Regional Marine Gravity Field Refinement by Integrating Region-Adaptive Fusion and Multi-Relational Graph Residual Learning

Satellite-altimetry-derived marine gravity models often exhibit limited regional adaptability and region-dependent residual errors relative to shipborne observations, particularly in coastal, shelf, and slope areas. This study proposes a regional marine gravity refinement method that integrates region-adaptive background-field fusion with multi-relational graph neural network residual learning in the northern South China Sea and adjacent waters. Four background models—SIO/UCSD, SDUST2022GRA, NSOAS24, and SWOT05—were used together with shipborne gravity, bathymetry, distance-to-coast, and survey-line information. A region-adaptive initial field was first constructed according to the error characteristics of the background models in different subregions, and its difference from the shipborne observations was taken as the residual-learning target. The matched shipborne points were then represented as graph nodes, with spatial, terrain, model-response, survey-line, and regional relations used to construct a multi-relational graph. Spatially disjoint blocks were used to separate the training, validation, and test samples. The Multi-relational GNN predicted local residuals, which were added back to the region-adaptive initial field to obtain the refined gravity anomalies. On the held-out test set, the proposed method achieved an RMSE of 6.10 mGal, an MAE of 3.83 mGal, a 95th-percentile absolute error of 13.08 mGal, a bias of −0.19 mGal, and a squared Pearson correlation coefficient r2 of 0.945, outperforming the interpolation, machine-learning, and neural-network baselines. Ablation experiments showed that the survey-line relation provided the largest individual contribution. However, the present validation is limited to spatially held-out samples within the existing shipborne survey network; generalization to completely unseen cruises, other marine regions, and areas without nearby shipborne constraints remains to be further verified. Overall, the results indicate that combining regional background-model adaptation with structured residual learning can improve regional marine gravity field refinement in spatially heterogeneous environments.