Coral reefs are under unprecedented threat due to global anthropogenic climate change as well as local stressors from human activities including pollution and sedimentation, or destructive fishing practices. Therefore, it is crucial to monitor reefs at the highest resolution to better understand the degradation and recovery of reefs after exposure to stressors. However, current coral monitoring methods are often limited in scale due to requiring expensive equipment or substantial expert time needed for data analysis, leading to data-deficient regions, particularly in countries with restricted research resources.

This thesis aims to design scalable coral reef monitoring tools based on machine learning and 3D computer vision, ultimately democratizing reef monitoring by lowering the cost and expertise needed to conduct coral reef surveys. The main proposed methodology uses neural networks for semantic 3D mapping of coral reefs using videos from affordable consumer cameras in real-time. Four principal research contributions are presented: first, the feasibility of the general methodology called DeepReefMap, combining self-supervised neural networks for 3D mapping and general-purpose semantic segmentation of reef imagery, is demonstrated. Secondly, the 3D mapping neural networks are extended to alleviate unwanted artifacts, namely the flickering caustics and the backscatter, from underwater imagery. The third contribution is the first high-quality dataset for general-purpose semantic segmentation in coral reefs, deliberately designed for principled evaluation of machine learning models. Lastly, the components are combined into a practical system that is used in a transnational surveying endeavor, in which 365 transect videos from 45 transect sites are processed with DeepReefMap, showing high repeatability and robustness to environmental conditions.

The research of this thesis is concentrated on the Red Sea region, embedded in EPFL's Transnational Red Sea Center: data was collected in five Red Sea countries, closely cooperating with relevant local researchers and institutions, as part of expeditions and 3D mapping workshops to train local monitoring teams. As DeepReefMap is integrated into a committed monitoring program in three countries, it empowers local conservationists by giving them full ownership from data acquisition to analysis.

There are substantial implications of the presented research for coral reef monitoring: as machine learning enables flexible workflows from less curated data, high-quality annotation and cataloging efforts should be prioritized to maximize potential insights. Future work could improve the taxonomic resolution of the semantic segmentation system, incorporate fish identification, deploy DeepReefMap on underwater robots, or extend the methodology to other underwater ecosystems.