How allenai/planteur-benchmark works
PLANTEUR is a satellite time-series benchmark for agricultural parcel semantic and panoptic segmentation. Its principal workflow is to prepare crop labels and spatial splits, export aligned Sentinel and Landsat observations plus annotations, then load those patch-level samples or derive parcel pixel sets for model training; variable-length batches can be padded separately. The README also documents monthly mosaics and precomputed embeddings as benchmark products. Generation and loading relationships are grounded in the shown scripts and README; some ancillary generation and product-export details are not sampled.
Main entry points: dataloader.py Patch loader · export_planteur_raw.py S2/S1 export · dataloader_pixelset.py Parcel pixel sets · collate.py Temporal padding
Components 16
Dataset generation
- S2/S1 export
data_generation/export_planteur_raw.py - Landsat export
data_generation/export_landsat_raw.py - Metadata rebuild
data_generation/rebuild_meta.py - S1 orbit metadata
data_generation/add_orbits_s1.py
Labels and splits
- RPG class mapping
data_generation/build_pastis_rpg_map.py - Spatial splits
data_generation/assign_splits.py - Panoptic labels
data_generation/export_planteur_panoptic.py
Dataset access
- Patch loader
code/dataloader.py - Parcel pixel sets
code/dataloader_pixelset.py - Temporal padding
code/collate.py
Benchmark products
- PLANTEUR dataset
- Monthly mosaics
- Precomputed embeddings
- Class distribution reports
data_generation/plot_class_distributions.py
Other components
- Benchmark user
- Hugging Face Hub
Connections 17
- RPG class mapping → PLANTEUR dataset: defines classes
data_generation/build_pastis_rpg_map.py - Spatial splits → PLANTEUR dataset: assigns splits
data_generation/assign_splits.py - S2/S1 export → PLANTEUR dataset: exports S2/S1
data_generation/export_planteur_raw.py - Landsat export → PLANTEUR dataset: exports Landsat
data_generation/export_landsat_raw.py - Panoptic labels → PLANTEUR dataset: writes instances
data_generation/export_planteur_panoptic.py - Metadata rebuild → S2/S1 export: reuses helpers
data_generation/rebuild_meta.py - S1 orbit metadata → PLANTEUR dataset: adds orbit data
data_generation/add_orbits_s1.py - Hugging Face Hub → Benchmark user: hosts dataset
README.md - Benchmark user → Patch loader: loads patches
README.md - Patch loader → PLANTEUR dataset: reads arrays
README.md - Benchmark user → Parcel pixel sets: loads parcels
README.md - Parcel pixel sets → Patch loader: uses loader API
code/dataloader_pixelset.py - Benchmark user → Temporal padding: pads batches
README.md - Temporal padding → Patch loader: pads sequences
README.md - Benchmark user → Monthly mosaics: loads mosaics
README.md - Benchmark user → Precomputed embeddings: uses embeddings
README.md - Class distribution reports → PLANTEUR dataset: reads labels
data_generation/plot_class_distributions.py