Model Purpose: An RF-DETR (Roboflow DEtection TRansformer) model was trained and deployed to analyze European lobster (Homarus gammarus) populations from videos recorded in shallow (8-25m) underwater habitats in Western Sweden. This study addressed critical limitations in traditional trap-based monitoring by providing non-invasive assessment of lobster abundance, size distribution, and behavioral patterns, particularly focusing on individuals too large to enter conventional lobster pots. Taxonomic Scope: The model was trained to identify: Homarus gammarus (European lobster) Model input: The video material consisted of continuous recordings from baited camera systems (Mobius ActionCams) positioned 2m above lobster pots, with recording periods ranging from 24 hours (Kåvra MPA) to 120 hours (Stora Kornö). Footage collected by Peter Ljungberg.Temporal and Geographic scope: The model was trained and validated on footage collected in August 2022 (Kåvra MPA) and August 2024 (Stora Kornö), two contrasting management regimes with varied substrates/turbidity: Kåvra Marine Protected Area (fishing prohibited since 1989): 21 stations, 8-25m depth Stora Kornö (commercially fished area): 9 stations, 13-22m depth Camera setup: Mobius ActionCams with wide-angle lenses (132° FOV) mounted above Carapax brand lobster pots, baited with thawed herring. Video specifications: 720p at 10 FPS (2022) and 720p at 25 FPS (2024). Platforms used: Annotation: Roboflow (polygon-based instance segmentation) Model Training & Evaluation: Roboflow and SUBSIM (Swedish platform for subsea image analysis) Performance Analysis: Weights & Biases Model performance: Mean Average Precision (mAP): 89.8% Precision: 87.1% Recall: 93%

Files included: model_dataset: contains the trained YOLOv11 model and the dataset of 932 images, split into train (76%), test (13%), and validation (15%) subsets. model_metrics: contains graphs (curves) showing model accuracy and efficiency (for detailed explanations, visit YOLO Performance Metrics) model_performance: contains example tracking videos demonstrating the model's performance on footage with either single or multiple lobsters. An additional batch example of the frames used in the training is included.