Description

Technical info

Record 1 — Geospatial Labels (Oil Palm Landscapes, Indonesia)

Wall-to-wall, human-annotated land-use / land-cover reference polygons for oil-palm landscapes across ten provinces of Indonesia (Sumatra, Kalimantan, Sulawesi), digitised on a systematic 6 × 6 km grid. This record contains the author-created label data only. The satellite imagery the labels were drawn on is distributed separately (see Related records below), because it carries different, source-specific licenses.

License

All contents of this record are released under Creative Commons Attribution 4.0 International (CC-BY-4.0) — see LICENSE.txt. You are free to share and adapt the data for any purpose, including commercially, provided you give appropriate credit.

Cite as: Wafiq, M. W., et al. Open Land Use Reference Dataset for Palm Oil Landscapes in Indonesia. Zenodo. https://doi.org/10.5281/zenodo.15618532

Contents

record1_geospatial_labels/
├── README.md
├── LICENSE.txt                 CC-BY-4.0 legal code
├── data_dictionary.csv         field-by-field description of every label attribute
├── class_schema.csv            canonical class_ID → name, variants, oil-palm flag
├── grid_index.csv              per-grid province, island, CRS, polygon count, label year
├── grid_province_lookup.csv    grid → province → island
├── province_summary.csv        per-province polygon / area / cell totals
├── labels/
│   └── grid_XX/                (XX = 01 … 72)
│       ├── user_land_use_grid_XX.{shp,gpkg,geojson} (+ shapefile sidecars)  annotation polygons
│       ├── user_boundary_grid_XX.shp (+ sidecars)                          cell boundary + QA
│       └── land_cover_classification_grid_XX.tif (+ sidecars)              0.3 m rasterised map
└── notebooks/                  runnable benchmark notebooks + their own README
    ├── README.md               how to run the benchmark end-to-end
    ├── 00_Prepare_Shared_Tiles │ 01_AlphaEarth_Benchmark │ 02_Clay_Sentinel2_Benchmark
    ├── rf_baseline.py          classical Random-Forest baseline + shared metric code
    └── environment.yml         conda environment

Key files

• user_land_use_grid_XX — the core deliverable: wall-to-wall polygons, provided in three mutually consistent formats (Shapefile, GeoPackage, GeoJSON). Attributes:
  • id, plotid — internal and Collect Earth Online identifiers;
  • class_ID, class_ENG, class_BAH — canonical class code and English / Bahasa names (class_ID is authoritative; class_ENG names vary across grids and are reconciled in class_schema.csv);
  • class_l1, class_l2, class_l3 — the hierarchical typology (broad category → thematic subdivision → detailed class), derived from class_ID;
  • img_date — reference-image date for the cell, propagated to its polygons;
  • source_sensor — interpretation imagery stack (campaign-level);
  • interpreter_id — annotator identifier (recorded value where present, else the team-level MoHE_intern_team).

  See data_dictionary.csv for full definitions and allowed values.

• class_schema.csv — 17 classes. Oil palm = class_ID 1 (Palm, initial planting) and 2 (Palm, mature/young). Coconut (ID 5) is a separate class and is not oil palm.
• land_cover_classification_grid_XX.tif — a 0.3 m rasterised land-cover product derived from the polygons. Provided for completeness; the paper's benchmark rasterises the vectors directly rather than using this file.

Coordinate reference system

The annotation polygons (user_land_use_*) are distributed in WGS 84 (EPSG:4326) in all three formats. grid_index.csv records each grid's native imagery UTM CRS (province-dependent, e.g. EPSG:32750 for West Sulawesi, EPSG:32647/32648 for Riau), which is the CRS of the user_boundary_* files and of the separately distributed imagery.

Shapefile note. The .dbf format limits field names to 10 characters, so in the Shapefile only source_sensor appears as src_sensor and interpreter_id as interp_id. The GeoPackage and GeoJSON carry the full field names. All other names are identical across formats.

Benchmark notebooks (notebooks/)

This record also ships the runnable benchmark code — the same notebooks used for the Technical Validation in the data descriptor, written as a standalone tutorial for readers with no remote-sensing or deep-learning background. See notebooks/README.md for the full walkthrough (setup, hardware, and how to run it end-to-end).

• 00_Prepare_Shared_Tiles — downloads this labels record plus the companion imagery record and builds the shared 128×128 tiles and train/val/test split every model reuses.
• 01_AlphaEarth_Benchmark — Google Satellite Embedding V1 (AlphaEarth) features under a U-Net head and a Random Forest.
• 02_Clay_Sentinel2_Benchmark — Clay v1 (frozen ViT + FPN head) on Sentinel-2, paired against the AlphaEarth row.
• rf_baseline.py — classical Random-Forest baseline and the shared metric code every model is scored with; environment.yml pins the conda environment.

The notebooks pull the annotation polygons from this record's concept DOI (10.5281/zenodo.15618531), so they always track the latest published version.

Provenance & quality

This record contains the paper's 72 canonical grid cells (Sumatra 31, Sulawesi 27, Kalimantan 14). All 72 carry wall-to-wall land-use polygons (106–7,000+ polygons each) and Planet imagery. Imagery coverage of the other sensors is not uniform: grids 03, 18, 46 have no Sentinel-2 10 m composite (so no AlphaEarth), and grid 06 has no Landsat — see the imagery record's coverage table. Labels were produced by trained interpreters in Collect Earth Online / QGIS with multi-interpreter consensus and field validation (reported Overall Accuracy ≈ 83 %, reflecting internal thematic consistency and inter-interpreter agreement). See the accompanying data descriptor for the full methodology.

Related records

• Grid-Aligned Open Satellite Imagery for Indonesian Oil Palm Mapping (Sentinel-2, Landsat 8/9, AlphaEarth) — Open satellite imagery and embeddings aligned to these grid cells.
• Grid-Aligned Planet NICFI Imagery for Indonesian Oil Palm Mapping (Non-Commercial) — High-resolution Planet NICFI mosaics aligned to these grid cells (distributed under a non-commercial license).