Bathymetry of the Philippine sea with convolution neural network from multisource marine geodetic data

This study developed a deep learning-based method for high resolution bathymetry prediction in the Philippine sea, aiming to improve the accuracy of seafloor depth estimation using multi-source marine geodetic data. The method integrates geographic coordinates with auxiliary features, such as bathymetry, sea-land marks, seafloor slope and orientation, gravity anomaly, vertical gravity gradient, mean dynamic topography, deflection of the vertical, mean sea surface, and sedimentary thickness. These inputs were extracted from an 8 × 8 arcminute region around each training point and the model was trained to predict depth residuals. The trained model was applied to 1 × 1 arcminute grids to generate a detailed bathymetric map. Evaluation against ship-borne measurements shows that the developed method significantly improves prediction accuracy compared with existing models of similar resolution and performs comparably to a state-of-the art high-resolution model. This work demonstrates the potential of deep learning to enhance large-scale, cost-effective seafloor mapping using widely available geophysical datasets.