OceanCV/montereybaycoastalfauna

Detecting Coastal Fauna in Monterey Bay

By Christian Mendiola

Scientific Context

Avian and marine mammal populations are an integral part of the coastal ecosystem and maintaining its health. Monitoring of these fauna can help identify changes in the ecosystem over time.

This project aims to create a model capable of measuring avian and marine mammal populations utilizing a 24/7 coastal livecam previously not utilized for science.

---

Dataset Description

The Monterey Bay Live Cam is a 24/7 live camera of Monterey Bay, provided by MBARI. The camera is set to pan to a multitude of different views of the Monterey Bay coast with a random zoom setting, from which we can observe several objects of interest, such as birds, marine mammals, and ships. The dataset used in this study was obtained by screen recording the Monterey Bay Live Cam livestream on YouTube. To achieve a balance of ease of computation and sample size, I have collected 4 hours of footage from 10am - 2pm and have sped it up 10x to both train and assess the model.

---

Model Selection

In this study, I decided to use YOLOv11 object detection model in order to identify and locate classes with the least time and computational cost as I chose a very large dataset to process. Using Roboflow, I applied augmentations to account for the high level of variability possible in animal shapes, size, and location. These augmentations include:

• Clockwise
• Counter-Clockwise
• Upside Down
• Crop
• Rotation
• Brightness
• Blur

---

Classes

When annotating images, I specified a total of 5 classes:

• Avians (count: 1882)
• Pinnipeds (count: 243)
• Boats (count: 4)
• Mustelids (count: 1)
• Humans (count: 21)

---

Model Assessment

Confusion Matrix

Normalized Confusion Matrix

Precision-Confidence Curve

Precision-Recall Curve

Recall Confidence Curve

Interpretation of Results

The model seems to exhibit certain strengths and weaknesses. For example, it is able to correctly identify many classes with decent accuracy, but also has a strong tendency to misidentify objects in the background (such as rocks and waves) as fauna. Additionally, it doesn't have perfect recall as it also has good chance to not count visibly present fauna such as avians and pinnipeds. In addition, the model did not identify any mustelids or boats during testing, which means that it would likely be unable to do so if deployed as is. Overall, a model of this size will need more rigorous training protocols (higher amount of images for training and validation), and will likely need to be deployed using video that does not change as frequently and randomly as the Monterey Bay Live Cam. As for its strengths, the model is excellent at detecting avians perched on the coast, even in cases where there is a high density of them concentrated in an area. Additionally, the metrics of the model suggest that it has a good balanced between Precision and Recall (see Precision-Recall Curve).

---

Model Use Case

In its current state, the model will need to be used on data with a primary focus on avians with a fixed view where they can easily be detected. With these modifications, the model can be deployed and used for further studies regarding specific densities of the fauna observed in Monterey Bay.

Example Study: In the worlds current state, the Earth's climate is warming. A warming climate has been linked to seabird die-offs as they lose habitat and food availability. Though, the effect of warming on seabirds is very likely not equal in all parts of the world. Monterey Bay is a hotspot for seabirds due to its high biodiversity and nutrient richness. Loss in seabird life here will likely indicate the decline of the entire ecosystem, which incentivizes further efforts to monitor seabird populations here.

• Hypothesis: The amount of avians on the Monterey Bay coast will decrease as time passes as food availability is reduced as a product of warming climate.
• Methods: Model run on footage of the Monterey Bay coast utilizing track and count for several years. The model will sum the amount of birds counted in one year. After many years, data analysis can be performed to assess trends in avian population.

---

Disclaimer

All of the data used in this study is property of Monterey Bay Aquarium Research Institute. The Monterey Bay Live cam can be found at https://www.youtube.com/watch?v=fVa6-zCBR7A.