ßÉßɱ¬ÁÏ

Topic

The Creation, Validation, and Comparison of a Novel Computer Vision-Assisted Image Monitoring Method for Assessing Long-Term Underwater Diversity

Department of Biology

Date & location

• Tuesday, August 26, 2025
• 1:00 P.M.
• Clearihue Building, Room B007

Examining Committee

Supervisory Committee

• Dr. Francis Juanes, Department of Biology, ßÉßɱ¬ÁÏ (Supervisor)
• Dr. Rodney Rountree, Department of Biology, UVic (Member)
• Dr. Colin Bates, Senior Research Scientist, Cascadia Seaweed (Outside Member)

External Examiner

• Dr. Chris Rooper, Pacific Biological Station, Fisheries and Oceans Canada

Chair of Oral Examination

• Dr. Caetano Dorea, Department of Civil Engineering, UVic

Abstract

As marine ecosystems face rapid change, there is a growing need for accurate and efficient tools to support long-term biodiversity monitoring. This thesis evaluates the application of computer vision (CV) methods for monitoring epipelagic macrofauna using unbaited remote underwater video collected via FishCams, a novel low-cost camera system developed for nearshore environments. A ten-step workflow was developed to generate an annotated object detection training dataset, which included salient motion filtering, image classification, bounding box annotation, and iterative model training and validation. Using this method, we trained a YOLOv8 object detection model on over 240,000 annotations spanning 54 pseudo-taxonomic categories (representing both taxa and visually similar groups of fauna) from footage collected at two kelp farms and adjacent reference sites (N = 4 sites) on the west coast of Vancouver Island, British Columbia. Model performance was highest for species- and genus-level pseudo-taxa with distinctive morphology, while broad taxonomic groups exhibited lower accuracy and systematic misclassification. A re-training experiment found that increased annotation effort improves model accuracy, though performance improvements were more with taxonomic resolution (e.g., species-, genus-, or family-level) than the type of organism (e.g., fishes, gelatinous zooplankton).

Another goal of this thesis was to compare census methods across sites over a seven-month period. CV-assisted monitoring (YOLOv8 object detection model) results were compared to conventional census approaches (diver transects, Remotely Operated Vehicle (ROV) surveys, and manual annotation of FishCam footage) across seven months of sampling. The CV method that included training annotations (YOLO With Annotations) consistently recorded the highest taxonomic richness (34 unique pseudo-taxa), substantially exceeding that of ROV (17), dive (23), and the FishCam Subset (22). Monthly richness comparisons indicate that sampling frequency likely plays a substantial role in pelagic biodiversity detection; hourly sampling methods (YOLO With Annotations, and YOLO without annotations (YOLO Only)) captured higher richness than the low-frequency FishCam Subset, despite using the same camera system. Proportionally, YOLO-based methods and dive surveys observed a broader range of fish and gelatinous pseudo-taxa, while ROV surveys primarily detected small gelatinous zooplankton. Community composition varied significantly by both site (farm versus reference) and census method. Across all methods, richness was significantly higher at the kelp farm than the reference site (p = 0.024), driven by an increased proportion of solitary fish pseudo-taxa at the farm site. Together, our results suggest that CV-assisted underwater video monitoring has the potential to outperform dive and ROV surveys in an epipelagic environment, though differing sampling frequencies prevent equal effort comparisons across methods.