Machine and deep learning methods for satellite-derived bathymetric mapping in Canadian coastal waters

Chifuniro Ngalande, Costas Armenakis,
Machine and deep learning methods for satellite-derived bathymetric mapping in Canadian coastal waters,
Geomatica,
Volume 77, Issue 2,
2025,
100088,
ISSN 1195-1036,
https://doi.org/10.1016/j.geomat.2025.100088.
(https://www.sciencedirect.com/science/article/pii/S1195103625000448)
Abstract: The performance of machine and deep learning (ML/DL) classification models is evaluated for depth range mapping in shallow freshwater and saltwater coastal environments in Canada using satellite-derived bathymetry (SDB). The models were trained on data from several Canadian sites, and their transferability was tested on geographically distinct, unseen sites. When compared to depth ranges derived from empirical and physics deterministic methods, the DL models, including U-Net, SegNet, and DeepLabv3 + achieved more than twice the F1 scores of in saltwater and freshwater environments, such as Graham Island and Athol, with more modest improvements observed in mixed or complex water types, such as Rimouski. From the class-wise scores, ML/DL models can moderately predict depth ranges up to 8 m in freshwater due to greater water transparency, and up to 3 m depth ranges in saltwater, where suspended sediments limit light penetration. The deterministic methods struggled to derived water depth classes in deeper and turbid waters but performed similarly to the machine learning Random Forest classifier model in both environments. In addition to the classification performance metrics (precision, recall, F1), visual assessment of the predicted bathymetric maps showed that DL models captured shoreline features, water depth gradients, and seafloor morphology more accurately, especially in shallow waters. In general, the ML/DL models faced challenges in unseen geographic domains in the 2–8 m water depth classes. Overall, this study highlights the potential of machine learning and deep learning over deterministic methods while also emphasizing the need for diverse training data improvements in model transferability.
Keywords: Satellite-derived bathymetric mapping; Sentinel-2 satellite imagery; Machine/deep learning; Model transferability; Canadian coastal waters; Fresh/salt shallow waters; Water depth range classification