Hu et al. (2025) CycloneWind: A Dynamics-Constrained Deep Learning Model for Tropical Cyclone Wind Field Downscaling Using Satellite Observations

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Remote Sensing
• Year: 2025
• Date: 2025-09-10
• Authors: Yuxiang Hu, Ke Deng, Qingguo Su, Di Zhang, Xinjie Shi, Kaijun Ren
• DOI: 10.3390/rs17183134

Research Groups

Not specified in the provided text.

Short Summary

This study introduces CycloneWind, a novel deep learning framework designed to downscale tropical cyclone surface wind fields, achieving an 8-fold spatial resolution increase. It significantly improves the accuracy of wind component reconstruction and key dynamical metrics by integrating a high-quality dataset and a dynamically constrained Transformer-based architecture.

Objective

• To propose a novel deep learning framework, CycloneWind, for downscaling tropical cyclone surface wind fields, addressing the scarcity of high-quality datasets and the limitations of general deep learning frameworks in capturing TC dynamic characteristics.

Study Configuration

• Spatial Scale: Fine-scale within tropical cyclone regions, achieving an 8-fold spatial resolution increase.
• Temporal Scale: Not specified in the provided text.

Methodology and Data

• Models used: CycloneWind (a dynamically constrained Transformer-based architecture), Adaptive Dynamics-Guided Block (ADGB), Filtering Transformer Layers (FTLs).
• Data sources: Cyclobs satellite observations, ERA5 reanalysis data. Auxiliary variables include low cloud cover, surface pressure, and top-of-atmosphere incident solar radiation.

Main Results

• CycloneWind successfully achieves an 8-fold spatial resolution reconstruction in tropical cyclone regions.
• Compared to the best-performing baseline model, CycloneWind reduces the Root Mean Square Error (RMSE) for the U wind component by 9.6% and for the V wind component by 4.9%.
• Substantial improvements are achieved in key dynamical metrics: divergence difference (23.0%), vorticity difference (22.6%), and direction cosine dissimilarity (20.5%).

Contributions

• Development of CycloneWind, a novel deep learning framework specifically designed for downscaling extreme tropical cyclone wind fields, addressing a gap in existing general wind field downscaling methods.
• Construction of a high-quality dataset for tropical cyclone wind fields by integrating Cyclobs satellite observations with ERA5 reanalysis data and auxiliary variables.
• Introduction of a dynamically constrained Transformer-based architecture incorporating three wind field dynamical operators and a wind dynamics-constrained loss function to enforce consistency in wind divergence and vorticity.
• Design of an Adaptive Dynamics-Guided Block (ADGB) to explicitly encode tropical cyclone rotational dynamics and Filtering Transformer Layers (FTLs) for modeling small-scale wind field details.

Funding

Not specified in the provided text.

Citation

@article{Hu2025CycloneWind,
  author = {Hu, Yuxiang and Deng, Ke and Su, Qingguo and Zhang, Di and Shi, Xinjie and Ren, Kaijun},
  title = {CycloneWind: A Dynamics-Constrained Deep Learning Model for Tropical Cyclone Wind Field Downscaling Using Satellite Observations},
  journal = {Remote Sensing},
  year = {2025},
  doi = {10.3390/rs17183134},
  url = {https://doi.org/10.3390/rs17183134}
}