An Attention-GraphSAGE Algorithm for Marine Gravity Anomaly Inversion Using Denoised Photon Point Cloud Data of ICESat-2

The ICESat-2 (Ice, Cloud, and land Elevation Satellite-2) provides abundant ocean satellite altimetry data. Photon point clouds data denoised by the official ATL03 algorithm exhibit insufficient continuity, and classical gravity anomaly inversion algorithms suffer from high computational complexity. To address these issues, we propose a two-step denoising method for ATL03 data to obtain instantaneous sea surface heights (SSHs): photon point clouds data are denoised using the adaptive OPTICS algorithm, followed by secondary denoising using the Linear-Interquartile algorithm. The resulting SSHs demonstrate superior continuity and larger data volume than those of ATL12 data. This paper proposes the Attention-GraphSAGE algorithm—a graph neural network approach based on neighbor node sampling and self-attention-weighted feature aggregation. An Adam optimizer with L2 regularization is employed for iterative training to achieve nonlinear fitting. The architecture incorporates two-layer neighbor node sampling and aggregation, with residual connections between layers to prevent gradient explosion and preserve original data features. During model training, each input contains 81×81×4 feature values. Nodes consist of shipborne measurement points and surrounding grid points; edges represent connection relationships between shipborne points and first-layer neighbor nodes (grid points), as well as connections between grid points and second-layer nodes (grid points). Edge weights are determined by calculating the correlation coefficient of geoid height between each neighbor node and the shipborne measurement point (using self-attention mechanism). The output corresponds to the difference between shipborne gravity anomaly data and SIO V32.1 gravity anomaly data. The gravity anomaly model inverted utilizing denoised photon with the Attention-GraphSAGE algorithm (AGP-GRA model) demonstrated a correlation coefficient of 0.99 and a standard deviation of 3.36 mGal with shipborne gravity anomaly data. Compared to the model inverted directly utilizing ATL12 data with the same algorithm (AGA-GRA model), this represents a standard deviation reduction of 0.09 mGal. Experiments confirm the algorithm’s effectiveness for gravity anomaly inversion demonstrating with favorable model performance.