4DVarNet Schemes for the Reconstruction of the Underwater Surface Duct Using Argo Profiles

The variability of the ocean's superficial layers significantly influences acoustic propagation, crucial for sonar operations. This study introduces and evaluates neural inversion schemes using the 4DVarNet framework, which integrates variational data assimilation with machine learning techniques to enhance the reconstruction of Surface Duct Width (SDW). By combining Sea Surface Temperature (SST) data and Argo profiles with the 4DVarNet framework, our proposed method, particularly the multimodal 4DVarNet-SST, of Fers substantial improvements over traditional approaches like Optimal Interpolation and U-Net in terms of reconstruction accuracy and spatial-temporal resolution. Our findings indicate that the integration of SST data not only enhances spatial resolution but also introduces challenges in balancing the influence of SST, which may overshadow other dynamic ocean features. The study highlights the potential of these frameworks for real-time applications, providing a robust approach to adaptively integrate multimodal data for improved sonar operations. Future work will focus on refining these models to optimize performance across varying seasonal dynamics, ensuring reliable predictions in operational marine environments.