Building Footprint Extraction for Urban Planning
Custom pipeline development
DINOv3 building detection from VHR or Sentinel-2 imagery, fine-tuned on local structures.
Extract building footprints across your city using DINOv3 ViT-L/14 fine-tuned on local structure types — including the informal compound structures common across West Africa that global models miss. We deliver a GeoJSON of all detected buildings with area, perimeter, compactness, and estimated floor count.
Accuracy benchmarks on 5 African cities: 64.7% mean IoU zero-shot, 78.3% after 2,500-sample local fine-tuning. After fine-tuning to your specific city, expect 80–85% IoU on structured urban areas and 65–75% on dense informal settlements.
Deliverables: GeoJSON of building footprints, fine-tuned model weights (ONNX + PyTorch), inference pipeline for future updates, accuracy report with per-zone breakdown.
What's included
Data acquisition
Imagery sourced from 22+ providers, matched to your area and time window.
AI analysis
Segmentation, detection or change analysis through our pygeovision pipelines.
Clear deliverables
Maps, GeoJSON layers and reports you can use directly or hand to stakeholders.
Full transparency
Auditable, reproducible workflows — every result traces back to its source scene.
How delivery works
Scope
We agree objectives & AOI.
Build
We develop the pipeline.
Review
You validate the results.
Deliver
Final outputs & handover.
One-off project, quoted precisely to your scope.
- Built on open-source tooling
- No proprietary lock-in
- Reproducible & auditable