Toward accurate building footprint extraction from high-resolution imagery using open and reproducible spatial-AI workflows.
Research question
Can open spatial-AI pipelines extract geometrically accurate building footprints from high-resolution imagery and support repeatable municipal-scale building intelligence?
Problem
Conventional land-cover classification and change-detection workflows can identify built-up areas without accurately representing individual building boundaries. The outputs typically produce raster masks or coarse polygons that do not follow actual roof geometry.
Property intelligence, urban planning, infrastructure monitoring, taxation and disaster-exposure workflows all require building-level geometry: individual polygons that trace the physical footprint of each structure. The gap between "there are buildings here" and "here is the shape of each building" is the core problem this research addresses.
Why it matters
Accurate building footprints are foundational objects for a wide range of municipal and engineering applications:
- Property intelligence — enumeration, area estimation, unauthorized construction detection
- Urban growth monitoring — temporal comparison of footprint inventories
- Change detection — identifying new construction, demolition, or modification
- Infrastructure planning — road network, drainage, utility corridor design
- Disaster exposure analysis — flood, earthquake, cyclone impact estimation
- Building inventory creation — census-independent enumeration from imagery
- Digital twins — 2.5D/3D city models derived from footprints plus height estimation
When footprints are inaccurate, every downstream workflow inherits the error. When they are accurate and reproducible, they become a platform-level asset that multiple applications share.
Research approach
The methodology follows a progressive pipeline from imagery to validated geometry:
- AOI selection — Select a compact area (1–3 km²) containing a representative mixture of building types: dense urban, detached residential, irregular roofs, closely spaced structures, large institutional/commercial buildings, and areas with vegetation/shadow interference.
- Reference acquisition — Obtain building footprints from multiple established sources for the same AOI to enable meaningful comparison.
- Imagery acquisition — High-resolution optical satellite or aerial imagery covering the AOI at a resolution suitable for individual building extraction (sub-meter to 1–2 m GSD).
- Segmentation — Apply building segmentation models to produce pixel-level building masks.
- Polygonization — Convert raster masks into vector polygons.
- Boundary regularization — Straighten edges, enforce right angles where appropriate, remove slivers and artifacts.
- Topology cleanup — Remove overlaps, fill gaps, enforce valid geometry.
- Quality assessment — Compare extracted footprints against reference sources using established metrics.
- Failure analysis — Manually inspect cases where extraction diverges from reference, categorize failure modes.
Comparison framework
Building Intelligence compares footprint sources rather than evaluating a single extraction in isolation. For any given AOI, the comparison includes:
- Authoritative/local data — Government survey or municipal records where available
- OpenStreetMap — Community-contributed footprints
- Microsoft Building Footprints — ML-generated global dataset
- Google/Open Buildings — Where applicable and licensing permits
- Nebula Cloud extraction — Studio-generated footprints using open models and the research pipeline
The research question is not "which source is best" but rather: how does an open, reproducible extraction pipeline compare against established global and local datasets in terms of geometry quality, completeness and fitness for municipal use?
Evaluation methodology
Quality assessment uses established geospatial accuracy metrics:
- Precision — Proportion of extracted footprints that correspond to real buildings
- Recall — Proportion of real buildings that are detected
- F1 — Harmonic mean of precision and recall
- IoU (Intersection over Union) — Geometric overlap between extracted and reference polygons
- Boundary F1 — Edge-level accuracy measuring how well boundaries align
- Completeness — Coverage of the building inventory
- Commission/Omission — False positives and false negatives
- Geometry validity — Self-intersections, ring direction, topology
- Area difference — Absolute and relative area deviation per building
- Positional error — Centroid and boundary displacement
Critically, this research evaluates shape quality — not merely whether the model found a building, but whether the extracted polygon accurately represents its physical boundary.
Environment
Building Intelligence uses Nebula Cloud’s Spatial Intelligence Workbench:
- Platform — QGIS, GDAL/OGR, PostGIS, Python (rasterio, geopandas, shapely)
- Segmentation — Open building-extraction models (evaluated per experiment)
- Comparison — Custom quality-assessment scripts producing per-building metrics
- Visualization — QGIS for spatial comparison, overlay and inspection
- Orchestration — Nebula Cloud Studio for pipeline execution and reproducibility
Current status: Active research. AOI selection and comparison-source acquisition in progress. The first controlled benchmark experiment (Experiment 01) will cover a compact 1–3 km² urban/peri-urban area with multiple reference sources. Results will be published on this page when validation is complete.
Planned experiments
Experiment 01 — Controlled benchmark
A small-area comparison using a carefully selected AOI containing:
- Dense urban buildings with shared walls
- Detached residential structures
- Irregular and complex roof geometries
- Closely spaced buildings where individual separation is challenging
- Large institutional/commercial footprints
- Areas with heavy vegetation and shadow interference
The experiment will produce a reference benchmark dataset for the selected AOI, run the extraction pipeline, and generate per-building quality metrics and visual comparison outputs.
Subsequent experiments
Based on Experiment 01 findings, subsequent work may address:
- Boundary regularization quality (right-angle enforcement, edge simplification)
- Performance at different imagery resolutions (30 cm vs 50 cm vs 1 m)
- Temporal change detection using multi-date footprint inventories
- Scaling behavior across larger AOIs
- Model comparison (different segmentation architectures on the same AOI)
Potential applications
If validated, building footprint extraction at sufficient quality enables:
- Municipal property intelligence — automated building enumeration and area estimation for property tax, land records, unauthorized construction detection
- Urban planning — growth pattern analysis, density calculation, land-use characterization
- Disaster preparedness — building exposure databases for flood, earthquake, cyclone risk models
- Infrastructure design — building context for road, drainage and utility planning
- Change monitoring — repeatable temporal inventories showing new construction and demolition
- Digital twin creation — footprints as the 2D foundation for 3D city models
Connection to Nebula Cloud
Building Intelligence research connects to the broader Nebula Cloud Labs program. Research that reaches validated maturity may be operationalized as capabilities within Nebula Cloud Studio, the Spatial Intelligence Workbench, or Capability Packs — making reproducible building extraction available as a platform workflow rather than a one-off analysis.