PASTIS-HD / README.md
Khlaifiabilel's picture
fix task categories in model card
ff4c626 verified
|
Raw History Blame Contribute Delete
11.9 kB
---
language: [en]
license: other
size_categories:
- 100G<n<1T
task_categories:
- image-classification
- image-segmentation
tags:
- remote-sensing
- agricultural
- sentinel-2
- sentinel-1
- spot-6-7
- time-series
- crop-type-classification
- parcel-segmentation
- panoptic-segmentation
- nextav
pretty_name: PASTIS-HD
source_datasets:
- original
---
# PASTIS-HD
## Panoptic Agricultural Satellite TIme Series — High Definition
Optical time series, radar time series, and very high resolution (VHR) imagery for agricultural parcel segmentation.
**NextAV IP.** All rights reserved. © 2026 NextAV. All intellectual property in this dataset repository, including the organization, curation, alignment, annotations, normalization statistics, split definitions, and documentation, is owned exclusively by NextAV. This dataset is hosted under the NextAV organization: [https://huggingface.co/datasets/nextav/PASTIS-HD](https://huggingface.co/datasets/nextav/PASTIS-HD).
## Overview
PASTIS-HD is a benchmark dataset for panoptic and semantic segmentation of agricultural parcels from satellite time series over the French metropolitan territory. It contains **2,433 patches** of 128×128 pixels, each accompanied by panoptic annotations (instance index + semantic label per pixel), aligned multi-temporal Sentinel-2 optical, Sentinel-1 radar (ascending and descending), and SPOT 6–7 very high resolution imagery.
PASTIS-HD is an extension of [PASTIS](https://github.com/VSainteuf/pastis-benchmark) and [PASTIS-R](https://arxiv.org/abs/2112.07558v1), enriched with aligned SPOT 6–7 VHR images resampled to 1 m resolution.
**PASTIS-HD** is used to evaluate multi-modal fusion methods (optical time series, radar time series, and VHR images) for parcel-based classification, semantic segmentation, and panoptic segmentation.
This dataset is used in the paper **[MAESTRO: Masked AutoEncoders for Multimodal, Multitemporal, and Multispectral Earth Observation Data](https://huggingface.co/papers/2508.10894)**. Code for MAESTRO: [https://github.com/ignf/maestro](https://github.com/ignf/maestro)
---
## 📊 Dataset in numbers
| Component | Count / detail |
|-----------|----------------|
| Total patches | 2,433 |
| Total size | ~81 GB |
| Sentinel-2 time series | 2,468 files (`DATA_S2/`) |
| Sentinel-1 ascending (`DATA_S1A/`) | 2,468 files |
| Sentinel-1 descending (`DATA_S1D/`) | 2,468 files |
| SPOT 6–7 VHR images (`DATA_SPOT/`) | 2,434 TIFF files |
| Parcel-ID annotations (`ANNOTATIONS/`) | 4,866 files |
| Instance heatmap annotations (`INSTANCE_ANNOTATIONS/`) | 5,829 files |
| Individual parcels annotated | 124,422 |
| Area covered | ~4,000 km² |
| Total annotated pixels | >2 billion |
| Crop classes | 18 (+ background + void = 20 labels) |
| Sentinel-2 tiles | 4 (t30uxv, t31tfj, t31tfm, t32ulu) |
| Cross-validation folds | 5 |
### Splits
| Split | Patches | Filtered (min 5% cover) | Filtered (min 20% cover) |
|-------|---------|-------------------------|--------------------------|
| Train | 1,455 | 72 | 292 |
| Validation | 482 | 23 | 95 |
| Test | 496 | 27 | 100 |
| **Total** | **2,433** | **122** | **487** |
Split CSVs live at the repository root (`train.csv`, `val.csv`, `test.csv` and `*_filtered_5.csv`, `*_filtered_20.csv`). Fold assignments (1–5) and per-patch metadata are in `metadata.geojson`.
