Abstract
This study refines the gravity anomaly model derived from altimetry data by employing a multilayer perceptron (MLP) neural network to integrate multi-source geophysical data (longitude, latitude, gravity anomaly, geoid height, bathymetry, and sediment thickness) based on shipborne gravity. To reduce the impact of land on gravity anomaly inversion, the experimental area is divided into nearshore and offshore regions, with separate inversions for each. The model is trained using differences between shipborne gravity control points and 8'×8' grid points as input data, and differences between control point gravity anomalies and SDUST2022GRA model values as output data. The trained model predicts gravity anomalies at grid centers, and SDUST2022GRA values are applied to restore the predicted anomalies. The Gulf of Mexico region (81°W-99°W, 15°N-32°N) is selected to establish a high-resolution (1'×1') MLP Gravity Anomaly model (MLP_GRA). Compared to the SDUST2022GRA, SIO_V32.1, and DTU21GRA models, the MLP_GRA model reduces the standard deviation (STD) and mean absolute error (MAE) by 0.4 mGal and 0.32 mGal, 0.54 mGal and 0.37 mGal, and 0.39 mGal and 0.27 mGal, respectively. These results confirm the reliability and effectiveness of the proposed method.