🌍 EuroSAT Land Use Classification with CNN

A deep learning project that classifies satellite images into 10 land use categories using a custom-built Convolutional Neural Network (CNN). Trained and evaluated on the EuroSAT RGB dataset.

---

📁 Dataset

• Source: EuroSAT (RGB only)
• Size: 27,000 images
• Classes:
  • AnnualCrop
  • Forest
  • HerbaceousVegetation
  • Highway
  • Industrial
  • Pasture
  • PermanentCrop
  • Residential
  • River
  • SeaLake
• Image Size: 64×64 pixels

---

🧠 Model Architecture

A lightweight CNN trained from scratch:

[Conv2d(3→32) → ReLU → MaxPool] →
[Conv2d(32→64) → ReLU → MaxPool] →
[Conv2d(64→128) → ReLU → MaxPool] →
Flatten → FC(8192→256) → ReLU → Dropout(0.5) → FC(256→10)

---

🛠️ Training Configuration

Setting	Value
Optimizer	Adam
Learning Rate	0.001
Batch Size	32
Epochs	15
Dropout	0.3
Hardware	Tesla T4 (Colab)

---

📊 Results

Metric	Score
Accuracy	91%
Macro F1	~0.90
External Img	Supported ✅
Explainable	Grad-CAM ✅

Classification Report (excerpt):

Class	Precision	Recall	F1-score
Forest	0.97	0.99	0.98
Residential	0.98	0.97	0.97
PermanentCrop	0.76	0.91	0.83
...	...	...	...

---

🔍 Features

• ✅ CNN trained from scratch
• 📦 Preprocessing and normalization via torchvision
• 🖼️ External image inference support
• 🎯 Grad-CAM visualization for explainability
• 💾 Model saving/loading in .pth format

---

🚀 Getting Started

Prerequisites

• Python 3.7+
• PyTorch
• torchvision
• matplotlib
• numpy
• scikit-learn

Installation & Usage

git clone https://github.com/your-username/landuse-cnn-eurosat.git
cd landuse-cnn-eurosat
pip install -r requirements.txt

Open eurosat_cnn.ipynb in Google Colab or Jupyter and run all cells to:

• Download the EuroSAT dataset
• Train the CNN model
• Evaluate on test data
• Run inference on external satellite images
• Visualize Grad-CAM heatmaps

---

📁 File Structure

landuse-cnn-eurosat/
│
├── eurosat_cnn.ipynb         # Main notebook  
├── requirements.txt          # Python dependencies  
├── assets/  
│   ├── sample_predictions.png  
│   └── gradcam_visuals.png  
└── README.md

---

💡 Future Improvements

• Experiment with deeper architectures like ResNet18 or Vision Transformers (ViT)
• Deploy as a web app with Streamlit or Gradio
• Convert model to ONNX or TFLite for mobile deployment
• Test on higher-resolution or multi-spectral satellite datasets
• Add more explainability methods beyond Grad-CAM

---

🧑‍💻 Author

Sanjay Sakhinala
AI/ML Enthusiast | Open to AI Roles
📫 LinkedIn • 📧 sanjaysakhinala@gmail.com

---

📚 References

• EuroSAT Dataset Paper
• PyTorch Documentation
• Grad-CAM: Selvaraju et al., 2017

---

⭐ If you found this project useful, please give it a star on GitHub!