Generalized satellite-derived bathymetry across spatial and temporal domains: a domain-adaptive deep learning approach with multi-source remote sensing data

Accurate bathymetric data are crucial for the protection of marine ecosystems and for various human activities. Satellite-derived bathymetry (SDB) based on an empirical approach has been widely employed to estimate shallow water depths. However, the empirical approaches are regression models dependent on specific data, facing challenges in generalizing across different spatial and temporal domains. These challenges can be attributed to domain shifts caused by variations in water quality, substrate, and atmospheric conditions. This limitation hampers their scalability for large-scale or long-term monitoring applications. In this paper, we propose a domain adaptation-based deep learning model for satellite-derived bathymetry (DA-SDB) to address the limitations. Specifically, the DA-SDB model comprises three components: a feature extractor, a bathymetry predictor, and a domain aligner. The feature extractor integrates a pyramid-like block (PLB) and a physical-assisted block (PAB) to improve the data utilization and extract domain-invariant features. The domain aligner mitigates domain shift by aligning the pseudo-inverse Gram matrices. We conducted spatial and temporal transfer experiments in five diverse study areas (Dongsha Atoll, Bimini Island, South Warden Reef, Hadrah Island, and Mubarraz and Bu Tinah Islands). The DA-SDB model demonstrated significant improvements in generalization ability (with the root-mean-square error (RMSE) and mean absolute percentage error (MAPE) reduced by 0.27  m and 21.51 %, respectively) and stability (achieving 11 out of 12 best results) over the state-of-the-art methods. Notably, the DA-SDB model maintains robust accuracy across a wide range of depths, as shown by its agreement with reference bathymetry. It employs a top-of-atmosphere (TOA) reflectance dataset and achieves effective results without atmospheric correction. Utilizing this deep learning method, the bathymetric mapping of remote areas and long-term monitoring of critical regions can be conducted rapidly and cost-effectively.