### Per-sensor statistics
**Sentinel-2 (optical, `DATA_S2/`)**
- Array shape per patch: `(dates, 10 bands, 128, 128)`, dtype `int16`
- 10 spectral bands, 10 m / pixel resolution
- 38–61 acquisitions per series (max 61 dates per patch)
- Acquisition dates: 2018-09-17 to 2019-10-27 (159 unique dates across the dataset)
**Sentinel-1 ascending (radar, `DATA_S1A/`)**
- Array shape per patch: `(dates, 3 channels, 128, 128)`, dtype `float16`
- 3 channels: VV, VH, VV/VH ratio
- ~70 acquisitions per series (max 71 dates per patch)
- Acquisition dates: 2018-10-01 to 2019-12-04 (207 unique dates)
- 10 m / pixel resolution
**Sentinel-1 descending (radar, `DATA_S1D/`)**
- Array shape per patch: `(dates, 3 channels, 128, 128)`, dtype `float16`
- 3 channels: VV, VH, VV/VH ratio
- ~70 acquisitions per series (max 70 dates per patch)
- Acquisition dates: 2018-10-02 to 2019-12-02 (209 unique dates)
- 10 m / pixel resolution
**SPOT 6–7 VHR (`DATA_SPOT/`)**
- One TIFF image per patch, 1280×1280 pixels
- 3 spectral bands (RGB), 8-bit
- Natively 1.5 m resolution, oversampled to 1 m to align pixel-perfect with Sentinel data
### Annotation statistics
- Parcel count per patch: min 3, max 597, mean ≈ 62.64
- Parcel cover fraction per patch: min 0.0303, max 0.9883, mean ≈ 0.617
- Annotations stored as `(128, 128)` `int32` Parcel-ID arrays and `(128, 128)` `float64` instance heatmaps
### Per-patch normalization
Per-patch mean and standard deviation for each modality are provided in:
- `NORM_S2_patch.json`
- `NORM_S1A_patch.json`
- `NORM_S1D_patch.json`
---
## 🗂️ Dataset components
| Folder / file | Description |
|---------------|-------------|
| `DATA_S2/` | Sentinel-2 L2A optical time series, one `(T, 10, 128, 128)` `int16` `.npy` per patch (`S2_<ID>.npy`) |
| `DATA_S1A/` | Sentinel-1 GRD ascending-orbit time series, one `(T, 3, 128, 128)` `float16` `.npy` per patch (`S1A_<ID>.npy`). Channels: VV, VH, VV/VH |
| `DATA_S1D/` | Sentinel-1 GRD descending-orbit time series, one `(T, 3, 128, 128)` `float16` `.npy` per patch (`S1D_<ID>.npy`). Channels: VV, VH, VV/VH |
| `DATA_SPOT/PASTIS_SPOT6_RVB_1M00_2019/` | SPOT 6–7 VHR orthoimages, one 1280×1280 3-band 8-bit `.tif` per patch |
| `ANNOTATIONS/` | Parcel-ID segmentation maps, one `(128, 128)` `int32` `.npy` per patch (`ParcelIDs_<ID>.npy`) |
| `INSTANCE_ANNOTATIONS/` | Instance heatmaps, one `(128, 128)` `float64` `.npy` per patch (`HEATMAP_<ID>.npy`) |
| `metadata.geojson` | Patch index (2,433 entries): `Fold`, `ID_PATCH`, `N_Parcel`, `Parcel_Cover`, `TILE`, acquisition dates per sensor |
| `NORM_*_patch.json` | Per-patch normalization statistics (mean/std) per modality |
| `train.csv`, `val.csv`, `test.csv` | Standard split patch IDs |
| `*_filtered_5.csv`, `*_filtered_20.csv` | Splits filtered by minimum parcel cover (5% / 20%) |
| `documentation/label_names.json` | Mapping of label id → crop name (20 classes) |
| `documentation/colormap.txt` | RGB colormap for the 20 classes |
| `documentation/pastis-documentation.pdf` | Full dataset documentation |
> **Folder naming remark:** `DATA_S1A` contains Sentinel-1 **ascending** images; `DATA_S1D` contains Sentinel-1 **descending** images. `DATA_S2`, `DATA_S1A`, and `DATA_S1D` contain more than 2,433 files because some patches are not labelled and not used for training. The relevant information is in `metadata.geojson` (2,433 entries), which is used as the index by the dataloader.
---
## 🏷️ Classes (label id → name)
| Id | Class | Id | Class |
|----|-------|----|-------|
| 0 | Background | 10 | Winter triticale |
| 1 | Meadow | 11 | Winter durum wheat |
| 2 | Soft winter wheat | 12 | Fruits, vegetables, flowers |
| 3 | Corn | 13 | Potatoes |
| 4 | Winter barley | 14 | Leguminous fodder |
| 5 | Winter rapeseed | 15 | Soybeans |
| 6 | Spring barley | 16 | Orchard |
| 7 | Sunflower | 17 | Mixed cereal |
| 8 | Grapevine | 18 | Sorghum |
| 9 | Beet | 19 | Void label |
Background = non-agricultural land. Void label = parcels mostly outside their patch.
