The nearshore area is of paramount importance in the ecosystem. Accurate bathymetric maps which depict underwater topography play a key role in supporting activities such as coastal research, environmental management and marine spatial planning. The new generation of Ice, Cloud, and Land Elevation Satellite 2 (ICESat-2) is equipped with an advanced Terrain Laser Altimeter System (ATLAS), which delivers considerable benefits in providing accurate bathymetric data across extensive geographical regions. ATLAS data can be combined with passive optical remote sensing imagery to realize efficient bathymetry estimation using a machine learning approach. Therefore, this study proposes a Convolutional Neural Network (CNN) model for physical radiation transmission information, which combines optical radiation transmission information with CNN models. The physical radiation transfer data highlights the spectral characteristics of shallow water areas, while the CNN structure takes into account the surrounding information of water depth measurement point pixels well. An adaptive elliptic density segmentation algorithm approach is applied to generate training and test samples based on the spectral reflectance characteristics and radiative transfer properties of Sentinel-2, using the reference bathymetry points of ICESat-2 as priori training data. The training datasets are generated based on the spectral reflectance and radiative transfer features of Sentinel-2. Next, a convolutional neural network model is appied to establish a link with the reference bathymetric point of ICESat-2. Finally, a complete bathymetric map would be generated by feeding the spectral feature data of the entire Sentinel-2 image into the trained convolutional neural network model. The obtained results are analyzed to validate the methodology, and comprehensively explores the effects of ICESat-2 extracted bathymetry point accuracy, inversion model and atmospheric correction on the performance of satellite-based remote sensing bathymetry inversion results. The continuously updated digital elevation model field data on the island of St. Croix are used to verify the accuracy and robustness of the water depth maps generated by the physical radiation transfer CNN model. The experimental results show that the adaptive elliptical density segmentation algorithm can better track water depth information compared to the standard fixed parameter density clustering algorithm. The adaptive elliptical density segmentation algorithm well eliminates the noise points and reduces the impact of noisy bathymetric points on the subsequent bathymetric inversion. The CNN model containing physical radiation transmission information exhibits higher accuracy and the RMSE using the CNN model containing physical radiation transmission information is reduced by 10% compared to the model without physical radiation transmission information in St. Thomas. The accuracy of the inversion results of the physical radiation transfer CNN model for water depth exceeds 95%, with an error of less than 1.6 m in all three study areas. In addition, the RMSE of the error evaluation of the bathymetric results using the data including the diffuse attenuation factor is 1.59 m, with an accuracy of 97%, which are better than those of the bathymetric results trained without the a priori diffuse attenuation factor, and the inclusion of the diffuse attenuation factor of the optical nature in the inversion process is favorable for the shallow water depth inversion. The ICESat-2 reference bathymetry data are used as the field data to validate the simulation estimation results, and the RMSE of the error evaluation is 1.78 m and the accuracy could reach 95%, which shows that the method is still valid and stable when using different data sources. The above results demonstrate the potential of the convolutional neural network modeling approach based on physical radiative transfer information in obtaining high-precision bathymetric information, which is expected to play an active role in the large-scale application of satellite-mounted LiDAR.