Date of Award
2025
Degree Type
Dissertation
Degree Name
Doctor of Philosophy in Oceanography
Specialization
Marine Geology and Geophysics
Department
Oceanography
First Advisor
Adam Soule
Abstract
The deep sea remains one of the most crucial frontiers in Earth sciences, with vast unknowns regarding its biodiversity, ecosystem dynamics, and geological features and processes, and impact and interaction with humans. Ocean exploration seeks to address this gap by focusing data collection in some of the least known places on the planet in order to significantly expand our knowledge of marine ecosystems, geological formations, and oceanographic processes. Ocean exploration datasets provide crucial insights into a wide range of scientific domains including: deep sea biodiversity, seafloor morphology, and deep-sea soundscapes. The advent of autonomous and remotely operated underwater vehicles (AUVs and ROVs) has enabled large-scale data collection, that include high-resolution imagery, video recordings, and hydroacoustic datasets. However, the sheer volume and complexity of these datasets present significant challenges in processing, analysis, and interpretation. This research highlights the importance of these challenges and addresses them by integrating machine learning methodologies to enhance the automated analysis of three critical types of ocean exploration datasets, including underwater acoustics, imagery, and video recordings. Through the application of advanced computational approaches, this study enhances the efficiency, accuracy, and scalability of deep-sea data processing.
In Manuscript 1, we report findings from hydro-acoustic data recorded by an array of ocean-bottom seismometers (OBSs) deployed offshore near the southern flank of Kīlauea volcano in Hawaii. The OBSs recorded numerous earthquakes and lava-water interactions as molten lava entered the ocean from an eruption on the Lower East Rift Zone (LERZ). These interactions resulted in hydrovolcanic explosion that occasionally produced lava bombs, one of which struck a sightseeing boat on July 16, 2018, injuring 23 people.
To investigate this phenomenon, I analyzed hydrophone data recorded between July to mid-September 2018, to identify the hydro-acoustic signatures of lava bombs. Finally, an automated detection technique, short-term average vs. long-term average (STA/LTA) was applied, to identify additional events in this active time period and create a catalog of seismo-acoustic events associated with the volcanic eruption and entrance of lava into the sea.
In Manuscript 2, we present the development of ROVIA (Remotely Operated Vehicle Intelligent Annotator), a portable, field-deployable Convolutional Neural Network (CNN) model designed to automatically identify highlights in long-duration deep-sea dive videos collected by Autonomous Underwater Vehicles (AUVs) and Remotely Operated Vehicles (ROVs). These videos serve as valuable resources in marine science research but are often sparse, containing high-value clips interspersed with hours of underwater footage, posing significant challenges for managing large-scale video datasets.
ROVIA addresses this issue by efficiently extracting spatiotemporal features related to camera zooming, organism movement, and optical flow variations to detect highlights. The current model achieves an accuracy of over 80% with a specificity of 94%, significantly improving the process of identifying and extracting relevant clips from extended underwater video recordings. Consequently, this tool not only enhances video archiving efficiency but also facilitates the easy utilization of these short video clips for scientific research and educational purposes.
In Manuscript 3, we explore the application of supervised machine learning techniques to automate seafloor morphology classification, reducing the need for extensive manual annotation. Deep-sea photo surveys conducted using towed, robotic, or autonomous vehicles play a crucial role in seafloor geological studies, particularly in submarine volcanic regions. Since, analysis of such voluminous datasets involves manual annotation efforts, processing and labelling of these datasets can be accelerated significantly by developing automation tools. With this goal in mind, in this study, we trained and tested different supervised learning algorithms on digital seafloor images collected from depths of ~2500m in the East Pacific Rise after the 2005 - 2006 volcanic eruption. The model classifies seafloor images by providing labeled categories as output. This approach serves as a powerful tool for geological studies and supports broader applications in oceanographic research.
In Manuscript 4, we investigate the use of unsupervised learning techniques for the classification of deep-sea soundscapes. For this analysis, we utilized passive acoustic data collected in 2018 by the Deep Autonomous Profiler (DAP), a hadal lander deployed at ~8.3 km depth for ocean profiling and acoustic data collection. The recordings were obtained using a custom-built, full-ocean-depth hydrophone capable of withstanding extreme pressure conditions. By applying unsupervised learning algorithms, we identify and categorize sound sources based on their acoustic signatures, demonstrating the automated classification of previously unknown sound events. This approach enhances soundscape analysis, offering a novel framework for studying deep-sea ecosystems and evaluating the impact of human activity in these remote environments.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Banerjee, Puja, "MACHINE LEARNING FOR OPTIMIZED OCEAN EXPLORATION DATA ANALYSIS" (2025). Open Access Dissertations. Paper 4439.
https://digitalcommons.uri.edu/oa_diss/4439