Bathymetry reconstruction from sidescan sonar images using neural networks

Abstract

We demonstrate a method to estimate bathymetry from sidescan sonar images using machine learning. The mapping of sonar image intensity values to bathymetry maps is learned by a convolutional neural network autoencoder. This network is trained with simulated images derived from the Lambertian model, and it is tested with never-before-seen simulated images and actual sidescan sonar images. After training, the network was able to estimate bathymetry maps from actual sidescan images with an error in the sub-meter level. This degree of error is useful for the navigation of ships, construction of submarine pipelines, and management of marine protected areas.