microsoft/GlobalMLBuildingFootprints

Worldwide building footprints derived from satellite imagery

What it is

Microsoft’s Global Building Footprints repository provides a massive, openly‑licensed dataset of building‑level polygon geometries (and, where available, height estimates) for almost the entire planet. The data are derived from aerial and satellite imagery processed with deep‑learning models that perform semantic segmentation and polygonization.

Why it matters

  • Scale – Over 1 billion building footprints, with hundreds of millions of height estimates, making it one of the largest publicly available 3‑D urban datasets.
  • Open licensing – Released under the CDLA Permissive 2.0 license, so anyone can download, remix, or integrate the data into commercial or research projects.
  • Geographic breadth – Coverage spans 225 regions (30 340 tiles) and includes detailed per‑country partitions, plus separate layers showing where height data are available.
  • AI‑driven creation – The footprints are generated by a two‑stage pipeline (semantic‑segmentation neural net → polygonization), and height values come from a neural network trained on imagery paired with ground‑truth elevations.

How to use it

  1. Download – The dataset-links.csv file lists URLs for gzipped line‑delimited GeoJSON (.csv.gz) files, each keyed by a Bing Maps quad‑key.
  2. Read – Files can be streamed with standard tools (e.g., gzip + jq) or processed in parallel with Spark/Dask. Example scripts are provided (scripts/read-large-files.py, scripts/make-gis‑friendly.py).
  3. Integrate – Load the polygons into GIS software (QGIS, ArcGIS) or spatial databases (PostGIS, BigQuery GIS) for mapping, analysis, or as input to downstream AI models (e.g., 3‑D city reconstruction, urban planning, disaster response).
  4. Height data – When present, the height attribute (meters) is the mean of per‑pixel height predictions inside the footprint; missing values are -1.
  5. Confidence scores – A confidence field (0‑1) indicates the model’s certainty for each footprint; -1 means “not available”.

Quality & limits

  • Precision/Recall varies by region (e.g., Europe ≈ 94 % / 86 %, Africa ≈ 94 % / 71 %).
  • False‑positive rates are generally around 1 % per country sample.
  • Resolution depends on the vintage of the source imagery (2014‑2024) and on tile availability; some remote squares are omitted.
  • Height coverage is far sparser than footprint coverage (≈ 174 M heights vs 1 B footprints).

Who might need it?

  • Researchers building urban AI models (e.g., 3‑D city generation, traffic simulation).
  • NGOs and governments conducting disaster risk assessments or infrastructure planning.
  • Companies creating location‑based services, augmented reality, or logistics routing that require building geometry.
  • OpenStreetMap contributors looking for bulk import candidates (with the usual community guidelines).

Getting started

# Grab the index of files
curl -O https://bfppub.blob.core.windows.net/%24web/2026-08-13/dataset-links.csv

# Example: download a single tile (replace <quadkey> with the desired key)
wget https://bfppub.blob.core.windows.net/global-buildings/<quadkey>.csv.gz

gunzip -c <quadkey>.csv.gz | head   # view a few GeoJSONL lines

Related projects

  • USBuildingFootprints, AustraliaBuildingFootprints, etc. – country‑specific releases.
  • RoadDetections – a companion dataset of AI‑derived road vectors.

All details are taken directly from the repository’s README; no additional features are inferred.

Related

  • Project
    e-sensing/sits

    SITS is an open‑source R package (with a Python wrapper) for building, regularising, and classifying satellite‑image time‑series data cubes. It connects to many public EO collections, offers a suite of ML/DL models (including GPU‑accelerated CNNs and attention encoders), and provides end‑to‑end tools for training, classification, smoothing, uncertainty estimation, and active learning.

  • Project
    google-deepmind/tapnet

    Google DeepMind’s TAP‑Net repository offers datasets, pretrained models (TAPIR, BootsTAPIR, TAPNext/Next++), training code, and Colab/real‑time demos for the “Tracking Any Point” task—precise point‑level video tracking useful for vision research and robot imitation.

  • Project
    haoyi-duan/WorldScore

    WorldScore is an open‑source benchmark for evaluating world‑generation models (3‑D, 4‑D, video). It provides a dataset, generation adapters, a suite of spatial‑temporal metrics, and a public leaderboard. Users register their model via a small YAML config and a Python class, generate videos, run the evaluation script, and submit the resulting JSON score.

  • Dispatch
    Microsoft AI MAI-Thinking-1 Release

    Microsoft AI has introduced MAI-Thinking-1, a medium-sized reasoning model featuring a 35B-active parameter MoE architecture designed for high-performance software engineering and mathematical reasoning.

  • Project
    zubair-trabzada/geo-seo-claude

    A Claude Code skill that audits a website for AI‑search (GEO) visibility, scoring citability, brand signals, content quality, technical SEO and schema, then generates markdown or PDF client reports.