Object detection model for epibenthic species in low-trophic aquacultures (LTA): A case study on the Swedish west coast

Model purpose: A YOLOv8 model was trained and employed to extract species observations from videos recorded on the Swedish west coast. The study was part of the COOL BLUE project and focused on determining the relative abundance and species composition of mobile epibenthic fauna in low-trophic aquacultures (LTA), particularly associated with the cultivation of blue mussels (Mytilus edulis).

Taxonomic scope: The model was trained to identify the following taxa:

1. Gastropoda sp. (Class)
2. Gobiidae spp. (Family; Gobies)
3. Hyas araneus (Species; Spider King Crab)
4. Paguridae sp, (Family; Hermit Crab)
5. Pleuronectiformes sp. (Order; Flatfish)
6. Portunoidea spp. (Superfamily; Crabs)* 

Model input: The video material included 30 minutes-recordings using a baited remote underwater video system (BRUV).

Temporal scope: The model was trained and used on movies collected in 2024 (February to May).

Geographic scope: The model was trained on footage from mussel aquacultures along the Swedish west coast. Two different farms were sampled: one at ~8 m depth and the other at ~18 m depth. 

BRUV Setup: A downward-facing BRUV was built for this study.

• A GoPro 9 was used, a light was mounted alongside the camera, facing the same direction, and a black mat was placed at the bottom of the BRUV to enhance the background for optimal and standardized conditions to train a model.

Platforms used: 

1. Annotation: Roboflow
2. Model Training & Evaluation: SUBSIM (Swedish platform for subsea image analysis)

Files included:

• model_dataset: contains the trained YOLOv8 model and the dataset of 910 images, split into train (70%), test (15%), and validation (15%) subsets.

• model_performance: contains the performance of the model on both seen footage (videos used for training) and unseen footage (videos from an older setup the model analyzed for the first time).

• model_metrics: contains graphs showing model accuracy and efficiency (for detailed explanations, visit YOLO Performance Metrics).