Sea Surface Height Inversion Model Based on Multimodal Deep Learning for the Fusion of Heterogeneous FY-3E GNSS-R Data

Sea surface height (SSH) is of great significance in oceanography and meteorology. Traditional physical altimetry methods based on delay–Doppler mapping (DDM) are subject to errors that are difficult to correct computationally. The current deep-learning-based SSH inversion techniques primarily relying on single-modal data are unable to fully leverage the rich feature information from global navigation satellite system reflectometry (GNSS-R) remote sensing data, therefore limiting the potential accuracy improvement. This study proposes a physics-informed multimodal deep-learning framework, physical-informed multimodal heterogeneous altimetry network (PIMFA-Net), to fuse heterogeneous spaceborne GNSS-R data to retrieve SSH. The GNSS-R data are acquired from the GNOS II instrument onboard the Fengyun-3E satellite, which can receive reflected signals from both global positioning system (GPS) and BeiDou navigation satellite system (BDS). GNSS-R parameters are used to construct the PIMFA-Net, which includes cropped DDM images, signal parameters, and system parameters in combination with environmental parameters (wind speed and convective rain rate from the European Centre for Medium-Range Weather Forecast, and SSH derived from physical altimetry models). The global sea surface dataset Danmarks Tekniske Universitet 2018 is used as ground truth for model training and evaluation. Data from 1 to 31 July 2022 are used to train PIMFA-Net, while data from August to October 2022 are used to evaluate the general ability of PIMFA-Net. Results demonstrate that the PIMFA-Net not only improves the accuracy and generalization by integrating heterogeneous data sources but also achieves all-weather, all-day, and wide-area SSH inversion with a precision of less than 40 cm for both GPS and BDS signals. This outcome holds significant potential for applications in marine ecological security monitoring and research.