Abstract: Objectives: High-resolution bathymetric modeling plays a critical role in marine tectonic studies, resource exploration, and navigation safety. Traditional bathymetry inversion methods based on satellite altimetry-derived gravity data are often limited by insufficient nonlinear representation capability, reduced adaptability in areas with complex geomorphology, and relatively high computational cost. To overcome these limitations, a convolutional neural network (CNN)-based bathymetry inversion approach integrating multi-source spectral band-decomposed differential gravity features is proposed for 15″ resolution seafloor modeling in the Yap Trench region. Methods: Gravity anomaly, vertical gravity gradient, and deflection of the vertical data are first transformed into the frequency domain using a two-dimensional Fourier transform. Spectral bands are partitioned according to coherence characteristics to characterize gravity responses to bathymetry at different spatial scales. Multi-source differential features are then constructed by integrating band-decomposed gravity components, spatial location information, a prior bathymetric model (SDUST2023BCO), and sediment thickness data. Feature samples are organized around multibeam shipborne sounding control points and their surrounding 64 × 64 grids. Based on these inputs, a 15″ × 15″ bathymetric model, named the Yap Trench Bathymetric Chart of the Oceans (YT_BCO), is established. To evaluate the effectiveness of the spectral band decomposition strategy, a control model (YP_BCO2) without band partitioning is constructed under identical network architecture and training settings. Independent single-beam sounding data are adopted for validation, and comparative analyses are conducted against topo_27.1, SRTM15+V2.7, and GEBCO_2025. Results: The YT_BCO model achieves a standard deviation of 111.80 m with respect to independent check data, representing improvements of 4.66% and 10.25% compared with topo_27.1 and SRTM15+V2.7, respectively. Overall performance surpasses that of the non-band-decomposed model YP_BCO2, confirming the effectiveness of spectral band processing. Although slightly inferior to GEBCO_2025-likely due to its integration of more extensive in situ datasets-the proposed model demonstrates stable and reliable performance in the complex trench environment. SHAP-based interpretability analysis reveals that band-decomposed vertical gravity gradient features contribute most significantly to the predictions, accounting for 41.0% of total feature importance. Conclusions: The integration of spectral band-decomposed multi-source gravity differential features within a CNN framework effectively enhances nonlinear representation capability and inversion accuracy in regions with complex bathymetry. The proposed strategy improves the robustness and precision of gravity-based bathymetry inversion and provides technical support for high-resolution seafloor topographic modeling.