| dc.description.abstract | The modern aquaculture industry suffers from low observability and control of conditions in the aquaculture cages. In order to better understand fish behaviour in both normal and stressful conditions, more knowledge of how fish behave is needed. The purpose of this project was to use machine learning methods to identify different modes of individual salmon behaviour with positional data from acoustic telemetry tags. Positional data from six different fish from two different cages at two different times of year was processed in order to create discrete fish swimming trajectories. Additional variables were calculated based on positional data: average depth, depth difference per second, track length per second, angle change per second, average distance from cage center per second and distance moved in relation to cage center. Every trajectory had one value for each of these variables. These trajectories were analysed based on both traditional methods and principal component analysis, in addition to being clustered with the HDBScan algorithm. In general, fish were more active at day and swam closer to the surface at night, and this was the case for every fish except one. Average depth distribution was the variable that differed most betweeen individuals. The variables that differed the least between individuals were the two variables based on distance from center, and these did not contribute to the clustering. Correlation structure at night was somewhat similar for most fish, as the variables average depth, depth difference, and track length were more correlated at night. Multiple modes of behaviour were detected, including feeding, circular swimming, and third mode consisting of idle, non-circular swimming. For most fish, circular swimming was the most prevalent behavioural mode in the day, while short, high angle change, non-circular, idle trajectories close to the surface was the dominating swimming pattern at night. The results show that salmon behaviour has definable modes that can be detected from positional data, and that the prevalence of these modes differs from night to day. Moreover, the results show the potential of the application of machine learning methods in aquaculture. | |
