🛰️ Land Use and Land Cover (LULC) Classification with Vision Transformers

This project revisits and rebuilds a previous Land Use and Land Cover (LULC) classification task on the EuroSAT dataset, using Vision Transformers (ViT) and modern PyTorch best practices.

It covers everything from dataset loading, preprocessing, training with transfer learning, to saving and using the model for inference — all in a clean, modular structure powered by Pipenv and reproducible configs.

📜 Read my blog post here

Part 1 : How I Rebuilt My First LULC Project with a Vision Transformer and Clean Code

Part 2 : From Model to Map: Automating Land Use Classification with Vision Transformers and Google Earth Engine

📂 Project Structure

├── config.yaml           # All training and data parameters
├── data/                 # Folder for the EuroSAT dataset (structure preserved, content ignored by Git)
├── datafactory.py        # Custom Dataset and transforms
├── engine.py             # Training and evaluation loops
├── LULC\_3BIMG\_VIT.ipynb  # Main training notebook
├── outputs/              # Folder to save models, plots, reports (content ignored by Git)
├── utils.py              # Utilities (plotting, seeding, inference tools)
├── Pipfile / Pipfile.lock # Reproducible environment with Pipenv
└── .gitignore

---

🚀 Getting Started

1. Clone the Repository

git clone https://github.com/LiganiumInc/LULC-ViT-EuroSAT3B.git
cd LULC-ViT-EuroSAT3B

2. Set up the Environment

# Make sure Pipenv is installed
pip install pipenv

# Create and activate virtual environment
export PIPENV_VENV_IN_PROJECT=1
pipenv install
pipenv shell

3. Download the EuroSAT RGB Dataset

• Link: https://github.com/phelber/EuroSAT
• Extract into the following structure:

data/
└── EuroSAT/
    └── 2750/
        ├── AnnualCrop/
        ├── Forest/
        └── ...

---

⚙️ Configuration

All settings can be found in config.yaml. Example:

data_dir: "./data/EuroSAT/2750/"
batch_size: 32
num_epochs: 10
lr: 0.001
weight_decay: 0.05
percentage_per_class: 0.3

---

🧠 Model: Vision Transformer (ViT)

We use the pretrained vit_b_16 from torchvision.models, and fine-tune it to classify 10 LULC categories.

📊 Outputs

After training, you'll find:

• best_model.pth: The saved fine-tuned ViT model
• loss_acc_curves.png: Visual training metrics
• train_report.txt: Final performance logs

---

🧑‍💻 Author

Bernardin Ligan PhD Student in AI & Remote Sensing

🌍 Passionate about geospatial ML, open science, and SDGs

---

📜 License

This project is open-source and available under the MIT License.