Satellite-Derived Bathymetry Combined With Sentinel-2 and ICESat-2 Datasets Using Deep Learning

Accurate bathymetric data is critical for marine ecological balance and resource management. Deep learning algorithms, known for their capacity to model complex, multivariate, and nonlinear relationships, have been increasingly applied to satellite-derived bathymetry (SDB). However, existing deep learning models are limited by simple architectures and low efficiency in hyperparameter optimization, resulting in suboptimal training performance. This paper proposes a Convolutional Neural Network and Bidirectional Long Short-Term Memory hybrid model based on the Bayesian Optimization Algorithm (BOA-CNN-BILSTM) to enhance bathymetric inversion accuracy and efficiency. The model employs BOA to optimize the key hyperparameters of the CNN-BILSTM architecture, thereby improving inversion performance. Bathymetric inversion experiments were conducted using fused ICESat-2 and Sentinel-2 data, focusing on Coral Island and Dong Island in the South China Sea, as well as Midway Island and Oahu Island in the Pacific Ocean. Comparative experiments demonstrated that BOA significantly outperforms conventional random search (RS) by achieving near-optimal hyperparameter configurations with fewer evaluations, accelerating convergence and reducing computational costs. The BOA-CNN-BILSTM model reduced the root mean square error (RMSE) by 28.6%–56.5%, 29.6%–53.7%, 34.1%–52.6%, and 28.9%–57.1% across the study areas compared with the Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BILSTM), and CNN-BILSTM models. Other evaluation metrics also showed varying degrees of improvement. These results demonstrate that the proposed approach is effective and highly accurate for bathymetric inversion in shallow waters.