Benchmark-Dataset-and-Deep-Learning-Method-for-Global-Tropical-Cyclone-Forecasting

This repository introduces the contents of our research work, which consists of two main parts:

1. A benchmark dataset for global tropical cyclone (TC) research – TropiCycloneNet Dataset
2. A deep learning-based method for predicting TC track and intensity – TropiCycloneNet Method

Both parts are released to foster open research and encourage reproducible results in the field of tropical cyclone forecasting.

---

🌪️ TropiCycloneNet-Dataset

🔍 Overview

The TropiCycloneNet Dataset (TCN_D) is a comprehensive dataset designed for studying tropical cyclones worldwide. It spans from 1950 to 2023 and includes:

• Data_1d: Tabular records of cyclone center location, pressure, wind speed, etc.
• Data_3d: Gridded meteorological fields (e.g., geopotential height, wind components, SST) around the cyclone center.
• Env-Data: Structured environmental features that support deep learning applications.

It enables the analysis and modeling of both spatial and temporal cyclone behavior and provides a foundation for building data-driven forecasting systems.

📥 Download

We offer two download options:

• Full Dataset (1950–2023, ~25.7GB)
• Test Subset (2017–2023, ~3.34GB)

📊 Visualization

We provide easy-to-use scripts for loading and visualizing each part of the dataset (read_TCND.py). You can generate visualizations of Data_1d, Data_3d, and Env-Data with a single command.

Example:

python read_TCND.py dataset_path Haiyan 2001101418 WP train

This will produce three images showing the meteorological and environmental context of Typhoon Haiyan at a specific timestamp. 📎 Features

Resolution: 0.25°, every 6 hours

Ocean Regions: WP, EP, ATL, NI, SI, SP

Pressure levels: 200, 500, 850, 925 hPa

Additional features: SST, movement velocity, intensity change, subtropical high data

📘 Documentation

See full documentation and usage guide in the GitHub repository:

👉 TropiCycloneNet-Dataset

🧠 TropiCycloneNet-Model

🔧 Introduction

TropiCycloneNet (TCN_M) is our proposed deep learning model for predicting TC track and intensity using multi-modal data (e.g., environmental features, 1D/3D meteorological data).

This repository contains:

Preprocessed subset of the TCN_D dataset

Test code for inference and visualization

Forecast results rendered on Himawari-8 satellite imagery

🔍 Red circles: Ground truth tracks
🌟 Green stars: Our model’s best prediction
🟩 Green area: Multiple trajectory predictions by our model
🟥 Red area: MMSTN’s prediction range

⚙️ Requirements

Python 3.8.5

PyTorch 1.11.0+

OpenCV, NumPy, Matplotlib

Training code and model weights will be released soon. See full documentation and usage guide in the GitHub repository:

👉 TropiCycloneNet-Model

Citing TropiCycloneNet

@article{Huang2025,
  author    = {Huang, Cheng and Mu, Pan and Zhang, Jinglin and Chan, Sixian and Zhang, Shiqi and Yan, Hanting and Chen, Shengyong and Bai, Cong},
  title     = {Benchmark dataset and deep learning method for global tropical cyclone forecasting},
  journal   = {Nature Communications},
  volume    = {16},
  number    = {1},
  pages     = {5923},
  year      = {2025},
  publisher = {Nature Publishing Group},
  doi       = {10.1038/s41467-025-61087-4},
  url       = {https://doi.org/10.1038/s41467-025-61087-4},
  issn      = {2041-1723}
}