A Physics-Constrained Network for Daily Gap-Free Gridded Wind Speed Product Considering High Wind Speeds From CYGNSS

To address key challenges including Cyclone GNSS (CYGNSS) sparse point observations, insufficient coverage in coastal regions, and systematic underestimation of high wind speeds, this study proposes a PIDW-E2E deep learning framework that integrates U-Net and Graph Convolutional Network (GCN), achieving for the first time end-to-end generation of daily gap-free gridded wind speed products based on CYGNSS observations. The model effectively extracts spatially correlated features of normalized bistatic radar cross section (NBRCS), addressing the coupling challenge between sparse point observations and complex wind field structures while simultaneously accomplishing coastal wind speed retrieval and data gap filling in unobserved areas. By further incorporating a physics-constrained loss function and a high wind speed label enhancement strategy, the model significantly improves retrieval performance in high wind speed regions. The results show that, when evaluated against daily ERA5 wind speeds, the gridded product generated by the model achieves an overall RMSE of 0.92m/s. Compared to CYGNSS L2 FDS data, the RMSE is further reduced by 0.67m/s. In coastal wind speed reconstruction tasks, PIDW-E2E demonstrates significant advantages, producing spatially complete coverage in coastal regions (RMSE ∼1.10m/s, Bias ∼-0.01m/s) with accuracy comparable to that of open ocean areas. For tropical cyclone monitoring, the model effectively mitigates systematic underestimation in high wind speed regions. Under conditions where the maximum retrievable wind speed reaches 45m/s, the model achieves an RMSE of 3.47m/s and a Bias of -1.39m/s relative to the daily HWRF wind speed reference data. This study provides a new approach for CYGNSS wind speed retrieval that combines gap-free coverage capability with enhanced responsiveness in high wind speed regions.