A curriculum-inspired Hierarchical model for GNSS-R sea surface wind speed retrieval using coarse-to-fine learning

Abstract Accurate retrieval of sea surface wind speed from global navigation satellite system reflectometry is crucial for enhancing ocean monitoring and weather prediction capabilities. However, the accuracy of such retrievals is often limited by the wide dynamic range of wind speeds, nonlinear scattering effects, and data imbalance across different wind regimes. To overcome these issues, this paper proposes a curriculum-inspired coarse-to-fine learning framework, named as ‘C2FWindNet’, which employs a cascaded spatial–temporal fusion architecture. In the framework, a convolutional neural network (CNN) captures spatial features and categorizes wind speeds into coarse-grained intervals, while a Transformer refines temporal dependencies within each interval. A residual correction module is incorporated to mitigate nonlinear biases, and a dual-curved sigmoid loss function is introduced to enhance robustness under extreme wind conditions. Extensive evaluation on the Cyclone GNSS dataset demonstrate the superiority of the proposed approach. It achieves a root-mean-square error of 1.201 m s −1 , outperforming the best state-of-the-art model by 9.1%. Under high-wind conditions, the dual-curved sigmoid weighting further enhances the retrieval performance, yielding a 24.62% improvement over the best baseline. Additional experiments confirm the framework’s strong generalization across regions and time periods, while the flexible design of the dual-curved sigmoid loss enables targeted optimization for different wind speed intervals, improving its adaptability to diverse oceanic conditions and operational scenarios.