Full colormap: `documentation/colormap.txt`.
---
## 📥 Data loading
Time series lengths are variable across patches. Sentinel-2 and Sentinel-1 time series are stored as NumPy arrays; SPOT images as TIFF; annotations as NumPy arrays.
```python
import numpy as np
s2 = np.load("DATA_S2/S2_10000.npy") # (43, 10, 128, 128) int16
s1a = np.load("DATA_S1A/S1A_10000.npy") # (65, 3, 128, 128) float16
s1d = np.load("DATA_S1D/S1D_10000.npy") # (70, 3, 128, 128) float16
lab = np.load("ANNOTATIONS/ParcelIDs_10000.npy") # (128, 128) int32
```
A PyTorch dataset class is available in the [OmniSat repository](https://github.com/gastruc/OmniSat/blob/main/src/data/Pastis.py) and can be used directly to load PASTIS-HD.
---
## 🏛️ Organization alignment
This repository is maintained by **NextAV** and follows the [NextAV Handbook](https://huggingface.co/nextav/nextav-handbook) for:
- Engineering standards and development workflow
- Data handling, security, and compliance policies
- AI/ML model governance and responsible AI
- Documentation and code review standards
All intellectual property in this repository is owned by NextAV.
---
## 🔖 Source data provenance
The underlying source imagery and reference annotations originate from public data providers:
- **Sentinel imagery** retrieved from [THEIA](www.theia.land.fr): "Value-added data processed by the CNES for the Theia www.theia.land.fr data cluster using Copernicus data. The treatments use algorithms developed by Theia's Scientific Expertise Centres."
- **Annotations** stem from the French [Land Parcel Identification System (RPG)](https://www.data.gouv.fr/en/datasets/registre-parcellaire-graphique-rpg-contours-des-parcelles-et-ilots-culturaux-et-leur-groupe-de-cultures-majoritaire/) produced by IGN.
- **SPOT images** are opendata via the [Dataterra DINAMIS](https://dinamis.data-terra.org/opendata/) "Couverture France DINAMIS" program.
NextAV's IP covers the patching, alignment, normalization, annotation curation, split definitions, and documentation of PASTIS-HD as published in this repository.
---
## 📚 References
If you use PASTIS, please cite:
```
@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 PASTIS-R (optical–radar fusion):
```
@article{garnot2021mmfusion,
title={Multi-modal temporal attention models for crop mapping from satellite time series},
author={Sainte Fare Garnot, Vivien and Landrieu, Loic and Chehata, Nesrine},
journal={ISPRS Journal of Photogrammetry and Remote Sensing},
year={2022},
doi={https://doi.org/10.1016/j.isprsjprs.2022.03.012}
}
```
For PASTIS-HD (optical–radar–VHR, OmniSat):
```
@article{astruc2024omnisat,
title={OmniSat: Self-Supervised Modality Fusion for Earth Observation},
author={Astruc, Guillaume and Gonthier, Nicolas and Mallet, Clement and Landrieu, Loic},
journal={ECCV},
year={2024}
}
```
If you use the MAESTRO model or its evaluation on this dataset:
```
@article{labatie2025maestro,
title={MAESTRO: Masked AutoEncoders for Multimodal, Multitemporal, and Multispectral Earth Observation Data},
author={Labatie, Antoine and Vaccaro, Michael and Lardiere, Nina and Garioud, Anatol and Gonthier, Nicolas},
journal={arXiv preprint arXiv:2508.10894},
year={2025}
}
```
---
## 📄 License & IP
All intellectual property in this dataset repository — including the dataset organization, curation, alignment, annotations, normalization statistics, split definitions, colormaps, and documentation — is owned exclusively by **NextAV** © 2026.
The underlying source imagery (Sentinel-1, Sentinel-2, SPOT 6–7) and reference parcel boundaries originate from public data providers (CNES/Theia, IGN RPG, Dataterra DINAMIS) and remain subject to their respective terms of use as described in the Source data provenance section.
For commercial use or licensing inquiries, contact NextAV through the organization page: [https://huggingface.co/nextav](https://huggingface.co/nextav)
---
## 📞 Contact
- Organization: [nextav](https://huggingface.co/nextav)
- Repository: [nextav/PASTIS-HD](https://huggingface.co/datasets/nextav/PASTIS-HD)
- Handbook: [nextav/nextav-handbook](https://huggingface.co/nextav/nextav-handbook)
- Security reports: privately via HF security channels (never in public issues)