Featured image of post Building Footprint Extraction with U-Net
Deep Learning Remote Sensing Building Extraction

Building Footprint Extraction with U-Net

Automatic extraction of building footprints from NAIP imagery using a ResNet-50 U-Net and OpenStreetMap labels

This post demonstrates how we trained a U-Net (with a ResNet-50 encoder pretrained on ImageNet) to extract building footprints from high-resolution NAIP imagery using OpenStreetMap polygons as labels.

Data

  • Input imagery: NAIP 1 m RGB tiles over Orange County, CA
  • Labels: Rasterized building footprints from OpenStreetMap

Model & Training

  • Architecture: U-Net with a ResNet-50 encoder, skip-connections, and sigmoid output for binary masks
  • Loss: Dice loss to handle class imbalance
  • Optimizer: Adam (lr = 1e-4) with CosineAnnealingWarmRestarts
  • Augmentations: Random flips, 90° rotations, affine jitter
  • Training: 20 epochs on 512×512 patches (10 initial + 10 resumed)

Convergence Curves

Loss vs Epoch / IoU vs Epoch
Loss vs Epoch / IoU vs Epoch

  • Left: Dice loss decreases on both training (blue) and validation (orange).
  • Right: Intersection-over-Union rises from ~0.52 up to ~0.74 over 20 epochs.

Sample Predictions

Sample 1

Sample 1
Sample 1

Sample 2

Sample 2
Sample 2

Sample 3

Sample 3
Sample 3

Sample 4

Sample 4
Sample 4

Vectorized Footprints

After cleaning and simplifying the predicted masks, we convert them to vector polygons and overlay on the original imagery:

Cleaned Building Footprints Overlay
Cleaned Building Footprints Overlay

References

© 2020 - 2025 Yinan's Personal Website
Built with Hugo
Theme Stack designed by Jimmy