Abstract
Norwegian Atlantic salmon aquaculture produced 1.5 million metric tons of fish in 2024. Such large production comes with its challenges, and the industry experienced a loss of 57.8 million fish (15.4 %) during the sea phase that year. In response, the Norwegian government has set the ambitious goal to reduce the sea phase mortality to 5% within the next decade. Achieving this goal requires substantial improvements in fish health and welfare, beginning with the development of reliable and representative monitoring methods. Assessing welfare at commercial scale is inherently challenging, with up to 200 000 salmon inhabiting a single sea cage of up to 40 000 m³. Current welfare assessments rely on the monitoring of Operational Welfare Indicators (OWIs) informing on whether the fish welfare needs, such as adequate nutrition, appropriate water quality, good health, behavioral freedom, and safety, are met.
In this doctoral work, a selection of OWIs was evaluated to monitor salmon health, growth, loser fish prevalence, and salmon behavior, using non-invasive methodologies at commercial
scale. It is now possible to monitor salmon directly underwater with the help of smart farming technologies, such as computer vision to assess physical health, stereovision to measure the fish size, and sonar to monitor salmon group behavior, preventing potential harm caused by handling and increasing the number of individuals assessed. However, even if these technologies provide precise results, their accuracy might highly depend on the placement of the monitoring equipment, since the whole cage can rarely be assessed as a whole and salmon might be stratified based on their size or health status, or might not use the space in the cage homogeneously. In this doctoral thesis, some non-invasive operational methodologies were explored for more representative assessments of fish welfare in commercial sea cages.
Evidence for vertical stratification in size and health was found, highlighting the limitations of relying solely on automatic sensor measurements without carefully considering sensor placement within the sea cage. Furthermore, three computer vision models were created and compared on the ability to differentiate loser fish from healthy individuals using readily available machine learning techniques, enabling to create a low-cost monitoring solution for the prevalence of loser fish in commercial sea cages. Finally, a mechanical 360-degree singlebeam scanning sonar was evaluated and demonstrated significant value in monitoring salmon group behavior, providing detailed insights into spatial usage across both vertical and horizontal planes.
All together, these findings highlight the potential of new technologies for assessing fish welfare in commercial sea cages and provide indications on how to use them in a way to increase representativity. With the replication of such studies, more targeted guidelines for effective welfare monitoring in commercial sea cages using new technologies can be developed. With robust and representative welfare assessments, evidence-based management decisions aimed at maintaining and enhancing fish health and welfare can be made, increasing the quality of life of the fish while enhancing production efficiency.