PLGF-NET: A Physics-Guided Local-Global fusion network for GNSS-R high wind speed retrieval

Accurate global sea surface wind speed retrieval under high-wind conditions (> 15 m/s) remains a persistent bottleneck for spaceborne Global Navigation Satellite System Reflectometry (GNSS-R). Conventional data-driven models often experience severe performance degradation in these regimes due to inherent physical limitations, including long-tail data distributions, low signal-to-noise ratios (SNR), and scattering signal saturation. To transcend these limitations, this study proposes the Physics-Guided Local-Global Fusion Network (PLGF-NET), establishing a novel physics-data dual-driven paradigm. At the data level, a Comprehensive Difficulty Score (CDS) is introduced to quantify the intrinsic physical complexity of Delay-Doppler Maps (DDMs). This metric drives a difficulty-aware hierarchical sampling strategy that rectifies long-tail biases and dynamically mines the scarce, high-wind scattering signatures. Within the network, a Physics-Guided Spatial Attention (PGSA) mechanism translates specular reflection theory into a structural inductive bias. This enforces a rigorous feature purification process, locking the receptive field onto the effective scattering window while suppressing peripheral thermal noise. Furthermore, geophysical consistency is guaranteed through a selective regularization mechanism governed by Sample Prediction Difficulty (SPD). Rather than applying uniform constraints, this approach adaptively enforces Geophysical Model Function (GMF) constraints exclusively on high-uncertainty samples, providing a robust physical anchor against signal saturation. Comprehensive evaluations demonstrate the exceptional robustness of PLGF-NET. The framework achieves a state-of-the-art overall root mean square error (RMSE) of 1.38 m/s and drastically compresses the high-wind RMSE to 2.40 m/s, effectively correcting systematic negative biases. Ultimately, this scalable and physically interpretable architecture provides a highly reliable Earth observation tool for high-wind marine environment monitoring.