Aquaculture systems are complex and it is critical to combine comprehensive statistical and data-driven techniques with deep domain expertise to inform analysis. In this section, we present a variety of statistical (boxplot, density plots, correlation and summary metrics) and machine learning (feature explained, accumulated local effects) analysis to help inform drivers of caged salmon behaviour.

Boxplot analysis of the data can be used to summarise key points about the farm. Firstly, despite the deep waters (approximately 60m), and the large depth of the cage (27m), fish tended to congregate in the upper third, and spent most of the time at depths of 3-9m. The box plot does not show any pronounced daily pattern. However, it is worth highlighting that the northerly latitude of the site (67 N) means it is characterised by 24-hour sunlight for most of the summer months which likely impeded the development of daily patterns during some of the periods. Further, fish in the cage was changed between 26-29th July and new stock was introduced, which naturally modifies recurrent patterns of behaviour. These factors are likely contributory to observed conditions and more pronounced diurnal patterns were observed at other sites.

Feature importance metrics are extremely valuable for the interpretation and fine-tuning of any machine learning experiment. It allows us to quantify which features are the most important contributors (and also to eliminate superfluous features that do not contribute significantly) and provides a valuable tool to explain model results. Further, the temporal evolution of feature importance as provided by Accumulated Local Effects can be extremely informative to guide analysis in the marine environment with complex spatial and temporal variabilities.

Read more in our paper!