Date of Award
2026
Degree Type
Thesis
Degree Name
Master of Science in Oceanography
Department
Oceanography
First Advisor
Christopher Roman
Abstract
Effective monitoring of benthic ecosystems requires scalable approaches capable of simultaneously identifying key species and quantifying habitat composition. In Narragansett Bay Rhode Island, bay scallops, Argopecten irradians, serve as important ecological indicators, yet traditional survey methods remain labor-intensive and spatially limited. This study presents an automated deep learning framework for scallop detection and habitat segmentation using underwater benthic imagery.
The proposed system combines three components: (1) YOLO based object detection for identifying live scallops, (2) U-Net semantic segmentation for quantifying benthic habitat classes (eelgrass and sand/silt), and (3) a geospatial contrastive learning framework (GeoCLR) to leverage spatial continuity in imagery and enhance feature representation under limited labeled data. Detection performance was evaluated using precision, recall, and mean Average Precision (mAP) statistics, while segmentation performance was assessed using Intersection over Union (IoU) and pixel accuracy.
Results demonstrate that a single class detection strategy with structured augmentation significantly improves scallop detection performance, while habitat segmentation enables quantitative estimation of benthic composition at the image level. The integration of detection and segmentation outputs further allows spatial analysis of scallop occurrence relative to habitat type.
This framework provides a scalable, automated approach for habitat informed scallop monitoring and establishes a foundation for broader benthic ecosystem assessment using machine learning.
Recommended Citation
Odubayo, Waheed, "ENHANCING BAY SCALLOP MONITORING USING MACHINE LEARNING AND SEMANTIC SEGMENTATION WITH BENTHIC VISUAL IMAGERY" (2026). Open Access Master's Theses. Paper 2719.
https://digitalcommons.uri.edu/theses/2719