MapSpaceORNL/remote-territory-landuse-11class-v1

MapSpace LSG 2026: Remote Territory 11-Class Land Use

Dataset Details

This dataset contains POI-based land-use estimates for LandScan Global grid cells in Remote Territory. It was generated and is shared by the Geospatial Science and Engineering Division at Oak Ridge National Laboratory.

• Curated by: Geospatial Science and Engineering Division, Oak Ridge National Laboratory
• Version: v1
• License: CC BY 4.0
• Geographic coverage: Remote Territory
• Spatial grid: 30 arc seconds (1/120 degree)
• CRS: EPSG:4326
• Format: GeoParquet, one country or territory per file
• Files: 14
• Rows: 383,379
• Contact: mapspace@ornl.gov

The Spratly Islands file is intentionally excluded from this release.

Each row represents one grid cell and contains probabilities for 11 land-use classes, an argmax label, prediction diagnostics, and the cell geometry. These values are model outputs or a documented no-feature fallback; they are not ground-truth land-use observations.

This is an 11-class, 30-arc-second product. It is distinct from the earlier three-class MapSpace dataset cited under Related Work.

Dataset Structure

The repository root contains 14 GeoParquet files. The current schema is:

Column	Type	Description
id	string	Grid-cell identifier.
global_row	int64	Global LandScan grid row.
global_col	int64	Global LandScan grid column.
ls23_country	string	Country or territory attribute inherited from the LandScan grid.
geom	binary	Polygon or MultiPolygon encoded as GeoArrow/GeoParquet WKB in EPSG:4326.
agricultural	float32	Agricultural probability.
commercial	float32	Commercial probability.
education	float32	Education probability.
industrial	float32	Industrial probability.
institutional	float32	Institutional probability.
military	float32	Military probability.
open_space	float32	Open-space probability.
recreation	float32	Recreation probability.
residential	float32	Residential probability.
retail	float32	Retail probability.
transportation	float32	Transportation probability.
landuse_label	string	Class with the largest probability.
landuse_probability	float32	Largest class probability.
confidence_level	string	Probability bin: high, medium, low, or mixed.
prediction_entropy	float32	Natural-log entropy, -sum(p * ln(p)).
label_source	string	model_prediction or no_feature_fallback.

Class columns appear in the order shown above. Exact probability ties are resolved by that order.

Confidence levels are assigned from landuse_probability as follows:

• high: at least 0.70
• medium: at least 0.50 and below 0.70
• low: at least 0.35 and below 0.50
• mixed: below 0.35

Label provenance

label_source	Rows	Percent	Meaning
model_prediction	7,960	2.076%	Probabilities produced by the land-use model.
no_feature_fallback	375,419	97.924%	No POI/tag feature was available; release construction assigns a one-hot open_space vector.

no_feature_fallback does not mean that the model or an independent observation established that the cell is open space. Filter on label_source == "model_prediction" when analyses require model-derived labels only.

Source Data and Method

Inputs include OpenStreetMap and PlanetSense Program POI/feature tags joined to the LandScan Global grid. Training labels are derived from OpenStreetMap land-use, POI, and natural-feature polygons mapped to the 11 classes. The pipeline uses TF-IDF features and XGBoost multiclass probability models. Models are fitted by source area; areas without sufficient training data may use a donor model.

Exact snapshot dates for OpenStreetMap, PlanetSense, and LandScan inputs are not supplied in this release.

Uses

Potential direct uses include urban planning, transportation planning, population modeling, and disaster-risk assessment. Users should validate fitness for their geography and application before operational use.

This dataset should not be treated as parcel-level land use, legal zoning, surveyed ground truth, or a substitute for local authoritative data.

Bias, Risks, and Limitations

• OpenStreetMap-derived training labels are weak labels and vary in completeness, positional accuracy, tagging practice, and class balance by geography.
• Model performance can therefore vary across countries, regions, and classes. Sparse areas may depend on geographically transferred donor models.
• POI and feature coverage is uneven. A large share of rows may use no_feature_fallback; these rows require separate interpretation.
• The 30-arc-second grid is angular, not equal-area. East-west cell width decreases with latitude.
• Independent input snapshots can create temporal mismatch.
• A consolidated accuracy estimate for this regional release is not provided. Do not infer accuracy from landuse_probability or confidence_level alone.

License and Attribution

The dataset is released under the Creative Commons Attribution 4.0 International license. Attribute the Geospatial Science and Engineering Division at Oak Ridge National Laboratory and cite the repository used.

Citation

No DOI has been assigned to this v1 regional release. A suggested citation is:

@dataset{mapspace_lsg_2026_remote_territory,
  author    = {{Geospatial Science and Engineering Division, Oak Ridge National Laboratory}},
  title     = {{MapSpace LSG 2026: Remote Territory 11-Class Land Use}},
  year      = {2026},
  version   = {v1},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/MapSpaceORNL/remote-territory-landuse-11class-v1}
}

Related Work

The following references describe prior MapSpace datasets and methods; the 2024 DOI refers to the earlier three-class dataset, not this release.

• Fan, Junchuan, and Gautam Thakur. (2024). Three-class Global POI-based land use map. Dataset. https://doi.org/10.17605/OSF.IO/395ZF
• Fan, J., and Thakur, G. (2023). “Towards POI-based large-scale land use modeling: spatial scale, semantic granularity and geographic context.” International Journal of Digital Earth, 16(1), 430–445.
• Thakur, G., and Fan, J. (2021). “MapSpace: POI-based Multi-Scale Global Land Use Modeling.” GIScience Conference 2021.

Contact

mapspace@ornl.gov