Met2Net: A Decoupled Two-Stage Spatio-Temporal Forecasting Model for Complex Meteorological Systems

🏆 ICCV 2025 Accepted Paper · 🛰️ Spatio-Temporal Forecasting · ⏳ Two-Stage Learning

---

🆕 Updates

• [June 26, 2025] 🎉 Met2Net accepted to ICCV 2025! Paper available on arXiv: 2507.17189

---

📂 Datasets & Models Overview

We release both the processed datasets and pretrained model weights on Hugging Face for full reproducibility:

Dataset Name	Variables	Shape (C×H×W)	Seq (Input→Output)	Samples (Train/Test)	Dataset Repo	Model Repo
ERA5-Cropped	T2M, U10, V10, MSL	4×128×128	12 → 12	43,801 / 8,737	📂 HF Dataset	🧠 HF Model
MvMmfnist	3 synthetic channels	3×64×64	10 → 10	10,000 / 10,000	📂 HF Dataset	🧠 HF Model

---

🚀 Getting Started

1. Clone and setup environment

git clone https://github.com/ShremG/Met2Net.git
cd Met2Net
bash create_env.sh

2. Run training

bash run.sh

---

📄 Citation

If you find this project useful in your research, please cite:

@InProceedings{Li_2025_ICCV,
    author    = {Li, Shaohan and Yang, Hao and Chen, Min and Qin, Xiaolin},
    title     = {Met2Net: A Decoupled Two-Stage Spatio-Temporal Forecasting Model for Complex Meteorological Systems},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {5458-5468}
}

---

📬 Contact

For questions, collaborations, or implementation issues, please open a GitHub issue or contact the corresponding author.

---

🙏 Acknowledgements

This work is built upon foundations from OpenSTL.
We thank the open-source community for enabling reproducible weather forecasting research.

---