Shallow Sea Bathymetry Mapping from Satellite SAR Observations Using Deep Learning

This study proposes a deep-learning model to map underwater topography from spaceborne synthetic aperture radar (SAR) images. For the model development, 18 high-resolution Sentinel-1 dual-polarization SAR images covering the southern Cape Cod area from 2014 to 2022 are collected, among which twelve images were used for model training, while the rest were used for testing. In order to facilitate the model learning complex and nonlinear relations among various parameters, the model input is designed to consist of 8 variables, which involve radar backscatter coefficient, radar observation geometry, geography information and marine dynamic environmental parameters. The deep learning network consists of a convolutional layer, two residual blocks, and a fully connected layer, facilitating feature extraction and regression prediction. The final model performances in retrieving shallow water depth are shown as follows: For the training set, the model achieved a root mean square error (RMSE) of 1.02 m and a relative error of 4.32%, with the maximum detectable water depth reaching 48.17 m. For testing, the best performance of the model presented with an RMSE of 1.59 m and a relative error of 5.29%, while the worst performance corresponds to an RMSE of 4.53 m with a relative error of 15.61%. Thus, the proposed model has the capacity of detecting the shallow water depth up to 50 meters with high precision.