Sentinel-2 Land Cover Classification - Hanoi, Vietnam

A pixel-based land cover classification project using Sentinel-2 satellite imagery to map urban areas in central Hanoi. Three machine learning models were compared: SVM, Random Forest, and 1D CNN.

🎯 Quick Results

Model	Accuracy	Performance
1D CNN	95.26%	⭐ Best
Random Forest	91.75%	✅ Good
SVM	89.84%	✅ Good

Winner: 1D CNN - Best at handling complex urban land covers

🗺️ What We Classified

• Water 💧 - Lakes, rivers
• Built Area 🏢 - Buildings, roads
• Vegetation 🌳 - Parks, trees
• Others 🔶 - Bare land, construction sites

📊 Detailed Results

Model Performance by Class

Class	SVM	Random Forest	1D CNN
Water	0.99	0.98	0.99
Built Area	0.94	0.95	0.98
Vegetation	0.59	0.63	0.77
Others	0.76	0.75	0.88

Key Findings

• CNN excels at mixed/complex areas 🧠
• All models good for water detection 💧
• Vegetation vs bare land remains challenging 🌱

Confusion Matrices

🚀 Applications

• Urban Planning - Monitor city growth
• Environmental Tracking - Watch green space changes
• Smart City - Support data-driven decisions
• Disaster Management - Risk assessment and planning

💻 Tech Stack

• Data: Sentinel-2 satellite imagery (Feb 2020)
• Area: Central Hanoi (6 districts)
• Tools: Python, TensorFlow, Scikit-learn, QGIS
• Best Model: 1D Convolutional Neural Network

📁 Repository Structure

├── assets/                   
└── Notebooks/                     
    ├── rf_svm.ipynb
    └── CNN.ipynb

---

📖 Want More Details?

This README shows the key results. For complete methodology, technical details, literature review, and full analysis, please read the full project report: Project2_Report.docx

The report includes:

• Detailed methodology for each model
• Complete literature review
• Step-by-step data preprocessing
• In-depth analysis and discussion
• Future work suggestions
• Full references

---

Student: Vu Duc Thang - Hanoi University of Science and Technology
Supervisor: PhD Tran Nguyen Ngoc