Satellite-Derived Bathymetry Using Machine Learning Methods

Abstract

This study evaluates the performance of machine and deep learning (ML/DL) models for satellite-derived bathymetry (SDB) compared to traditional empirical and physics-based methods in high-latitude freshwater and saltwater environments. The models were trained on data from several Canadian sites, and their transferability was tested on unseen sites. DL models, U-Net, SegNet and DeepLabv3 +, achieved about two times higher F1 scores than traditional methods. From the class-wise scores, ML/DL models can reliably predict up to a depth of 8m in freshwater, due to greater water transparency, and up to 3m in saltwater, due to sediments that prevent light penetration. Traditional methods struggled in deeper and turbid waters, but performed similarly to the machine learning Random Forest model in both environments. Overall, the ML/DL models have difficulties generalizing to unseen data or data from geographic locations not included in the training process, particularly in saltwater due to poor light penetration. The depth results met the Zones of Confidence (ZOC) categories, CATZOC C and CATZOC D, of the International Hydrographic Organization (IHO) depth accuracy standards. Overall, this study highlights the advantages of machine learning over traditional methods while identifying challenges in model generalization and data diversity.