This repository contains the GlobalBuildingAtlas (GBA) Level of Detail 1 (LoD1) building footprints dataset, converted from the original 922 GeoJSON files into Parquet format for improved performance and accessibility.

Dataset Overview

The GlobalBuildingAtlas LoD1 dataset contains building footprints for approximately 2.75 billion buildings worldwide. This represents the most comprehensive global building dataset available, covering nearly all inhabited areas on Earth.

The dataset combines multiple existing building footprint datasets with new footprints generated using open source deep learning models applied to Planet Labs satellite imagery.

This Parquet version was converted from the original 922 GeoJSON files (1.1 TB) to improve performance and accessibility. For detailed information about the conversion process and technical analysis, see Mark Litwintschik's comprehensive blog post.

Data Format

The dataset is organized as Parquet files, each containing building footprint data with the following key attributes:

• Building footprints: Vector polygons representing building outlines
• Building heights: Height information for each building
• Geographic coordinates: Precise location data
• Metadata: Additional building characteristics

Parquet Structure

Each Parquet file contains:

• Building geometry (polygons)
• Height attributes
• Spatial indexing for efficient querying
• Optimized columnar storage for fast analytics

Data Sources and Methodology

The dataset combines multiple building footprint sources:

1. Google Open Buildings (2023): 1.62 billion buildings
2. OpenStreetMap (2025): 490 million buildings
3. Microsoft Building Footprints (2024): 432 million buildings
4. TUM Deep Learning (ours2): 135 million buildings from Planet Labs satellite imagery
5. 3D Global Footprints: 68 million buildings from previous research

The TUM deep learning component uses open source computer vision models applied to Planet Labs satellite imagery. Planet Labs operates several constellations totaling hundreds of satellites in low Earth orbit, capturing images of the entire Earth's landmasses daily.

Data Limitations

• Temporal Accuracy: No timestamps for AI-generated footprints
• Data Conflicts: Some footprints may conflict with newer OpenStreetMap data
• Construction Areas: Recent construction may not be captured
• Validation: Manual verification recommended for critical applications

Usage Recommendations

• Map Making: Use as base layer with selective integration of newer OSM data
• Construction Areas: Prioritize latest OSM data for areas with recent development
• Research: Suitable for large-scale analysis and modeling
• Applications: Urban planning, disaster response, infrastructure analysis

Technical Specifications

• Coordinate System: WGS84
• File Format: Parquet
• Compression: Optimized for fast querying
• Spatial Indexing: H3 and other spatial indices included
• Total Size: 210 GB (compressed from 1.1 TB original)

Data Quality Notes

The dataset combines multiple data sources as detailed above. The TUM deep learning component (ours2) represents new building footprints generated from Planet Labs satellite imagery, while the other sources provide existing building data from various organizations.

For areas with significant recent construction, consider using the latest OpenStreetMap data or Overture's building dataset, which prioritizes recent OSM updates.

Citation

If you use this Parquet version of the dataset hosted on Source Cooperative, please cite both the original dataset and this hosted version:

Original Dataset:

Source Cooperative Hosted Version: