Shallow-water Bathymetry Mapping from Satellite SAR Imagery Using Deep Learning with Multiple Feature Inputs

In response to the demand for high-precision shallow sea topography inversion and to improve the limitations of optical remote sensing, this study proposes a deep-learning model utilizing multiple feature inputs to inverse shallow sea topography from spaceborne Synthetic Aperture Radar (SAR) images. For the data acquisition and dataset construction, six high-resolution Sentinel-1 dual-polarization SAR images covering the waters northeast of Hainan Island of China in 2024 under different phases and sea conditions were collected, among which six images were used for model training, while the rest were used for testing. The reference depth was obtained from ETOPO. In dataset creation, SAR images are segmented into 8×8 sub-images. The model input is designed to consist of 8 feature variables, which involve the VV polarization backscattering coefficient $ {\mathrm{\sigma }}_{0}^{\mathrm{V}\mathrm{V}} $, radar incidence angle $ \theta $, geography information (latitude and longitude), and marine dynamic environmental parameters. The model output is reference depth from ETOPO with spatial consistency. The deep learning network comprises a convolutional layer, two BottleNeck modules from ResNet, and a fully connected layer. 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 1.57 m, and the average absolute percentage error is 6.56%, with the maximum detectable water depth reaching 49.05 m. The model presented with an RMSE of 1.95 m, and the average absolute percentage error is 11.55% for the testing dataset. Additionally, there is little difference between the two scenes with different temporal and sea conditions, indicating that the model is stable and robust. Thus, the proposed model was based on the brightness patterns observed in SAR imagery, which can detect shallow-water depths up to 50 m with high precision.