Zhang et al. (2025) Physics-Informed Transformer Networks for Interpretable GNSS-R Wind Speed Retrieval

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

Identification

• Journal: Remote Sensing
• Year: 2025
• Date: 2025-11-24
• Authors: Zao Zhang, Jingru Xu, Guifei Jing, Dongkai Yang, Yue Zhang
• DOI: 10.3390/rs17233805

Research Groups

Not provided in the paper text.

Short Summary

This study develops a novel Transformer-Graph Neural Network (GNN) model for Global Navigation Satellite System Reflectometry (GNSS-R) wind speed retrieval, addressing limitations in interpretability and accuracy during high wind conditions. The model significantly reduces wind speed RMSE and provides physically grounded interpretations of spatiotemporal influence propagation.

Objective

• To overcome the lack of interpretability and accuracy degradation in high wind conditions of existing GNSS-R wind retrieval models by leveraging a mathematical equivalence between Transformers and graph neural networks to provide a physically grounded interpretation of self-attention as spatiotemporal influence propagation in GNSS-R data.

Study Configuration

• Spatial Scale: Local (25–100 km), mesoscale (100 km–500 km), and synoptic (>500 km) circulation patterns.
• Temporal Scale: Data from 2023–2024.

Methodology and Data

• Models used: Transformer–GNN model, which treats each GNSS-R footprint as a graph node and interprets multi-head self-attention weights as localized spatiotemporal interactions.
• Data sources:
  • Level 1 Version 3.2 GNSS-R data (2023–2024) from four Asian sea regions.
  • ERA5 reanalysis 10 m equivalent-neutral wind fields (primary training reference dataset).
  • Stepped Frequency Microwave Radiometer (SFMR) aircraft observations (independent validation during tropical cyclone events).
  • Moored buoy measurements (independent validation where spatiotemporally coincident data are available).

Main Results

• The Transformer–GNN model achieved an overall wind speed Root Mean Square Error (RMSE) reduction of 32%, from 1.98 m s⁻¹ to 1.35 m s⁻¹.
• Substantial accuracy gains were observed in high-wind regimes (winds >25 m s⁻¹), with an RMSE of 3.2 m s⁻¹.
• Interpretability analysis using SHAP revealed condition-dependent feature attributions and suggested coupling mechanisms between ocean surface currents and wind fields.

Contributions

• Advances both predictive accuracy and interpretability in GNSS-R wind retrieval, particularly in high-wind conditions.
• Introduces a novel approach by leveraging the mathematical equivalence between Transformers and GNNs to provide a physically grounded interpretation of self-attention as spatiotemporal influence propagation in GNSS-R data.
• Offers an operationally viable framework for interpretable, physics-aware Earth system AI applications.

Funding

Not provided in the paper text.

Citation

@article{Zhang2025PhysicsInformed,
  author = {Zhang, Zao and Xu, Jingru and Jing, Guifei and Yang, Dongkai and Zhang, Yue},
  title = {Physics-Informed Transformer Networks for Interpretable GNSS-R Wind Speed Retrieval},
  journal = {Remote Sensing},
  year = {2025},
  doi = {10.3390/rs17233805},
  url = {https://doi.org/10.3390/rs17233805}
}