MSc Oceanography thesis · University of Rhode Island, 2026 · with Rhode Island DEM
Bay Scallop Monitoring with Computer Vision
The problem
Bay scallops in Point Judith Pond, Rhode Island, were once a 100,000+ lb/year fishery; today numbers are too low to sustain one, and decades of restoration have had limited success. Monitoring relies on scuba surveys and manual image review — slow, small-scale, and hard to repeat. And because scallop survival is closely tied to eelgrass, counting scallops without measuring habitat tells only half the story.
What it revealed
Scallops occurred most often in moderate eelgrass cover (30–60%)— less in sparse or dense beds. That fits a balance between protection from predators (lacking in sparse beds) and water exchange and oxygen (reduced in dense ones): habitatdensity matters, not just presence.
What I built
An end-to-end deep learning pipeline that finds scallops and measures the habitat around them, from seafloor imagery collected by an autonomous surface vehicle along RI DEM-selected transects.
Detection
YOLOv5 detector for live scallops — small, camouflaged, often half-buried.
Habitat segmentation
U-Net classifying every pixel as eelgrass or sand/silt, giving each image a habitat composition.
Representation learning
GeoCLR — contrastive learning that uses each image’s GPS position to learn habitat structure from mostly unlabeled imagery.
Integration
Detections merged with habitat masks to analyze where scallops occur relative to habitat.
Engineering decisions
- Single-class over multi-class
- Live, dead, and face-up shell classes were too sparsely labeled; focusing on live scallops concentrated the model on what matters.
- Structured augmentation for underwater light
- CLAHE contrast correction, HSV color-space transforms, and rotations to counter strobe hotspots, dark corners, and color loss — applied to training data only.
- Spatially independent evaluation
- Entire survey lines held out, so results reflect new locations, not memorized ones.
- Right-sized models
- YOLOv5 generalized better than newer, larger variants on a small, specialized dataset.
- Bias mitigation
- Identified and corrected systematic eelgrass overestimation from vignetting (15% edge crop) and camera viewing geometry.
Tools
- Python
- OpenCV
- MATLAB
- CVAT
- YOLOv5
- U-Net
- GeoCLR
