Supervisor/s Gabriel Oliver Codina | Francisco Bonin-Font
Most of the land terrain on Earth is being constantly mapped using sub-meter resolution spaceborne
images. However, about eighty percent of the oceans’ seafloor remains unexplored and most of the benthic habitat and geological structure distribution continues unknown. The main problem is that the methods that work on land are not applicable underwater. Remote sensing (RS) methods using spaceborne or airborne images can be used to map shallow water environments up to 10m depth and acoustic RS methods using multibeam echosounders are now providing increased resolutions on the backscatter data that can be used for benthic habitat mapping (BHM). Nonetheless, the feature richness of data collected with these types of methods is very limited. Most of the methods used for BHM require the use of in situ (IS) data in order to validate or even train their mapping algorithms.
This Thesis aims to facilitate the acquisition of IS data by pushing the boundaries of autonomous underwater vehicles (AUVs). Nowadays, AUVs are increasingly used to acquire ocean data, bridging the gap between the use of remote operated vehicles (ROVs) and shipborne acoustic RS. However, their autonomy is usually limited by the condition of following preprogrammed paths, their behavior is blind with respect to benthic data collected and are autonomous only in the sense that they are not tethered and are able to estimate their location and control their motion. This Thesis presents three novel methods to (1) provide an online semantic perception of the environment by processing the online data flow and building a probabilistic model of the benthic environment, (2) enlarge the decision-making autonomy of AUVs providing an adaptive capacity to replan mission paths based on a semantic understanding of the physical variable under study that depends on the objective of the campaign, (3) improve the autonomous navigation for in situ image recording of AUVs.
All in all this thesis build a data-driven exploration architecture that automates the in situ sampling
process on unknown environments while maximizing data informativeness.
The algorithms described in this Thesis have been extensively validated in field using an AUV equipped with a stereo camera rig used to gather images of the seabed partially covered by Posidonia oceanica seagrass meadows.