Advancing Marine Autonomy: Perception, Classification, and Workforce Development for Sustainable Ocean Monitoring

Advances in robotics, artificial intelligence (AI), and image recognition have significantly enhanced the monitoring of marine wildlife populations, which serve as valuable indicators of environmental changes and impacts. However, existing marine wildlife datasets face substantial challenges, including limited realistic data, missing annotations, imbalanced (long-tailed) class distributions, fine-grained class variations, and hierarchical classification complexities.

In this thesis, a novel approach to realistic marine wildlife detection is proposed, using autonomous underwater vehicles (AUVs) designed to perform multiple vision-based tasks simultaneously. This approach optimizes resource utilization in complex perception systems and enables AUVs to perform wildlife recognition in tandem with core perception functions such as target localization. By leveraging an efficient object detection architecture and loss-based training methods, this work addresses key challenges such as incomplete annotations and the computational constraints of AUV platforms.

Building on this, this thesis introduces a fine-grained hierarchical classifier capable of adapting to unknown class distributions during testing. By integrating a self-supervised ensemble learning technique with a hierarchical classifier architecture, the proposed solution demonstrates better performance, surpassing baseline models in 7 of 9 evaluation metrics across three diverse test distributions.

In addition to these contributions to AUV perception systems, we develop methods to prepare engineers for using and developing AUV systems. These methods form a practical and comprehensive curriculum for autonomous systems development, focused on equipping engineers with the skills to deploy advanced perception, planning, and control systems. This curriculum bridges the gap between theoretical knowledge and practical implementation, laying a strong foundation for the next generation of autonomy engineers.

These three contributions address needs for autonomous AUVs through workforce development and a robust online-offline pipeline for wildlife image recognition and labeling, leveraging the persistent autonomy capabilities of AUVs. By addressing critical technological and educational gaps, this thesis advances marine autonomy and deep learning, making significant strides toward marine wildlife conservation and sustainable ocean monitoring.