Denoising bathymetric data through deep learning

Abstract

Bathymetric data is an extension of a point cloud and is used to describe the seafloor relative to the water’s surface. It can be used to create maps that portray what the seafloor looks like. It is important for the surfaces to be represented as accurately as possible. Bathymetric data often consists of errors due to challenges related to underwater communication. The errors deviate from the average depth of the surface around it and can be cleaned using statistical software tools. Unfortunately, cleaning it manually is both time-consuming and repetitive.

This thesis is about using deep learning as a tool to denoise bathymetric data. Points clouds are difficult targets for deep learning since they lack structure. The thesis proposes several methods to supply the necessary structure for deep learning, as well as capture the differences between errors and actual surface. The most prominent method treats the bathymetric data as a grid where each cell in the grid represents information about the depth of the area.

The resulting model is capable of identifying all of the obvious errors on the surface. The ability to identify errors comes with the downside of struggling with steep terrain. Yet, the method shows promise despite not being able to reach the safety mark. There are also other methods besides the most promising ones. These provide further insight as to whether deep learning can be used to denoise bathymetric data