🏙️ UrbanGraphSAGE

Building footprint extraction from Sentinel-2 satellite imagery using GraphSAGE — A case study on Algiers, Algeria.

---

Overview

UrbanGraphSAGE applies Graph Neural Networks to medium-resolution satellite imagery for urban building extraction. Instead of treating pixels independently, we construct superpixel graphs that capture spatial relationships, enabling the model to leverage neighborhood context for superior segmentation.

The Approach

graph LR
    A["🛰️ Sentinel-2<br/>13 bands"] --> B["🔲 SLIC<br/>Superpixels"]
    B --> C["📊 Feature<br/>Extraction"]
    C --> D["🔗 Graph<br/>Construction"]
    D --> E["🧠 GraphSAGE<br/>Message Passing"]
    E --> F["🏗️ Node<br/>Classification"]
    F --> G["🗺️ Building<br/>Footprint Map"]

Loading

Stage	Method	Details
Preprocessing	Sentinel-2 L2A	Cloud masking, tiling (256×256), band normalization
Superpixels	SLIC (scikit-image)	~500 superpixels/tile, compactness=20
Features	Spectral indices	13 bands + NDVI + NDWI + NDBI per superpixel
Graph	k-NN adjacency	k=8, spatial + spectral similarity
Model	GraphSAGE (PyG)	3 layers, mean aggregation, 256 hidden dim
Labels	OpenStreetMap	Building polygons → superpixel majority vote

Results

Model	F1-Score	IoU	Parameters
U-Net	0.81	0.69	31.0M
DeepLabv3+	0.83	0.72	41.0M
SegFormer-B2	0.85	0.74	27.5M
UrbanGraphSAGE	0.88	0.79	12.3M

Installation

git clone https://github.com/OMUZ9924/UrbanGraphSAGE.git
cd UrbanGraphSAGE
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt

Quick Start

# Preprocess Sentinel-2 tiles into superpixel graphs
python -m src.preprocessing --input data/raw/ --output data/processed/

# Train GraphSAGE model
python -m src.train --config configs/default.yaml

# Run inference
python -m src.predict --checkpoint checkpoints/best.pt --input data/processed/test/

Project Structure

UrbanGraphSAGE/
├── configs/default.yaml
├── src/
│   ├── preprocessing.py    # Sentinel-2 → superpixel graph pipeline
│   ├── graph_construction.py # Graph building from superpixels
│   ├── model.py            # GraphSAGE architecture
│   ├── train.py            # Training loop with validation
│   ├── predict.py          # Inference and evaluation
│   └── utils.py            # Metrics, visualization, seed
├── tests/
├── notebooks/
├── .github/workflows/ci.yml
├── Dockerfile
├── requirements.txt
└── README.md

Citation

@thesis{arbouz2025urbangraphsage,
  title   = {Graph Neural Networks and Deep Learning for Large-Scale Remote Sensing},
  author  = {Arbouz, Maamar},
  year    = {2025},
  school  = {University Djilali Bounaama, Khemis Miliana}
}

License

MIT — see LICENSE for details.

---

⭐ Star this repo if you find it useful!