<p>Convolutional Neural Networks (CNNs) have been widely used to model the nonlinear relationship between gravity anomalies and seafloor topography. However, most CNN-based methods operate only in the spatial domain, which limits resolution and hinders the recovery of fine-grained topographic features. To address this, we propose SGL-CNN, a novel framework that extracts multi-input features from both spatial and frequency domains. By integrating multi-component gravity anomalies with long-wavelength bathymetric data, our model simultaneously captures low-, medium-, and high-frequency seafloor components, enabling more detailed topographic reconstruction. We validate SGL-CNN in three representative regions of the Western Pacific–a slope, a seamount, and a trench–against baseline methods (ParkerO, SAS, GGM, LCNN). Accuracy and PSD results show that SGL-CNN outperforms others over seamounts and trenches. Across diverse terrains and depth ranges, its dual-domain three-branch architecture (Spatial, Global and Local Frequency) effectively handles multi-scale wavelength distributions, recovering low-, medium-, and high-frequency components corresponding to slope trends, seamount bodies, and trench fracture zones. Ablation studies confirm the necessity of the proposed architecture and the long-wavelength bathymetric input, and further validation on a high-latitude grid supports its generalizability. In summary, the synergistic fusion of spatial and spectral features in SGL-CNN overcomes spectral truncation issues in single-domain methods, achieving high-resolution bathymetric inversion.</p>