|
The ocean is a vital natural resource, providing food, transportation, energy, and recreation. Anthropogenic
and other stresses are impacting benthic communities and habitats residing on or near the seafloor. Biodiversity monitoring using benthic imagery provides insight into how widespread or localised stresses are impacting the ocean, but manual annotation of
this imagery remains a significant bottleneck, requiring up to 8 hours of expert labour per image. While machine learning (ML) offers a path toward automation, developing robust models is hindered by the scarcity of diverse training datasets and by the unique
environmental challenges of the underwater domain, such as light attenuation and domain shift. This thesis addresses these limitations by investigating whether alternative data sources, such as weakly annotated data and synthetically generated imagery, can
compensate for the lack of expert-labelled data. By developing a framework that integrates specialised training pipelines with unsupervised domain adaptation and model selection, this work demonstrates that high-performing species detectors can be trained
with minimal to no human annotation. These contributions ultimately provide an applied solution that performs at a level comparable to fully supervised models, enabling the timely and automated generation of ecological insights needed to manage the benthic
environment.
|
|
Heather Doig completed her PhD (currently under examination) at the ACFR in 2025, during which she developed new
machine learning approaches for environmental and ecological monitoring of the seafloor. Her research spans domain adaptation, synthetic data generation, and methods for learning from weak or sparse annotations in data‑limited environments. She holds a Master
of Data Science from the University of Sydney and a Bachelor of Engineering from the University of Melbourne. Prior to academia, Heather worked in technology and operations management at Macquarie Bank and at ReachOut Australia.
|
