PSO-optimized dual-channel BP neural network for bathymetric prediction using multisource marine geodetic data : a case study of the gulf of Mexico

Abstract

Accurate seafloor topography is essential for marine scientific research, resource exploration, and understanding geological processes. Traditional bathymetric surveying methods are constrained by limited spatial coverage and high operational costs, particularly in deep-sea environments. To overcome these challenges, we developed a Particle Swarm Optimization (PSO)-optimized dual-channel BP neural network (PSO_BP), integrating shipborne bathymetric data with satellite altimetry-derived gravity anomalies. These gravity anomalies were further decomposed into long-wavelength, short-wavelength, and residual components to enhance bathymetric prediction accuracy. We systematically evaluate the impact of different gravity data combinations, including gravity anomalies, gravity gradients, and vertical deflections, used individually, in pairs, or as a three-component combination, on bathymetric prediction accuracy. Results show that PSO_BP consistently outperforms existing models (GEBCO_2024, Topo_25.1, DTU18_BAT, and SRTM15 + V2.6), achieving the lowest RMSE (25.45 m), MAE (9.95 m), MAPE (3.70 %), and highest R2 (99.96 %) across various depth ranges and shoreline distances. The decomposition of gravity anomalies into long- and short-wavelength components and their residuals proves to be the most effective approach for improving bathymetric prediction accuracy, while PSO optimization enhances model convergence and reduces prediction errors. This study highlights the importance of integrating diverse gravity datasets and advanced optimization techniques to improve the accuracy and robustness of seafloor depth prediction, offering a reliable solution for global bathymetric mapping in deep and remote ocean regions.