SimCLR Remote Sensing Adaptation

This repository contains a PyTorch implementation of SimCLR (Simple Framework for Contrastive Learning of Visual Representations) adapted for Remote Sensing tasks, specifically Land Use Classification on the UC Merced dataset.

Project Structure

SimCLR-RemoteSensing/
├── app.py                 # Streamlit Production Demo
├── data/                  # Dataset storage
├── notebooks/             # Jupyter Notebooks (Colab)
├── scripts/
│   ├── train.py           # Training/Fine-tuning script
│   └── visualize.py       # t-SNE Visualization script
├── src/
│   ├── model.py           # SimCLR Model Definition
│   └── __init__.py
└── requirements.txt       # Dependencies

Setup

1. Install Dependencies:

   pip install -r requirements.txt

2. Download Pretrained Weights: Ensure simclr_model_RN101.pth (pretrained on EuroSAT) is in the root directory or scripts/ folder as needed.

Usage

1. Training / Fine-Tuning

To fine-tune the ResNet-101 backbone on the UC Merced dataset:

python scripts/train.py

This will:

• Download the UC Merced dataset automatically.
• Load the pretrained SimCLR weights.
• Fine-tune the model.
• Save the best model as finetuned_simclr_ucmerced.pth (or similar).

2. Visualization

To generate a 3D t-SNE animation of the embeddings:

python scripts/visualize.py

3. Production Demo

To run the interactive web interface:

streamlit run app.py

Upload an aerial image to get a classification prediction.

Deployment

1. GitHub Setup

This repository includes a .gitignore to exclude large model files and data.

1. Initialize a git repo: git init
2. Commit your code: git add . && git commit -m "Initial commit"
3. Push to GitHub.

2. Hosting the Model

Since the model (.pth) is too large for GitHub (>100MB):

1. Upload your finetuned_simclr_ucmerced.pth to a cloud provider (Google Drive, Dropbox, Hugging Face).
2. Get a direct download link.
3. Update MODEL_URL in app.py with this link.

3. Streamlit Cloud

1. Go to share.streamlit.io.
2. Connect your GitHub repository.
3. Deploy! The app will automatically download the model from your link on startup.

Model Details

• Backbone: ResNet-101
• Pretraining: SimCLR on EuroSAT (Sentinel-2)
• Downstream: Fine-tuned on UC Merced (Aerial Imagery)