Puget Sound Phytoplankton Detection Model
by Christopher Moon
This model was created for OCEAN 220 at the University of Washington School of Oceanography.
Scientific Context
Phytoplankton are the heart of any marine food web. Technologies such as FlowCams and IFCBs are able to collect thousands of images from a single sample. This creates large datasets requiring annotation and analysis. Manual annoation is time consuming and not cost efficient, leading to the need for computer vision models to count phytoplankton. Being able to automatically detect phytoplankton will allow scientist to spend their valuable time towards analyzing phytoplankton species abundance. If this model were deployed in-situ, it could allow scientists to track phytoplankton community dynamics in rreal time.
Dataset description
Images were taken using an Echo Rebel digital microscope. The slides were created from samples collected on the Ocean 220 cruise aboard the R/V Rachel Carson in Puget Sound. Additional images were added from WHOI IFCB dataset.
Classes:
- Chaetoceros - 152
- Coscinodiscus - 44
- Detonula Pumila - 17
- Eucampia - 27
- Odontella - 3
- Skeletonema - 221
- Stephanopyxis - 3
- Thallasiosira - 111
Before augmentations:
- 246 images
- 534 annotations
- 6 classes
- median image ration:2650x1998
- average image size: 5.29 mp
Model Selection
I used Ultralytics Yolov11 because its reported metrics for object detection are greater than YOLOv12 . I used object detection because it would allow me to both collect data on population numbers for multiple classes. #
Model implementation - code sample
model = YOLO('yolon.pt')
#path to Yaml
dataset_config = '/content/Dataset/data.yaml'
#Train the model
results = model.train(
data=dataset_config,
epochs=75,
batch=-1,
imgsz=640,
plots=True,
patience=50,
resume=True
)
print(results)
Model Assesment
F1-confidence curves shows us the ideal balance between precision and recall and how that changes with confidence. My model shows great F1 scores for many of the classes but struggles especially with Thallasiosira
Normalized Confusion Matrix shows the percentage of true predictions against all of the other classes in the dataset.
Model Use Case
This model is too be used by scientist with large phytoplankton datasets that need to be annotated.
Disclaimer
While this model has great metrics, some classes do struggle. Please check the output images before using in study.