Bathymetry prediction from gravity data using combined linear regression and neural networks for enhanced cross-regional transferability

Linear regression (LR) and neural network methods are widely used to infer seafloor topography from satellite altimetry-derived gravity fields. LR methods are effective in capturing large-scale topographic features but often fail to resolve fine-scale features, whereas neural network methods can generate high-resolution bathymetry but may suffer from limited cross-regional transferability. To overcome these complementary limitations, we propose a modeling strategy that combines LR with a convolutional neural network (CNN). The physics-informed LR provides robustness to noise in gravity-field data and reliably constructs a linear bathymetry model, while the CNN focuses on learning residual bathymetry associated with local seafloor features, making learned patterns less sensitive to regional geophysical conditions. Based on this framework, a bathymetry model, termed Bath-LRCNN, is constructed in part of the South China Sea using gravity anomalies, full-tensor gravity gradients, and shipborne bathymetry measurements. The model is evaluated using independent multibeam shipborne bathymetry data. Bath-LRCNN achieves a root‑mean‑square error (RMSE) of 87.28 m, which is generally lower than those of existing bathymetry models. Compared with a purely CNN-based modeling strategy, the proposed strategy achieves comparable accuracy (87.28 m versus 87.72 m in RMSE) and shows some improvement in the representation of topographic features. The learned CNN residual model also maintains relatively stable performance in cross-regional applications.