Multi-Resolution Inference of Bathymetry From Sidescan Sonar

Abstract: How to obtain complete and high-resolution bathymetry is an important research topic in the underwater domain. However, existing methods have certain shortcomings. Multibeam echosounder (MBES) can produce narrow beam range readings of the seafloor, but there is an interval between every two beams (between 10cm to 10m), and the resolution is low. Sidescan sonar can measure the seafloor in much higher resolution (down to below 1cm), but it is difficult to convert the sidescan into the bathymetry. Although several methods allow us to use physical models to estimate bathymetry from the sidescan, these methods are computationally difficult due to the growing amount of data. To bridge the gap, we propose a neural network-based system that can efficiently and accurately reconstruct high-resolution bathymetry from low-resolution bathymetry and the sidescan. In particular, the multi-resolution inference system can (1) efficiently extract the features of the sidescan sonar map; (2) reconstruct the high-resolution bathymetry using the extracted features and the input low-resolution bathymetry. Evaluations demonstrate that the inference system can reconstruct high-resolution bathymetry under different input settings.