SGL-CNN: A Dual-Domain Convolutional Neural Network harnessing Spatial and Frequency Features for Bathymetry Estimation

Bathymetric prediction using Convolutional Neural Networks (CNNs) to model the nonlinear relationship between gravity anomalies and seafloor topography has gained significant attention. However, most CNN-based methods operate primarily in the spatial domain, which often results in limited resolution and an inability to recover fine-grained topographic features. To address this, we propose a novel feature engineering framework, spatial global- and local-convolutional neural networks (SGL-CNN), that extracts multi-source input features from both spatial and frequency domains. This architecture enables the integrated "multi-source gravity + long-wavelength" data to simultaneously represent both low- and high-frequency components of the seafloor, facilitating the reconstruction of more detailed topography. The effectiveness of our SGL-CNN model is validated in representative regions of the Western Pacific, including a slope, a seamount, and a trench, against established baseline methods (ParkerO, SAS, GGM, CNN). Results demonstrate that SGL-CNN outperforms the baseline methods in predicting bathymetry over seamounts and trenches. By presenting results across diverse terrains and depth ranges, we show that SGL-CNN's dual-domain architecture effectively handles the multi-scale wavelength distribution of complex seafloor landscapes—recovering low-frequency (slope trends), medium-frequency (seamount bodies), and high-frequency (trench fracture zones) components. This capability overcomes the spectral truncation issue common in single-domain methods, enabling high-resolution bathymetric inversion.