An Attention Entropy-Driven Fusion Model for Sea Surface Wind Speed Retrieval

Abstract

Sea surface wind speed is a critical indicator of ocean–atmosphere interactions and climate change prediction. Global navigation satellite system reflectometry (GNSS-R) has been widely applied for wind speed retrieval, yet its accuracy is constrained by two key challenges: anomalous or contaminated satellite observations, and physical sensitivity saturation of GNSS observables at high wind speeds. Existing quality control strategies mitigate the first issue but reduce sample utilization and model generalization. Artificial intelligence-based meteorological foundation models offer an alternative, but suffer from limited spatiotemporal resolution and systematic underestimation bias at high wind speeds. This article proposes an attention entropy-driven fusion model that fuses cyclone GNSS satellite observations with Pangu-Weather predictions to improve the precision and spatiotemporal resolution of sea surface wind speed retrieval. To address anomalous observations, the proposed model introduces attention entropy features derived from bidirectional cross-attention to assess observation validity at the sample level, enabling discrimination between reliable and contaminated data. To overcome physical sensitivity saturation, two differentiated expert models are constructed: an observation fusion expert that leverages satellite observables to correct prediction errors in low wind speed regimes and a statistical correction expert that compensates for systematic underestimation bias in high wind speed regimes. A meta-strategy router based on entropy features dynamically fuses the two expert outputs at the sample level. Validation against ERA5 reference wind speeds yielded an RMSE of 1.27 m/s, a 17% improvement over the Pangu-Weather baseline, and accuracy improvements of up to 37% over existing physical and machine learning models, demonstrating strong robustness across global domains.