🌊 SEA-ViT: Sea Surface Currents Forecasting with Vision Transformers and GRUs

"SEA-ViT: Forecasting Sea Surface Currents Using a Vision Transformer and GRU-Based Spatio-Temporal Covariance Model"
📌 Accepted at IEEE KST 2025 | 🔬 Developed by Teerapong Panboonyuen (Kao)

🚀 Highlights

✅ Vision Transformer + BiGRU for capturing spatio-temporal ocean dynamics.
🌐 Forecasts U/V currents from HF-radar & ENSO time series.
📦 Modular PyTorch Code for fast deployment and experimentation.
🧠 30-Year Dataset Ready — built for long-term environmental forecasting.

🗂️ Project Structure

sea-vit/
├── src/
│   ├── model.py         # SEA-ViT Model: ViT + BiGRU
│   ├── train.py         # Training Pipeline
│   ├── inference.py     # Run Inference on Radar Sequences
│   ├── metrics.py       # Evaluation (RMSE, MAE, Corr)
│   └── utils.py         # Dataset I/O, Preprocessing, Logging
├── Dockerfile           # Reproducible Environment
├── requirements.txt     # Python Dependencies
├── README.md            # Project Overview (this file)
└── data/                # Example .npz radar data (optional)

📥 Installation

Install dependencies via pip:

git clone https://github.com/kaopanboonyuen/GISTDA-SeaViT.git
cd GISTDA-SeaViT
pip install -r requirements.txt

Or build the full environment via Docker:

docker build -t sea-vit .

💻 How to Run the Code

🔧 1. Train SEA-ViT on your dataset

Assuming your .npz dataset includes radar sequences + targets:

python src/train.py \
  --data-path ./data/train.npz \
  --batch-size 8 \
  --epochs 50 \
  --lr 1e-4

Input shape: [N, T, 2, H, W] for sequences
Target shape: [N, forecast_steps, 2] for (U, V)

🧠 2. Inference on a New Sequence

python src/inference.py \
  --model-path ./checkpoints/sea-vit.pth \
  --input-path ./data/sample_input.npz \
  --output-path ./output/forecast.npy

📏 3. Evaluate Metrics (RMSE, MAE, Correlation)

python src/metrics.py \
  --predictions ./output/forecast.npy \
  --ground-truth ./data/sample_ground_truth.npy

📈 Results Summary

Metric	Value (Example)
RMSE	0.087
MAE	0.065
Corr	0.91

SEA-ViT demonstrates high accuracy across diverse oceanic conditions, outperforming CNN-GRU and ConvLSTM baselines.

📚 Learn More

🔬 Citation

@inproceedings{panboonyuen2025sea,
  title={SEA-ViT: Forecasting Sea Surface Currents Using a Vision Transformer and GRU-Based Spatio-Temporal Covariance Model},
  author={Panboonyuen, Teerapong},
  booktitle={2025 17th International Conference on Knowledge and Smart Technology (KST)},
  pages={1--6},
  year={2025},
  organization={IEEE}
}

📫 Contact

Teerapong Panboonyuen (Kao)
Postdoctoral Researcher, Chulalongkorn University
Senior Research Scientist, MARSAIL
📧 teerapong.panboonyuen@gmail.com

🤝 Contributing

Pull requests welcome! For feature ideas or bug fixes, feel free to open an issue.

“SEA-ViT doesn’t just see the ocean — it foresees it.” 🌊📡