An et al. (2026) Physics-Aware Hybrid CNN–Transformer Network for GNSS-R Sea Surface Wind Speed Estimation

Identification

• Journal: Remote Sensing
• Year: 2026
• Date: 2026-03-31
• Authors: Baiwei An, Weiwei Qin, Weijie Kang, L Zhang, Hao Chi
• DOI: 10.3390/rs18071053

Research Groups

• Rocket Force University of Engineering, Xi’an 710025, China
• School of Electronic and Electrical Engineering, Baoji University of Arts and Sciences, Baoji 721016, China

Short Summary

This study proposes a Physics-Aware Hybrid CNN–Transformer Network (PA-HCTN) for accurate sea surface wind speed estimation from GNSS-R data, which integrates local feature extraction, global context modeling, and dynamic fusion of physical parameters. The model achieves a global Root Mean Square Error (RMSE) of 1.35 m/s and significantly mitigates high-wind-speed underestimation bias by incorporating a Geophysical Model Function (GMF)-constrained loss function.

Objective

• To address the challenges of accurate sea surface wind speed retrieval from Global Navigation Satellite System Reflectometry (GNSS-R) data, specifically the complex scattering mechanisms, nonlinear coupling between Delay–Doppler Maps (DDMs) and observation geometries, and the persistent underestimation bias in high-wind-speed ranges (>15 m/s).
• To develop a novel Physics-Aware Hybrid CNN–Transformer Network (PA-HCTN) that synergizes a Convolutional Neural Network (CNN) for local DDM feature extraction, a Transformer encoder for global context modeling, and a cross-attention module for dynamic fusion of auxiliary physical parameters, enhanced by a GMF-constrained loss function to ensure physical consistency.

Study Configuration

• Spatial Scale: Global ocean, focusing on the CYGNSS latitude range (approximately ±38°), which covers tropical cyclone formation and evolution regions. Validation includes specific ocean regions using NDBC buoy data.
• Temporal Scale: CYGNSS and ERA5 data for the entire year of 2024. Training data from 1 January to 30 June 2024, validation data from 1 July to 31 August 2024, and test data from the remaining period. NDBC buoy data for independent validation from 1 September to 31 December 2024.

Methodology and Data

• Models used:
  • Proposed: Physics-Aware Hybrid CNN–Transformer Network (PA-HCTN)
  • Comparative: Minimum Variance Estimator (MVE), MCNN, PA-CNN.
  • Underlying physical model: Geophysical Model Function (GMF) derived from the Zavorotny–Voronovich (Z-V) scattering model.
• Data sources:
  • Satellite: CYGNSS Level 1 Version 3.1 dataset, providing Delay–Doppler Maps (DDMs) and 14 auxiliary parameters (e.g., Normalized Bistatic Radar Cross Section (NBRCS), Leading Edge Slope (LES), Signal-to-Noise Ratio (SNR), incidence angle, ranges, velocities).
  • Reanalysis: European Centre for Medium-Range Weather Forecasts’ fifth-generation reanalysis (ERA5) 10 m sea surface wind speed (zonal and meridional components) with hourly temporal resolution and a 0.25° × 0.25° horizontal grid.
  • Observation: National Data Buoy Center (NDBC) buoy data from four stations (41002, 41049, 42060, 51000) for independent in situ validation, adjusted to a 10 m reference level.

Main Results

• The proposed PA-HCTN achieved a global RMSE of 1.35 m/s and an R2 of 0.75 for sea surface wind speed estimation from CYGNSS data, outperforming existing deep learning benchmarks.
• The model significantly mitigated the high-wind-speed underestimation bias, reducing it to −3.90 m/s for winds > 15 m/s.
• Ablation studies confirmed that ancillary parameters, the GMF-constrained loss function, the CNN module, and especially the cross-attention mechanism, each contribute significantly to the model's performance. The cross-attention mechanism alone reduced RMSE from 1.46 m/s to 1.32 m/s and increased R2 from 0.68 to 0.77 compared to simple concatenation.
• Independent validation against NDBC buoy data demonstrated consistent superior performance across four diverse stations, with RMSE ranging from 1.45 m/s to 1.51 m/s.

Contributions

• Proposed a novel Physics-Aware Hybrid CNN–Transformer Network (PA-HCTN) architecture that effectively combines the local feature extraction capabilities of CNNs with the global context modeling of Transformers for GNSS-R geophysical parameter retrieval.
• Introduced a cross-attention mechanism for dynamic and intelligent fusion of DDM image features with auxiliary physical parameters, enabling deeper interaction and improved performance compared to traditional concatenation methods.
• Integrated a Geophysical Model Function (GMF)-constrained loss function, explicitly enforcing physical consistency and significantly mitigating the persistent underestimation problem in data-sparse high-wind-speed regimes.
• Demonstrated superior accuracy (global RMSE of 1.35 m/s) and robustness, particularly in addressing high-wind-speed underestimation, compared to existing deep learning and traditional GNSS-R wind retrieval methods.
• Provided a novel architecture paradigm and a viable solution for enhancing the accuracy of extreme weather monitoring using spaceborne GNSS-R technology.

Funding

• National Natural Science Foundation of China (NSFC) under grant No. 72401292
• Young Talent Fund of Association for Science and Technology in Shaanxi, China under grant No. 2025019
• Provincial Outstanding Youth Science Fund in Shaanxi, China under grant No. 2025JC-JCQN-077

Citation

@article{An2026PhysicsAware,
  author = {An, Baiwei and Qin, Weiwei and Kang, Weijie and Zhang, L and Chi, Hao},
  title = {Physics-Aware Hybrid CNN–Transformer Network for GNSS-R Sea Surface Wind Speed Estimation},
  journal = {Remote Sensing},
  year = {2026},
  doi = {10.3390/rs18071053},
  url = {https://doi.org/10.3390/rs18071053}
}