Ochre Sea Star Object Detection Model
This project presents a computer vision model for detecting and counting ochre sea stars (Pisaster ochraceus) in intertidal imagery. It is designed to identify sea stars in still images and can also be applied to video frames to estimate counts over time.
Ochre sea stars are an ecologically important species along the west coast of North America. As a keystone predator, Pisaster ochraceus helps regulate mussel populations and thereby supports biodiversity in rocky intertidal communities. With experiencing population decline (Sea Star Wasting Syndrome (SSWS)), which has caused large-scale mortality and broader ecological shifts, including mussel bed expansion and altered community structure, monitoring this species population is considered essentially important. The motivation for this model is to support faster and more frequent monitoring of ochre sea star abundance estimation using images and videos, expanding the spatial and temporal scale of observations while reducing labor demands for ecological surveys.
Model Details & Selection
This model was trained using YOLOv11 for object detection.
Object detection was selected because the primary ecological goal is to identify and count individual sea stars within images or video frames while also preserving their approximate locations in the scene. With meeting the main population-monitoring objective, object detection features useful ecological output for abundance estimation, simpler annotation and implementation. The project workflow involved iterative model development: adding annotations, applying augmentations, making targeted improvements, testing on sample videos, and then adding a counting function for practical use. This iterative approach allowed performance and usability to be improved across multiple model versions.
Dataset
Exported/Edited Image Resolution:
Preprocesssing: Auto-orient, resize to 512 x 512
Augmentations:
- horizontal flip
- vertical flip
- rotation (15°)
- grayscale conversion 25%
Dataset size after Augmentations: 257
| Class Name | Number of Annotations |
|---|---|
| Pisaster-ochraceus | 113 |
Examples of the prediction of version 8 on the vallid batch.
Dataset Development
The model was developed iteratively across multiple versions, with later versions using more training data and more augmentation than earlier versions.
Examples from the project workflow include:
- V6: 53 train / 7 valid
- V7: 159 train / 7 valid
- V8: 257 train / 8 valid
Model Performance/Assessment
Model performance was evaluated primarily using F1-confidence curves and confusion matrices, consistent with the project’s goal of assessing how reliably the detector can identify ochre sea stars while limiting false detections. Across model versions, performance improved as the training dataset increased in size and diversity and as augmentation was added, later versions showing stronger overall behavior and more practical performance on sample videos.
Version Comparison
V6 achieved the highest peak F1 (0.97) and was the most sensitive model, performing best at a low confidence threshold. V7 also performed strongly (0.95) and showed a good balance between detection strength and confidence stability. V8 had a slightly lower peak F1 (0.85) but operated at a higher optimal confidence threshold, making it the most conservative of the three. In summary, V6 had the strongest peak performance, V7 was the most balanced, and V8 was the strictest in prediction confidence.
Evaluation Summary
- Earlier versions were trained on relatively small datasets.
- Later versions incorporated more training images and augmentation.
- Version 8 was used for validation plots and multiple sample video applications.
- Example testing thresholds included:
conf=0.05,iou=0.8conf=0.2,iou=0.8conf=0.2,iou=0.65
Model Potential Use Case
Practical use-case for this model includes comparing ochre sea star abundance across locations or through time using repeated intertidal imagery. The model could be applied to images from fixed shoreline cameras, repeat field photographs, or drone-based coastal surveys to estimate relative star density in different habitats or management areas. Hypothesis like ochre sea star density will decline less rapidly through time in protected sites than in comparable unprotected sites due to reduced human disturbance and stronger ecosystem integrity may support greater resilience could be proposed. Studies supporting the paper of Monaco et al. (2016) result, as well as monitoring sea star abundance helps us understand coastal ecosystem resilience to climate change could be conducted.
Limitations
- relatively small dataset size
- only one main target class
- limited metadata documentation in the current project version
- possible undercounting when sea stars are partly hidden
- reduced performance in complex or cluttered intertidal scenes
- no direct assessment of sea star health, wasting symptoms, or morphology
- no species differentiation among multiple sea star species
Future Improvements
Future extensions of this project could include:
- adding more sea star species as classes
- improving species identification in mixed-species scenes
- expanding the training dataset across more locations and conditions
- testing on more field videos and fixed-camera imagery
- moving from object detection to instance segmentation
- using morphology-based analysis to support health or disease studies
References / Citation
iNaturalist. Public species observations and imagery used in dataset development.
Course materials from OCEAN 462C: Computer Vision Across the Marine Sciences. [https://oceancv.org/]
[Hakai Wild: Ochre Sea Star]https://youtu.be/PfEz7yE4m0U?si=3F_NfYqvF-TBCyL7
Monaco, C. J., Wethey, D. S., & Helmuth, B. (2016). Thermal sensitivity and the role of behavior in driving an intertidal predator–prey interaction. Ecological Monographs, 86(4), 429–447.