Open Land Use Reference Dataset for Palm Oil Landscapes in Indonesia

Description

Technical info

The dataset was developed through an integrative data collection and validation approach:

• Wall-to-Wall Labeling:
  Training data was created through opportunistic sampling across 6x6 km grids overlapping known oil palm regions. Labels were digitized in QGIS and refined through multi-spectral analysis of Planet NICFI imagery from corresponding months. The land cover classification followed a two-tiered typology covering estate crops (e.g., oil palm, rubber, coconut), annual crops, forests, shrubland, wetlands, and other land uses.

• Satellite Imagery Pairing:
  Each labeled grid is paired with PlanetScope monthly composite images clipped to grid extent. These cloud-minimized mosaics are consistent with the date of interpretation to support supervised machine learning training.

• Validation Dataset (CEO):
  A stratified random sampling method was applied to generate verification points across diverse land cover classes. Interpreters labeled points in CEO using a standardized survey form and high-resolution imagery. Twenty percent of samples were cross-verified by multiple interpreters to calculate agreement scores as part of QA/QC procedures. Select ambiguous classes were also validated through targeted field checks.

• Metadata and Format:
  All spatial layers are delivered in GeoJSON format, accompanied by metadata following ISO 19115 standards. Image tiles are delivered as GeoTIFFs, pre-aligned to vector boundaries.

Data Type:

Vector spatial dataset (Polygons, WKT), tabular metadata (.csv/.geojson), JSON-compatible attributes, raster images (GeoTIFF)

Dataset Structure:

Land Cover Class Mapping:

Annotations:

Labels are expert-generated through remote sensing interpretation and supplemented by input from local field teams. CEO platform allowed interpreters to draw and describe features using temporal image stacks. Cross-checking and interpreter discussions were used to maintain consistency.

Relations to Existing Work:

This dataset was independently created for this project but complements existing regional datasets such as the Forest Data Partnership oil palm probability maps.

Considerations for Using the Data:

This dataset is intended to support machine learning model training and validation for land cover mapping in tropical agricultural regions. Users should consider potential limitations related to cloud cover in source imagery and differences in seasonal appearance of crops. It is particularly valuable for distinguishing oil palm from visually similar classes such as coconut, rubber, and forest.

Associated Imagery:

Each labeled polygon is linked to pre-processed, high-resolution satellite imagery that was used during interpretation. These images are cloud-free and temporally aligned with the label data to facilitate immediate use in AI/ML workflows.

The dataset release also includes multiple layers of reference imagery and data products designed to support advanced model training and validation:

1. Shapefile of labeled polygons

2. Raster file encoding Polygon ID (0.3-meter resolution)

3. Raster file encoding Land Cover Class ID (0.3-meter resolution)

4. Sentinel-2 composite (RGB + NIR, 10-meter resolution)

5. Sentinel-2 composite (Red Edge bands 1-3, SWIR1 & SWIR2, 20-meter resolution)

6. Landsat 8/9 composite (RGBN, 30-meter resolution)

7. PlanetScope composite (RGBN, 4.7-meter resolution)

8. Additional high-resolution commercial imagery (in selected sample grids)

All imagery products have been processed for spatial alignment with the labeled data and provided in GeoTIFF format for direct use in remote sensing applications and AI model development.