Building Footprint Detection

Extracts building footprints from satellite imagery using ESRI's pretrained Mask R-CNN model (usa_building_footprints.dlpk), without ArcGIS Pro.

Download the Model

1. Download from: USA Building Footprints — ArcGIS Online

2. Sign in when prompted. Georgia Tech users: use gtmaps as your organization URL:

   

3. Click Download to get usa_building_footprints.dlpk

4. Run 1_inspect_dlpk.py to extract the model files into model/

Requirements

• Python 3.12
• NVIDIA GPU recommended (CUDA 12.6+)
• uv for environment management

Setup

uv venv
uv add torch torchvision --index "https://download.pytorch.org/whl/cu126" --index-strategy unsafe-best-match
uv add arcgis rasterio Pillow numpy shapely scipy matplotlib

Model

Place the ESRI model files in model/ (extracted from usa_building_footprints.dlpk):

model/
├── usa_building_footprints.emd   # model metadata
└── usa_building_footprints.pth   # weights (Mask R-CNN ResNet50 FPN)

Run 1_inspect_dlpk.py to extract them automatically from the .dlpk file.

Pipeline

input/your_image.tif
        ↓
2_predict.py       → output/footprints.geojson
        ↓
4_regularize.py    → output/footprints_reg.geojson
        ↓
3_visualize.py     → output/overlay_raw.png
                      output/overlay_regularized.png

Step 1 — Inspect model (once)

uv run python 1_inspect_dlpk.py

Step 2 — Detect buildings

uv run python 2_predict.py --image input/your_image.tif

Parameter	Default	Description
--threshold	0.9	Confidence threshold (lower = more detections)
--cell_size	0.3	Target resolution in m/px (model training resolution)
--overlap	256	Tile overlap in pixels (reduces missed edge buildings)
--batch_size	4	Tiles per GPU batch (increase for large images, e.g. 16)

Step 3 — Regularize edges (optional)

Snaps polygon edges to right angles, similar to ArcGIS Pro's Regularize Building Footprint tool.

uv run python 4_regularize.py

Parameter	Default	Description
--tolerance	1.0	Simplification tolerance in map units before snapping

Step 4 — Visualize

uv run python 3_visualize.py

Produces output/overlay_raw.png and output/overlay_regularized.png (if regularized footprints exist).

Output

• output/footprints.geojson — raw detected footprints (geo-referenced polygons)
• output/footprints_reg.geojson — regularized footprints
• output/overlay_raw.png — visual overlay on satellite image
• output/overlay_regularized.png — regularized overlay

Notes

• Input imagery must be a GeoTIFF with a projected CRS (EPSG:3857 recommended)
• The model was trained on US imagery at 0.3 m/px; imagery is automatically resampled to this resolution
• GPU is used automatically if available; CPU fallback is supported