In the rapidly evolving field of aquaculture research, ensuring fish welfare is paramount. Traditional methods for monitoring fish behavior and welfare are often labor-intensive and time-consuming, limiting their utility in both experimental and commercial settings. However, an innovative study published in Scientific Reports presents a novel approach that leverages deep learning and ecological concepts to automate and enhance fish welfare monitoring.
The study, conducted by researchers from Nofima AS and Universidad de Las Américas, delves into an innovative methodology (deep learning) and its potential implications for aquaculture research.
The challenge of monitoring fish welfare
Fish welfare in aquaculture is generally assessed using a combination of outcome-based (biotic) and input-based (abiotic) welfare indicators. Behavioral indicators, while highly informative, are particularly difficult to quantify due to the manual and time-consuming nature of data collection.
The European Directive 2010/63/EU emphasizes the importance of refining and reducing the impact of experimental conditions on animal welfare, further highlighting the need for efficient monitoring tools.
Digitalization offers automated tools to quantify fish behavior, providing a broader and faster view of behavioral changes. Computer vision (CV) and artificial intelligence (AI) are transforming various fields, including fish research and aquaculture.
A new approach: Deep Learning and Distribution Area Analysis
The study, led by Santhosh K. Kumaran of Nofima AS, introduces a cutting-edge tool that combines deep learning techniques with the ecological concept of home range to monitor the spatial distribution of Atlantic salmon parr (Salmo salar L.) in experimental tanks.
The tool adapts the DeepLabCut framework, a computer vision and machine learning platform, to estimate fish posture and quantify their spatial distribution using metrics inspired by home range and core area concepts.
The “Home Range” concept and its application
Home range is an ecological concept defined as the spatial area an individual occupies during daily activities. This concept, along with core area (the area where an individual spends 50% of its time), can provide valuable insights into how fish interact with their environment.
A reduction in home range size may indicate stress, disease, or issues with rearing system design.
Key components of the methodology
- DeepLabCut Framework: The study employs DeepLabCut to identify key body parts (key points) of fish. This deep learning model, based on the ResNet50 architecture, was trained on 674 annotated frames and tested on 36 frames, achieving over 90% accuracy for most key points.
- Home Range and Core Area Metrics: Fish spatial distribution is analyzed using home range (the area where fish spend 95% of their time) and core area (the area where fish spend 50% of their time). These metrics are derived from the probability density function of fish distribution, estimated using Gaussian Kernel Density Estimation (KDE).
- Behavioral Variability Around Feeding: The study examines fish behavioral changes before, during, and after feeding. The analysis reveals distinct spatial distribution patterns during feeding, with metrics such as Water Inlet Preference (WIP) and Core Area Centrality (CAC) showing significant variations.
Findings and implications
The study demonstrates that the proposed methodology effectively captures differences in fish behavior across different tanks and feeding periods. Key findings include:
- Water Inlet Preference: Fish consistently preferred areas approximately 90° downstream from the water inlet, suggesting a potential link between water flow dynamics and fish distribution.
- Core Area Utilization: Fish occupied only 13-17% of the tank 50% of the time, indicating a tendency to swim together rather than disperse uniformly.
- Behavioral Changes During Feeding: During feeding, fish exhibited a more uniform distribution and lower group cohesion, likely due to foraging behavior.
Advantages of the new tool for salmon farmers
The integration of deep learning and home range analysis offers several advantages:
- Automation: The tool automates the fish behavior monitoring process, reducing the need for manual observation and minimizing human error.
- Real-Time Monitoring: Continuous 24/7 monitoring provides real-time data, enabling rapid intervention in case of abnormal behaviors.
- Non-Invasive: The method is non-invasive, reducing stress and potential harm to fish.
- Scalability: While the current study focuses on small-scale experimental settings, the methodology has the potential to scale up for larger aquaculture systems.
Limitations and future directions
Despite its promising results, the study acknowledges several limitations, including lack of depth information and the impact of camera angle and lighting on key point detection.
Future research should explore the use of stereo cameras to improve depth perception and investigate the scalability of the methodology for larger fish populations and more complex environments.
Conclusion
Kumaran’s study represents a significant advancement in fish welfare monitoring. By combining deep learning with ecological concepts, the proposed tool offers a powerful and efficient means to assess fish behavior and welfare in aquaculture research.
The study’s findings highlight the potential of computer vision and home range analysis for monitoring fish behavior. Automating this process can provide valuable insights into how fish interact with their rearing environment, contributing to the development of technologies for measuring and monitoring fish behavior in future research.
The study was funded by the DigitalAqua Project and the ML4AKVA Project by NOFIMA.
Contact
Santhosh K. Kumaran
Nofima AS,
Ås, Norway
Email: santhosh.kumaran@nofima.no
Reference (open access)
Kumaran, S.K., Solberg, L.E., Izquierdo-Gomez, D. et al. Applying deep learning and the ecological home range concept to document the spatial distribution of Atlantic salmon parr (Salmo salar L.) in experimental tanks. Sci Rep 15, 5976 (2025). https://doi.org/10.1038/s41598-025-90118-9
