Autonomous Underwater Vehicles (AUVs) have catalysed a significant shift in the way marine habitats are studied. It is now possible to deploy an AUV from a ship, and capture tens of thousands of georeferenced images in a matter of hours. There is a growing body of research investigating ways to automatically apply semantic labels to this data, with two goals. The task of manually labelling a large number of images is time consuming and error prone. Further, there is the potential to change AUV surveys from being geographically defined (based on a pre-planned route), to permitting the AUV to adapt the mission plan in response to semantic observations.udThis thesis focusses on frameworks that permit a unified machine learning approach with applicability to a wide range of geographic areas, and diverse areas of interest for marine scientists. This can be addressed through the use of hierarchical classification; in which machine learning algorithms are trained to predict not just a binary or multi-class outcome, but a hierarchy of related output labels which are not mutually exclusive, such as a scientific taxonomy.udIn order to investigate classification on larger hierarchies with greater geographic diversity, the BENTHOZ-2015 data set was assembled as part of a collaboration between five Australian research groups. Existing labelled data was re-mapped to the CATAMI hierarchy, in total more than 400,000 point labels, conforming to a hierarchy of around 150 classes.udThe common hierarchical classification approach of building a network of binary classifiers was applied to the BENTHOZ-2015 data set, and a novel application of Bayesian Network theory and probability calibration was used as a theoretical foundation for the approach, resulting in improved classifier performance. This was extended to a more complex hidden node Bayesian Network structure, which permits inclusion of additional sensor modalities, and tuning for better performance in particular geographic regions.
展开▼
