Research on improving the accuracy of remote sensing-based bathymetry on muddy coasts

The hydrodynamic conditions on shallow muddy coasts are complex, with high sediment concentrations and significant variations in underwater terrain. Several challenges, such as interference from spectral information from suspended sediments and low accuracy in model extrapolation, hinder remote sensing techniques for bathymetry inversion in these areas. Muddy coasts in Jiangsu Province, China, are typical throughout the country. Its offshore radial sand ridges (RSRs) are unique geomorphology among sand ridge systems worldwide, with a mean water depth of approximately 25 m. Taking the Jiangsu muddy coasts as the study area, we proposed a novel suspended sediment index named the three-band gradient difference suspended sediment index (TGDSSI) by analyzing the spectral characteristics of waters with high sediment concentration to enhance bathymetric information. Then, it was used as a key input variable for the back-propagation (BP) neural network model and the random forest (RF) model for bathymetry inversion experiments. We compared the performance of these two models and their inversion results on muddy coasts. Moreover, we quantitatively analyzed the contribution of TGDSSI to the accuracy of bathymetry inversion by evaluating the performance of the BP neural network without TGDSSI. Although the RF model had a slightly greater inversion accuracy than the BP neural network model, its bathymetric range was limited by the model training data, which were only up to 12 m and far less than the average water depth of the RSRs. The BP neural network performed well in large-scale extrapolation inversion, with a maximum bathymetric range of more than 30 m. TGDSSI can effectively highlight the modulation relationship between the underwater terrain and suspended sediment distribution, which enhances the underwater terrain features on images. By incorporating TGDSSI modeling, the root mean square error (RMSE) of the BP neural network model was reduced from 1.90 m to 1.67 m. We further compared the inverted water depth data from the BP neural network model with the General Bathymetric Chart of the Oceans (GEBCO) product in 2019 of the global digital bathymetric model (DBM). The R 2 increased from 0.05 in the GEBCO dataset to 0.74 in our inverted water depth results. The bathymetric accuracy improved by 66.40% in the shallow sea areas. This work showed the potential of providing underwater terrain data via satellite-based remote sensing, which will be helpful for the development, utilization, and spatial planning of muddy coasts. • Estimating underwater terrain with only remote sensing data. • Proposing TGDSSI to reflect the relationship between sediment content and bathymetry. • More robustness of RF than BP neural network for bathymetry on muddy coasts. • BP + TGDSSI performed better at large-scale extrapolation inversion for bathymetry. • The accuracy of BP + TGDSSI improved by 66.40% than GEBCO in shallow waters.