"Adaptive Sea Surface Wind Speed Retrieval for CYGNSS via a Hierarchical Gaussian Mixture of Experts with Anchor-Based Domain Ada"

"This comprehensive dataset provides over 12 million rigorously matched spatiotemporal data points between the Cyclone Global Navigation Satellite System (CYGNSS) Level 2 observables and Soil Moisture Active Passive (SMAP) Level 2B radiometer wind speeds. It is specifically designed to benchmark machine learning algorithms for sea surface wind speed retrieval, evaluate extreme weather (hurricane\/typhoon) tracking capabilities, and develop domain adaptation strategies to counteract inter-annual satellite payload degradation.The dataset is temporally divided into two major components:Historical Training Set (2021\u20132024): A carefully balanced subset of 954,483 samples processed via intelligent pyramid under-sampling. This set ensures a robust representation of extreme high-wind conditions while preventing low-wind variance contamination.Full-Year Independent Validation Set (2025): A massive collection of 11,141,721 highly precise matchups capturing the entire year of 2025. It includes a pre-defined 5% Anchor Set (557,086 points) for transfer learning and a 95% Blind-Test Set (10,584,635 points) for strictly independent evaluation.Key Features Included: Beyond standard empirical observables\u2014Normalized Bistatic Radar Cross Section (NBRCS) and Leading Edge Slope (LES)\u2014this dataset innovatively integrates auxiliary physics variables from ERA5 (Mean Square Slope [MSS] and Significant Wave Height [SWH]). Crucially, it features the pre-calculated Zilitinkevich-Voronovich (Z-V) physical scattering coupling term and surface roughness gradient vectors, enabling researchers to explore physics-informed machine learning architectures.Baseline Performance: The dataset includes the baseline wind speeds from the official CYGNSS Fully Developed Seas (FDS) operational algorithm. For benchmarking, researchers can reference the state-of-the-art Hierarchical Gaussian Mixture Model (H-GMM) trained on this dataset, which achieved an RMSE of 2.8997 m\/s and a near-zero bias of -0.0047 m\/s on the 2025 blind-test set (significantly outperforming the FDS benchmark of 3.0958 m\/s).This dataset offers immense value for researchers focusing on GNSS-R remote sensing, covariate shift\/domain adaptation in aging satellite sensors, and extreme ocean wave modeling."