A Geographically and Physics-Informed Convolutional Neural Network for Satellite-Derived Bathymetry Using ICESat-2 and Sentinel-2 Datasets

Abstract

Accurate bathymetric data are vital for the development of the blue economy and marine ecological civilization, but conventional surveys are costly and spatially limited. Satellite-derived bathymetry offers broader coverage at lower cost, yet its accuracy is often constrained by limited reference-depth data and spatial variations in water-column optical properties and bottom reflectance. To address these limitations, we propose a geographically and physics-informed convolutional neural network (GPI-CNN) that integrates Sentinel-2 multispectral imagery with ICESat-2-derived bathymetric samples and geographic coordinates. By incorporating geographic and physics-informed features derived from radiative-transfer theory, the model accounts for spatial variations in the spectral-depth relationship and improves bathymetric inversion accuracy. The proposed method was evaluated in two representative coastal areas: Coral Island in the Xisha Islands of the South China Sea and Guam in the western Pacific. The GPI-CNN model was compared with the traditional Stumpf model, Random Forest (RF), and U-Net. The results demonstrate that GPI-CNN outperforms the Stumpf model, RF, and U-Net, achieving root mean square errors of 0.53 m and 0.85 m, and mean absolute errors of 0.34 m and 0.64 m, respectively. These results confirm the advantage of GPI-CNN in improving the accuracy of shallow-water bathymetric estimation.