Advancing Coral Reef Bathymetry: A GAN-Augmented and Stratified CNN Analysis of Fused ICESat-2 and Sentinel-2 Dataset

High-resolution bathymetry mapping of coral reefs is essential to morphodynamic study of reef habitats, assisting reef monitoring and conservation under global climate change. However, the accuracy of conventional satellite-derived bathymetry (SDB) is reduced at depths over 15 m with optical signal attenuation and training data insufficiency. To address this gap, here, we present an approach that synergizes ICESat-2 advanced topographic laser (ATL24) photon-counting LiDAR data with Sentinel-2 multispectral imagery. A generative adversarial network (GAN) is implemented to offset dataset deficiency at deeper depths, and a stratified convolutional neural network (CNN) is adapted to distinct optical-depth regimes. Bathymetry derived at Jiuzhang Atoll is in good agreement with the in situ multibeam measurements, with a mean absolute error (MAE) of 0.75 m and a root-mean-squared error (RMSE) of 10% of the present maximum depth of 19 m, validating the effectiveness of GAN-driven sample synthesis to make up measurement inadequacy, and the enhancement of model generalizability across a wide depth range by stratified CNN. This approach could be applied to bathymetry mapping of coral reefs worldwide at depths of 15-30 m, where biodiversity generally increases the most with multisource satellite observations.