A CNN-BiLSTM-Attention model incorporating wind speed characteristics for enhanced significant wave height retrieval from CYGNSS data by spaceborne GNSS-R technology

Significant wave height (SWH) represents one of the most prevalent and crucial statistical parameters characterizing ocean waves, with extensive applications in oceanography, meteorology, navigation, and ocean engineering. Wind speed (WS) is a primary driver of SWH, yet their spatiotemporal relationship has been insufficiently explored. Existing multimodal CNN-based SWH retrieval methods mainly emphasize spatial features while overlooking temporal dynamics. In addition, auxiliary parameters are typically processed by fully connected networks (FCNs), which limits retrieval accuracy. To address these gaps, this study investigates the spatiotemporal relationship between WS and SWH. Subsequently, Bidirectional Long Short-Term Memory (BiLSTM) is employed to capture temporal features, complemented by an attention mechanism to augment the discriminative power for auxiliary parameters. An integrated CNN-BiLSTM-Attention SWH retrieval model incorporating WS information (termed CNN-BiLSTM-Attention-W) is then developed. Experimental results reveal a strong positive correlation between WS and SWH. Validation against the ERA5 SWH data product shows that the CNN-W model integrating WS information demonstrates superior accuracy compared to the CNN model. The CNN-W model yields a root mean square error (RMSE) of 0.492 m, while the CNN-BiLSTM-Attention-W model achieves a lower RMSE of 0.441 m. Comparison with buoy measurements shows that the RMSE values for the CNN-W model range from 0.392 to 0.643 m, and for the CNN-BiLSTM-Attention-W model range from 0.313 to 0.621 m. This provides an effective reference for addressing the issues of insufficient utilization of spatiotemporal information and weak feature extraction for auxiliary parameters in existing SWH retrieval methods.