Comparison of Satellite-Derived Bathymetry Algorithm Accuracy Using Sentinel-2 Multispectral Satellite Image

Abstract

The utilization of satellite image data and image data processing techniques has become an efficient alternative to obtain bathymetric data in a broad and complicated area. This study aimed to determine the algorithm's performance in the waters of Lambasina Island. Atmospheric and radiometric correction using the Dark Object Subtraction (DOS) method for initial processing of Sentinel-2 images. The multispectral channel used, namely the blue, green, and red bands, was tested by regression using field observation data. The algorithms used to estimate bathymetry include Lyzenga, Stumpf, and Support Vector Machine (SVM). The test results of the three algorithms showed that the support vector machine algorithm was the best algorithm for estimating bathymetry after the Stumpf and Lyzenga algorithms. The correlation results of the SVM algorithm in the waters of the small Lambasina island got a correlation coefficient of determination R2 = 0.81 and the large Lambasina waters area R2 = 0.82. The second-best algorithm was Stumpf, with a correlation coefficient of determination of R2 = 0.79 in the waters of the small Lambasina island and R2 = 0.80 in the waters of the large Lambasina island. Lyzenga's algorithm got the correlation coefficient of determination R2 = 0.78 on small Lambasina Islands and large Lambasina Islands with a determination correlation coefficient value of R2 = 0.79.

Methods to Improve the Accuracy and Robustness of Satellite-Derived Bathymetry through Processing of Optically Deep Waters

“…These models have the major advantage of not considering the optical characteristics of water bodies and automatically capturing the correspondence between complex image spectral values and depth through sample learning, resulting in fast predictions and low computational costs. Typical empirical models include those based on logistic regression [27], support vector machines [28][29][30][31], statistical models, and rapidly developing machine learning models [26,32]. Nevertheless, these models have inherent empirical biases, and their prediction accuracy relies on sample representativeness.…”