Four work packages, one welfare-monitoring pipeline

Research focus

• Semi-automatic labelling and supervised baselines for fish, fish parts, and known welfare indicators
• Unsupervised learning of recurring semantic units such as eyes, snout, fins, and gills
• Detection of wounds, deformation, lice, eye or fin damage, and other deviations from normal appearance
• Fish-behaviour analysis using motion tracking and pattern recognition
• Robustness to turbidity, marine snow, caustics, motion blur, back-scattering, and occlusion

Core research question

How can self-learning algorithms detect welfare indicators and learn high-level salmon features from large volumes of unlabelled video?

Expected contribution

Annotation-efficient methods that quantify appearance-based welfare indicators, identify potential rare welfare issues as outliers, and provide fish-part and behaviour information to WP2 and WP3.

WP1 datasets → · Related theses →

Research focus

• Evaluate melanin-spot approaches and other established biometric baselines
• Learn texture-oriented representations using machine learning, graph theory, and random-field concepts
• Combine spot patterns with gill-cover, jaw, fin, and other visual characteristics
• Achieve robustness to growth, viewpoint, scale, perspective, camera, and changing conditions
• Maintain individual fish-health journals combining identity, WP1 welfare and behaviour observations, and previous health status

Core research question

How can advanced visual biometric features enable robust one-to-many re-identification in populations that may contain up to 200,000 salmon?

Expected contribution

Highly accurate, long-term identity descriptors and a continuously updated health record for each observed fish, linking repeated visual observations across time.

WP2 dataset ↗ · Related theses →

Research focus

• Represent fish-welfare observations together with environmental and operational context
• Integrate historical time series, sensor measurements, observations, and operational logs
• Use a domain model, graph-based representation, and inference rules to connect heterogeneous data
• Build a fish-welfare journal that supports interpretation of welfare status over time
• Demonstrate welfare-score reasoning and forecasting using available project data

Core research question

How can heterogeneous welfare observations and contextual data be represented and reasoned over to improve understanding of fish welfare over time?

Expected contribution

Models that reveal relationships among welfare observations, operational factors, and environmental conditions, providing a structured basis for interpretation and future decision support. The published WP3 study demonstrates this through an integrated domain representation, inference rules, a fish-welfare journal, and a predictive example.

Research focus

• Evaluate methods using realistic underwater recordings and aquaculture data
• Assess the accuracy and robustness of welfare-indicator detection from WP1
• Assess long-term re-identification and fish-health journal creation from WP2
• Examine whether the relationships and interpretations developed in WP3 are methodologically supported
• Seek biological and domain-expert confirmation that detected indicators are relevant and useful for welfare assessment

Core research question

How well do the proposed methods perform under realistic conditions, and are their outputs meaningful for interpreting fish welfare?

Expected contribution

Evidence of technical performance, robustness, and practical usefulness, supported where possible by biological expertise confirming that the detected welfare indicators are meaningful.