Abstract: With the widespread use of active and passive satellite remote sensing technologies for bathymetric inversion, these technologies have been demonstrated their unparalleled advantages over traditional bathymetric measurement techniques in terms of large-scale, rapid, and efficient acquisition of water depth information. This is particularly important for regions with territorial disputes or border restrictions, where obtaining bathymetric data is crucial. This study addresses the issue of limited accuracy in bathymetric inversion in deep-water areas by proposing a BP neural network-based method that uses a polynomial fitting residual distribution model in collaboration with active and passive satellite remote sensing data. First, the method establishes a BP neural network model for preliminary bathymetric inversion using ICESat-2 lidar data and Sentinel-2 remote sensing data. It then creates a residual distribution model, and finally, overlays the preliminary bathymetric inversion results with the residual distribution results to obtain the final water depth inversion. To validate the effectiveness of the proposed method, experiments were conducted in three regions: Ganquan Island, Dong Island, and Biscayne Bay. The results show that the RMSEs of the proposed method are 1.14 m, 1.06 m, and 0.49 m, respectively. These values outperform those of existing bathymetric inversion models and demonstrate good inversion performance under different seabed topography conditions. The method proposed in this study provides new ideas and technical support for developing remote sensing bathymetric inversion technology, helping to improve the inversion accuracy in deep-water areas and the overall reliability of water depth inversion.