Owing to the development of satellites (e.g., LT-1, GF-1) (Li et al., 2023), we can rapidly acquire optical and SAR images in areas affected by the 2025 Dingri earthquake, interpret their source parameters, and assess their induced hazards. This provides strong support for timely scientific research and post-earthquake rescue-work, greatly improving the efficiency and scientific validity of disaster response.

This study combines data-driven DL with physics-driven tomography inversion to construct a more accurate subseafloor P-wave velocity model and applies it to OBS data from the SCS. The experimental results show that adding SSIM to the U-net can enhance the network’s ability to capture data details and improve learning efficiency. On the real data, this method shows good effectiveness and reliability in identifying stratigraphic interfaces and complex geological structures. Furthermore, applying the cosine transform for data preprocessing extracts key features, further improving neural network efficiency and effectiveness, and offering a feasible solution to the issue of limited sample size. Although some progress has been made, the accuracy improvement is still limited, and future research will incorporate physical constraints. By constructing an objective function containing physical constraints, the subseafloor velocity inversion results can follow the laws of geophysics and improve the accuracy and interpretability of the model.