Context and Need:

This object detection model has been created for the purpose of being able to identify glass eels in marine enviorments. This model aims to provide scientists with a way to monitor juvenile eels as they are still not fully understood.

Dataset:

The dataset used is available in the files and versions under the name ONCCameraFootage.zip. This footage was used to train the model on detection of glass eels in low light enviorments. All of the data is taken from MP4 files from the ONC Cambridge Bay underwater network, annotated by hand, then had the following augments applied to them:

Horizontal and vertical flips
+- 10 Degree sheers horizontal and vertical
Grayscale applied to 15% of the images
Noise on up to 1% of pixels
The data that support the findings of this study are openly available from Ocean Networks Canada at

10.34943/72a02242-b629-45ee-80a5-6e633885d7d6 and 10.34943/8a1fecd8-882f-4604-86eb-6b0e7f217631 *

In this dataset there are 957 annoted instances of glass eels through varying yearly conditions.

Choice of Model:

When choosing a model I wanted to choose YoloV11 because at the time it was the most up to date Yolo model and performed the most accurately and quickly out of the family of models. I choose object detection because I wanted to have a model that would allow for counting of glass eels for scientific purposes as they tend to be hard to spot.

Implementation:

You will need to do the following to use the model

pip install ultralytics

model = YOLO(r"Glass_Eel.pt")  # Load model with proper path

results = model.predict(
  source=r"sourcefile", # The file that you want to use the model to annotate
  conf=0.25,
  iou=0.7,
  save=True,
  project=r"/save-directory",  # Custom save directory
  name="predictions"
    )

Metrics and Graphs:

Below is the F1 confidence curve for this model:

A F1 confidence curve shows the balance between precision and recall over the confidence of the model. The higher the score F1 score the better however when plotted over confidence we tend to like to see that it has a nice rounded shape as this model does. This model tended to beat out the odds as with marine snow it can be very hard to tell what is an eel and what isnt. From this graph you can articulate that both precision and recal performed well with low and high confidence levels.

Potential Study:

Hypothesis

Glass Eel migration patterns are influenced by seasonal environmental factors such as tides, temperature, and salinity, with peak migration occurring during specific lunar and tidal phases.

Justification

A YOLO-based detection model is well-suited for this study due to its ability to process underwater video data efficiently, enabling automated identification and tracking of Glass Eels in diverse and dynamic aquatic environments. This approach minimizes human error, ensures consistency in detection, and allows for large-scale data analysis over extended time periods.

Study Design

To investigate eel migration patterns, underwater cameras and remotely operated vehicles (ROVs) will be deployed at key estuarine migration points. Daily video recordings will be collected over one year to track eel movements, with the YOLO model used for automated detection and counting.

Key Research Questions

When do Glass Eels migrate in the highest numbers?
How do tides, temperature, and salinity affect migration?
Are there long-term changes in migration patterns over the study period?

Data Analysis

Eel Count Trends: Analyze seasonal and daily fluctuations in migration.
Environmental Correlation: Compare eel activity with environmental parameters such as tides, temperature, and moon phases.
Migration Pathways: Identify preferred routes and potential barriers to migration.

Impact & Applications

Conservation Planning: Provide data to protect critical migration corridors.
Fisheries Management: Inform sustainable harvesting practices and regulatory decisions.

OceanCV
/

Glass_Eel