🌾 AgriSR-PASTIS — Agricultural Super-Resolution & Segmentation Benchmark

AgriSR-PASTIS is a multi-modal satellite dataset for optical super-resolution and agricultural (agritech) segmentation. It pairs low-resolution Sentinel-2 time series (10 m/px) with co-registered very-high-resolution SPOT 6-7 imagery (1 m/px) over 2,433 agricultural patches in metropolitan France, together with dense panoptic crop-parcel annotations (18 crop types, 124,422 parcels) and aligned Sentinel-1 radar time series.

This makes the dataset suitable for two workflows out of the box:

1. Super-Resolution (SR) — learn 10 m ➜ 1 m mappings using DATA_S2 (LR input, multispectral, multi-temporal) and DATA_SPOT (HR RGB target), pixel-aligned per patch (128×128 ➜ 1280×1280).
2. Agritech segmentation — semantic, instance, and panoptic segmentation of crop parcels using ANNOTATIONS and INSTANCE_ANNOTATIONS, from single images, time series, or SR-enhanced inputs. Radar (DATA_S1A/DATA_S1D) enables multi-modal and cloud-robust variants.

This dataset is a repackaged distribution of PASTIS-HD (IGNF / PASTIS benchmark), curated for super-resolution and agritech segmentation research. Original credits, license, and citations are preserved below, as required by the Licence Ouverte / Open License (etalab-2.0).

📊 Dataset in numbers

🛰️ Sentinel-2 (LR)	🛰️ Sentinel-1 (radar)	🛰️ SPOT 6-7 (HR)	🗻 Annotations
2,433 time series	2 × 2,433 time series (asc. + desc.)	2,433 images	124,422 individual parcels
10 m / pixel	10 m / pixel	1 m / pixel (native 1.5 m)	~4,000 km² covered
128×128 px / image	128×128 px / image	1280×1280 px / image	over 2B labeled pixels
38–61 acquisitions / series	~70 acquisitions / series	one observation (2019)	18 crop types + background/void
10 spectral bands	2 polarizations (VV, VH) + VV/VH ratio	3 bands (RGB, 8-bit)	semantic + instance + panoptic

⚠️ The SPOT data are natively 1.5 m resolution, over-sampled to 1 m to align pixel-perfect with the Sentinel grids (10 m ➜ 1 m = clean ×10 SR factor).

📁 Repository structure

AgriSR-PASTIS/
├── DATA_S2/                  # Sentinel-2 time series, npy (T×10×128×128) — SR input / segmentation
├── DATA_SPOT/                # SPOT 6-7 RGB 1m GeoTIFFs (1280×1280) — SR target
├── DATA_S1A/                 # Sentinel-1 ascending time series, npy — optional radar modality
├── DATA_S1D/                 # Sentinel-1 descending time series, npy — optional radar modality
├── ANNOTATIONS/              # TARGET_*.npy semantic labels, ParcelIDs_*.npy instance ids
├── INSTANCE_ANNOTATIONS/     # HEATMAP_*, INSTANCES_*, ZONES_* for panoptic training
├── metadata.geojson          # patch index: geometry, fold, acquisition dates (2,433 entries)
├── NORM_S2_patch.json        # per-fold normalization stats (S2)
├── NORM_S1A_patch.json       # per-fold normalization stats (S1 ascending)
├── NORM_S1D_patch.json       # per-fold normalization stats (S1 descending)
├── train.csv / val.csv / test.csv            # ready-made patch splits
├── *_filtered_20.csv / *_filtered_5.csv      # lightweight subset splits (20 / 5 patches per split)
└── documentation/            # original PASTIS documentation (PDF)

Patch IDs are consistent across folders: e.g. S2_10007.npy ↔ SPOT6_RVB_1M00_2019_10007.tif ↔ TARGET_10007.npy, so LR/HR/label triplets can be joined by ID.

⚠️ The S1 and S2 folders contain slightly more files than there are labeled patches; metadata.geojson (2,433 entries) is the authoritative index. DATA_S1A = ascending orbit, DATA_S1D = descending orbit.

🚀 Quick start

import numpy as np
import rasterio

pid = 10007

# Super-resolution pair
lr_ts = np.load(f"DATA_S2/S2_{pid}.npy")            # (T, 10, 128, 128) 10m multispectral series
with rasterio.open(
    f"DATA_SPOT/PASTIS_SPOT6_RVB_1M00_2019/SPOT6_RVB_1M00_2019_{pid}.tif"
) as src:
    hr = src.read()                                  # (3, 1280, 1280) 1m RGB target

# Segmentation labels
sem = np.load(f"ANNOTATIONS/TARGET_{pid}.npy")       # semantic crop-type mask
ins = np.load(f"ANNOTATIONS/ParcelIDs_{pid}.npy")    # parcel instance ids

A ready-made PyTorch dataset class is available in the OmniSat repository.

🌱 Crop classes

18 crop-type classes (plus background = non-agricultural land, and void = parcels mostly outside the patch), derived from the French land parcel identification system (RPG). See documentation/pastis-documentation.pdf for the full nomenclature and class distribution.

🎯 Suggested benchmarks

• ×10 optical SR: S2 (single date or temporal fusion) ➜ SPOT RGB, e.g. RRDB/ESRGAN, SwinIR, diffusion SR, temporal-fusion SR.
• SR-assisted segmentation: train segmentation on SR outputs vs. native 10 m inputs and measure the gain on parcel boundaries.
• Multi-modal segmentation: optical + radar time-series fusion for cloud-robust crop mapping.
• Panoptic parcel delineation: instance-level parcel extraction at 1 m using the panoptic annotation set.

📜 License

Distributed under the Licence Ouverte / Open License 2.0 (etalab-2.0). You are free to reuse, modify, and redistribute, including commercially, provided the paternity/attribution below is acknowledged.

🙏 Credits & provenance

This dataset is a curated redistribution of PASTIS-HD by IGNF, which extends the original PASTIS benchmark (PASTIS ➜ PASTIS-R ➜ PASTIS-HD). All data were produced by the original authors and institutions:

• The Sentinel imagery used in PASTIS was retrieved from THEIA: "Value-added data processed by the CNES for the Theia data cluster using Copernicus data. The treatments use algorithms developed by Theia's Scientific Expertise Centres."
• The annotations stem from the French land parcel identification system (RPG) produced by IGN.
• The SPOT images are open data thanks to the Dataterra Dinamis initiative under the "Couverture France DINAMIS" program.

📖 Citations

If you use this dataset, please cite the original papers:

@article{garnot2021panoptic,
  title={Panoptic Segmentation of Satellite Image Time Series with Convolutional Temporal Attention Networks},
  author={Sainte Fare Garnot, Vivien and Landrieu, Loic},
  journal={ICCV},
  year={2021}
}

For the radar extension (PASTIS-R):

@article{garnot2021mmfusion,
  title    = {Multi-modal temporal attention models for crop mapping from satellite time series},
  journal  = {ISPRS Journal of Photogrammetry and Remote Sensing},
  year     = {2022},
  doi      = {https://doi.org/10.1016/j.isprsjprs.2022.03.012},
  author   = {Vivien {Sainte Fare Garnot} and Loic Landrieu and Nesrine Chehata},
}

For the VHR extension (PASTIS-HD):

@article{astruc2024omnisat,
  title={Omni{S}at: {S}elf-Supervised Modality Fusion for {E}arth Observation},
  author={Astruc, Guillaume and Gonthier, Nicolas and Mallet, Clement and Landrieu, Loic},
  journal={ECCV},
  year={2024}
}