Boulder detection in a nearshore seabed through segmentation of bathymetric data gathered by a multibeam echosounder

This master’s thesis aims to deploy and test machine learning techniques to find areas of the seabed without submerged boulders using high-resolution bathymetric data. Undoubtedly, the growing interest in marine infrastructure development requires being able to hasten the planning and construction process of these facilities. Accordingly, the automatic localization of the most suitable areas to lay the foundations of these underwater structures is essential to boost the planning phase. In recent years, the usage of autonomous vehicles in submarine tasks, particularly for seabed surveying, has soared due to the required extended operating times and the need for high accuracy in performing the mission. Although underwater robots may guarantee a high level of autonomy in maneuvering control and path planning, the analysis of the collected data still proves to be an issue. Yet specialists are spending many hours gaining information from bathymetric maps to probe seafloor bed- forms and morphological features that could hinder any possible infrastructure construction. Different machine learning techniques are the objects of investigation in this thesis to find an automatic solution to deal with boulder detection. Unsupervised learning methodologies prove to be fast to deploy, but the accuracy obtained by the algorithms is rather low. Afterward, supervised techniques are investigated and they are demonstrated to be more reliable in detecting boulders if manually classified data are available. Still, some doubts may arise concerning which features are convenient for the classification and the expenses of dealing with large quantities of data. Finally, some artificial neural network architectures are proposed to solve the boulder identification problem. These implementations allow automatic feature generation and direct use of large native bathymetric data.