🛰️ Sentinel-2 Built-Up Area Detection Pipeline

An automated, modular ML pipeline for mapping built-up areas from Sentinel-2 satellite imagery — using spectral indices, Random Forest classification, and Microsoft Planetary Computer.

---

📌 What Is This?

Mapping built-up (urban) areas at scale is essential for urban planning, climate risk assessment, and land-use monitoring. Traditional methods require manual digitization or expensive commercial data.

This pipeline automates the full workflow — from satellite data acquisition to a classified built-up map — using free Sentinel-2 imagery and open-source ML tools. It is designed for city-scale and regional-scale automation, supporting both research and production use.

---

🔄 Pipeline Overview

1. Download Sentinel-2 tiles (Planetary Computer STAC API)
       ↓
2. Compute per-tile spectral index composites
       ↓
3. Build ML training dataset from training polygons / points
       ↓
4. Train Random Forest classifier
       ↓
5. Predict built-up probability & binary maps per tile
       ↓
6. Evaluate accuracy (F1, Precision, Recall, Confusion Matrix)

---

✨ Key Features

📥 Automated Sentinel-2 Downloader

• STAC API via Microsoft Planetary Computer
• Cloud filtering using scene metadata
• Downloads selected bands at 10m & 20m resolution
• Tile-aware search for large AOIs
• Resume-safe (skips already downloaded files)

🛰️ Spectral Index Processing

• Per-tile mean & median composites
• SCL cloud masking applied before aggregation
• Indices computed: NDVI, NDBI, BSI, NDWI, MNDWI

🤖 Random Forest Classification

• Supports both point and polygon training data
• Balanced sampling from polygon regions
• Outputs built-up probability raster + binary mask
• 3-fold cross-validation, F1, precision, recall reporting

🗺️ Optional OSM Building Integration

• Augment training with OpenStreetMap building footprints
• Strengthens urban class separation

---

🛰️ Spectral Indices Used

Index	Measures	Formula
NDVI	Vegetation density	(NIR - Red) / (NIR + Red)
NDBI	Built-up surfaces	(SWIR - NIR) / (SWIR + NIR)
BSI	Bare soil	(SWIR + Red) - (NIR + Blue) / ...
NDWI	Water bodies	(Green - NIR) / (Green + NIR)
MNDWI	Modified water	(Green - SWIR) / (Green + SWIR)

---

🗂️ Project Structure

sentinel2_builtup_pipeline/
│
├── scripts/
│   ├── download_s2_pc_by_tile.py      # Sentinel-2 downloader via STAC
│   ├── mean_indices.py                # Spectral index compositing
│   └── train_and_predict_builtup.py  # ML training & prediction
│
├── data/
│   ├── aoi/           # Area of interest shapefile
│   ├── training/      # Training polygons or points
│   ├── osm/           # Optional OSM building footprints
│   └── sentinel/      # Downloaded satellite tiles (auto-filled)
│
├── output/
│   ├── models/        # Saved Random Forest model (.joblib)
│   ├── prediction_tiles/  # Per-tile output rasters
│   └── logs/
│
├── run.ipynb          # Interactive notebook workflow
├── environment.yml
└── requirements.txt

---

⚙️ Setup

# Create environment
conda create --name s2builtup python=3.10
conda activate s2builtup
conda install jupyter nbconvert
conda install --file requirements.txt -c conda-forge

---

▶️ Usage

Step 1 — Download Sentinel-2 tiles

python scripts/download_s2_pc_by_tile.py \
  --outdir data/sentinel \
  --aoi data/aoi/your_aoi.shp \
  --year 2024 \
  --cloud 5 \
  --max-workers 6

Step 2 — Compute spectral index composites

python scripts/mean_indices.py

Step 3 — Train classifier and generate predictions

python scripts/train_and_predict_builtup.py

---

📤 Outputs

File	Description
*_MEAN.tif	Per-pixel mean spectral index composite
*_MEDIAN.tif	Per-pixel median spectral index composite
*_BUILTUP_PROB.tif	Built-up probability (0–1)
*_BUILTUP_MASK.tif	Binary classification (built-up / non built-up)
builtup_rf.joblib	Saved Random Forest model
training_summary.csv	Accuracy metrics per run

---

📈 Accuracy Metrics

The pipeline reports per-run:

• F1 Score
• Precision & Recall
• Confusion Matrix
• 3-fold Cross-Validation scores

---

🗺️ Roadmap

• UNet / DeepLab deep learning segmentation module
• Time-series built-up change detection
• Zonal statistics aggregation to admin boundaries
• Web map visualization with Leaflet or Kepler.gl

---

👤 Author

Athithiyan M R — Geospatial Data Scientist | Remote Sensing | Climate Analytics

---

🙏 Acknowledgements

• ESA Sentinel-2 Mission
• Microsoft Planetary Computer & STAC API
• OpenStreetMap contributors

---

📜 License

MIT License © 2026 Athithiyan M R