We　propose　a　method　for　enhancing　the　accuracy　of　bathymetry　models　based　on　a　multilayer　perceptron　(MLP)　neural　network　that　integrates　the　differences　in　multisource　marine　geodetic　data　(MMGD)　(longitude,　latitude,　reference　bathymetry,　slope,　the　meridional　and　prime　components　of　vertical　deflection,　gravity　anomaly,　vertical　gravity　gradient,　and　mean　dynamic　topography).　First,　we　use　the　MMGD　differences　between　the　shipborne　sounding　control　points　within　8　＇　x　8　＇　grid　points　and　shipborne　sounding　control　points　as　input　data,　as　well　as　the　differences　between　the　topo_24.1　model　and　the　measured　bathymetric　values　at　the　control　points　as　output　data　to　train　the　MLP　model.　Second,　we　feed　the　input　data　from　the　central　point　of　a　1　＇　x　1　＇　grid　into　the　MLP　model　to　obtain　predictions,　and　then　use　the　topo_24.1　model　to　recover　the　predicted　bathymetry　at　the　prediction　point.　We　focus　on　the　Caribbean　Sea,　and　construct　a　Caribbean　Bathymetric　Chart　of　the　Oceans　(CBCO1)　model　using　MLP　neural　network.　The　reliability　of　MMGD,　a　CBCO2　model　using　MMGD,　and　the　reliability　and　effectiveness　of　the　overall　method　are　demonstrated　through　comparisons　with　the　CBCO2,　GEBCO_2022,　topo_24.1,　DTU18　models　at　the　checkpoints.