Improving the Retrieval Accuracy of Spaceborne GNSS-R Ocean Surface Altimetry Based on a Novel Attention Mechanism Model

Measurements from global navigation satellite system reflectometry (GNSS-R) have proven effective for applications in sea surface height (SSH). Deep learning-based methods for spaceborne GNSS-R SSH retrieval can effectively leverage physical quantities related to SSH to establish a mapping relationship between multiple observables and SSH. However, current deep learning approaches for SSH retrieval suffer from issues such as low information utilization efficiency and susceptibility to being misled by marginal features. To address these challenges, this article proposes the novel attention mechanism-based feature-enhanced residual multimodal (FERM) model for retrieving SSH based on the GNSS-R. First, the FERM model is constructed by combining a feature enhancement residual (FER) network, which is based on an attention mechanism, with a fully connected neural network (FCNN). The FER network employs two distinct sub-blocks—channel and spatial—to differentiate and enhance the effective features extracted by the neural network in both the channel and spatial dimensions of the input data. This enhances the features by improving the utilization efficiency of the input data information and assigning attention weights to the features, thereby improving the input quality and subsequently enhancing the SSH retrieval accuracy of the FCNN. Second, the DTU18 validation model, composed of the DTU18 Mean Sea Surface (DTU18 MSS) model and the TPXO8 global tidal model, is employed to verify the SSH inversion capability of models. Through the analysis of global mid- and low-latitude data from 2020 and the first three months of 2021, it was found that the mean absolute error (MAE) values of the original residual model and the proposed FERM model are 3.84 and 2.56 m, respectively, representing a reduction of approximately 33.33%; their root mean square error (RMSE) values are 5.46 and 3.74 m, respectively, showing a decrease of 31.50%. Third, the waveform fitting (FIT) retracking algorithm is implemented to retrieve SSH, and the results are compared with those obtained using the FERM method. According to MAE and RMSE metrics, the FERM method demonstrates higher accuracy than the FIT retracking method, with reductions in MAE and RMSE of approximately 51.42% and 43.24%, respectively. The FERM method provides a new theoretical and methodological reference for future SSH retrieval using high spatial and temporal resolution data from GNSS-R altimetry satellites.