Bathymetric model generation in shallow waters with optical satellite images and machine learning algorithms

Abstract

Bathymetric mapping is essential for understanding ocean dynamics, mapping ecosystems,
and forecasting coastal erosion. The goal of this study is to produce bathymetric models in
shallow coastal area of rarely investigated Antarctic region using machine learning methods
and Sentinel-2 satellite image. Based on satellite imagery and multibeam echosounder data,
two algorithms, random forest (RF) and support vector machine (SVM), were used to predict
ocean depths. The accuracy criteria used to evaluate the models' performance included RMSE,
MAE, and R2. With an RMSE of 1.51, an MAE of 1.04, and an R2 of 0.77, the RF model produced
promising results. These metrics provided low errors and a good fit between projected and
observed water depths. The SVM model also provided promising results with slightly lower
performance with an RMSE of 1.58, an MAE of 1.13, and an R2 of 0.75. Overall, this study
showed that above mentioned algorithms can be used as viable approach for generating
bathymetric models in shallow coastal areas. These models can contribute to our
understanding of underwater topography, ecosystem dynamics, and the impacts of climate
change.