Comparison of Six Empirical Methods for Multispectral Satellite-derived Bathymetry

Abstract Satellite-derived bathymetry (SDB), an important technology in marine geodesy, is advantageous because of its wide coverage, low cost, and short revisit cycle. At present, several different kinds of SDB methods exist, and their inversion accuracy is affected by algorithm performance, band selection, and sample distribution, among other factors. But these factors have not been adequately quantified and compared. In the present study, we evaluate the performances and highlight the best scenarios for applying the six classical empirical methods including the log-transformed single band, band ratio (BR), Lyzenga polynomial (LP), support vector regression, third-order polynomial (TOP), and back propagation (BP) neural network. The results reveal that the number of training samples is important for the empirical SDB methods, and the TOP and BP methods need more training samples than other methods. Compared to the robust BR and LP methods, the TOP and BP methods can obtain high accuracy but are severely influenced by incomplete samples. In addition, experiments that prove the local minimum (poor robustness) problem of the BP method exist and cannot be ignored in the bathymetry field. The present study highlights the most suitable method for obtaining reliable SDB results and their applicability.