Satellite derived bathymetry using empirical and machine learning approaches: a case study in the highly dynamic coastal water

Abstract

The estimation of nearshore water depth is essential in many near-shore marine sciences and coastal engineering. However, direct measurements using fieldwork methods can be expensive and time-consuming. Also, near-shore waters are typically turbid and very dynamic, which decreases the accuracy of previous methods for clear waters. In this study, we compared the performance of two Satellite-Derived Bathymetry approaches, including empirical (in particular, the Stumpf and multi-linear regression models) and five non-linear regression of machine learning algorithms (Random Forest, Support Vector Machine, CastBoost (CB), Extreme Gradient Boosting, and Light Gradient Boosting Machine) to provide an efficient method for deriving accurate bathymetric data in a high dynamic coast, the Hasaki located in Ibaraki Prefecture, Japan, from the multi-spatiotemporal satellite images of the Sentinel-2 constellation. A total of 70% depth data points of each year (2016, 2018, and 2020) obtained from boat-based echo-sounding measurements are used for training, and the remaining 30% are used for testing each approach. CB revealed an outstanding performance (R2 ranging from 0.84 to 0.92) in coastal area of Hasaki down to ~15 m depth at all the 3 years of datasets, with mean RMSE values less than 0.5 m when the depth less than 7 m.