three).

Leave-one-island-out (LOIO) — splits/loio.json. For each held-out territory, test is all of its patches; train/val split the remaining territories' patches (about 88/12), as used in the paper. Unlike the default split, this train/val split is not block-aligned: train and val share 0.1° blocks, so val scores are optimistic. test never shares a block with train or val:

held-out (test)	train	val	test
Guadeloupe	1750	265	735
Martinique	1950	250	550
Guyane	1960	284	506
Réunion	1720	217	813
Mayotte	2296	314	140

Saint-Martin (6 patches) is labelled in metadata.geojson via Island but is never held out — it stays in train for every fold. Mayotte is the smallest fold (140 patches); treat its per-island numbers with corresponding caution.

import json
loio = json.load(open("splits/loio.json"))["folds"]["Réunion"]
train_ids, val_ids, test_ids = loio["train"], loio["val"], loio["test"]

Patch counts by territory

Territory	Patches	%
Réunion	813	29.6%
Guadeloupe	735	26.7%
Martinique	550	20.0%
Guyane	506	18.4%
Mayotte	140	5.1%
Saint-Martin	6	0.2%

Notes

• Temporal coverage: the retrieval window is 2018-09-01 to 2019-12-31 for all three sensors, so both the September-aligned agricultural year (2018-09 to 2019-08) and calendar year 2019 are fully contained. Every patch has observations through December 2019 (median final observation 2019-12-28 for Sentinel-2, 2019-12-29 for Sentinel-1 and 2019-12-16 for Landsat; the earliest, 2019-12-06, is Landsat). The first usable observation varies by patch and sensor, because it depends on actual acquisition and cloud availability rather than a fixed start: Sentinel-2 median 2018-09-04, but 10% of patches begin on or after 2018-12-03 (latest 2019-09-21); Sentinel-1 median 2018-09-04, 10% after 2018-11-19 (latest 2019-03-04); Landsat is effectively uniform (every patch starts by 2018-09-16). This is why the sequence length T varies per patch, as listed below.
• Time series are per-acquisition observations (not monthly composites): Sentinel-2 L2A and Sentinel-1 RTC from Microsoft Planetary Computer, Landsat-8 Collection 2 Level-1 from the USGS bucket on AWS. After same-day mosaicking the sequence length T varies per patch: S2 40–150 (median 87), S1 40–150 (median 123), Landsat 28–61 (median 30).
• Load with PLANTEUR_Dataset(folder=root, sats=['S2']) from code/dataloader.py (sats also takes 'S1', 'LS', 'LS_PAN' and 'S2_MONTHLY', the 10 DATA_S2 bands of the monthly mosaics); target='semantic' for crop-type segmentation, target='instance' for panoptic; annotations='ANNOTATIONS_BALANCED' for the class-balanced training labels. Select default splits with folds=[1] (train), [4] (val), [5] (test).
• The paper's all-observation input (AnySat) is the first 96 Sentinel-2 observations of 2019 per patch, taken from DATA_S2 with dates-S2.
• The Sentinel-2, Sentinel-1 and label arrays also load unchanged through the upstream pastis-benchmark/code/dataloader.py, for both semantic and instance targets.
• PASTIS-R splits Sentinel-1 into ascending and descending series (DATA_S1A, DATA_S1D). PLANTEUR has one DATA_S1 series per patch with both orbit directions in date order (2,746 of 2,750 patches have both; same-day ascending and descending passes, about 3% of dates, are averaged into one observation); orbits-S1 gives each observation's orbit, so an ascending or descending series can be cut out of DATA_S1. Load it as sats=["S2", "S1"]. The PASTIS-R pixel-set loader also needs parcel-level files that PLANTEUR does not ship; use PLANTEUR_Dataset_PixelSet from the GitHub repository.

Citation

If you use PLANTEUR, please cite:

@misc{planteur2026,
  title        = {PLANTEUR: A Crop-Type Segmentation Benchmark for Tropical Regions},
  author       = {Wolters, Piper and Tseng, Gabriel and Zhang, Yawen and Redmon, Joseph and
                  Bastani, Favyen and Sablon, Hadrien and Shah, Ando},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/allenai/planteur-benchmark}}
}

This dataset follows the format of PASTIS; if you use it, please also cite:

@inproceedings{garnot2021panoptic,
  title     = {Panoptic Segmentation of Satellite Image Time Series with Convolutional Temporal Attention Networks},
  author    = {Sainte Fare Garnot, Vivien and Landrieu, Loic},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year      = {2021}
}

License

This benchmark dataset is licensed under ODC-BY. It is intended for research and educational use in accordance with Ai2's Responsible Use Guidelines.

Attribution

The data is processed from satellite imagery and labels provided by the following sources:

• Sentinel-1 Synthetic Aperture Radar (SAR) - Radiometrically Terrain Corrected (RTC) dataset provided by the European Space Agency and accessed from Microsoft Planetary Computer at https://planetarycomputer.microsoft.com/dataset/sentinel-1-rtc
• Sentinel-2 Level 2-A dataset provided by the European Space Agency and accessed from Microsoft Planetary Computer at https://planetarycomputer.microsoft.com/dataset/sentinel-2-l2a
• USGS Landsat - Earth Resources Observation and Science (EROS) Center. (2020). Landsat 8-9 Operational Land Imager / Thermal Infrared Sensor Level-1, Collection 2 dataset. U.S. Geological Survey. https://doi.org/10.5066/P975CC9B. Accessed from https://registry.opendata.aws/usgs-landsat
• Le Registre Parcellaire Graphique (RPG) dataset at https://cartes.gouv.fr/rechercher-une-donnee/dataset/IGNF_RPG?redirected_from=geoservices.ign.fr

The file format is adapted from PASTIS (V. Sainte Fare Garnot & L. Landrieu), MIT-licensed.