Monitoring monthly mortality of maricultured Atlantic salmon (Salmo salar L.) in Scotland I. Dynamic linear models at production cycle level

The mortality of Atlantic salmon is one of the main challenges to achieving its sustainable production. This sector benefits from generating many data, some of which are collated in a standardized way, on a monthly basis at site level, and are accessible to the public. This continuously updated resource might provide opportunities to monitor mortality and prompt producers and inspectors to further investigate when mortality is higher than expected. This study aimed to use the available open-source data to develop production cycle level dynamic linear models (DLMs) for monitoring monthly mortality of maricultured Atlantic salmon in Scotland. To achieve this, several production cycle level DLMs were created: one univariate DLM that includes just mortality; and various multivariate DLMs that include mortality and different combinations of environmental variables. While environmental information is not collated in a standardized way across all sites, open-source remote-sensed satellite resources provide continuous, standardized estimates. By combining environmental and mortality data, we seek to investigate whether adding environmental variables enhanced the estimates of mortality, and if so, which variables were most informative in this respect. The multivariate model performed better than the univariate DLM (P = .004), with salinity as the only significant contributor out of 12 environmental variables. Both models exhibited uncertainty related to the mortality estimates. Warnings were generated when any observation fell above the 95% credible interval. Approximately 30% of production cycles and more than 50% of sites experienced at least one warning between 2015 and 2020. Occurrences of these warnings were non-uniformly distributed across space and time, with the majority happening in the summer and autumn months. Recommendations for model improvement include employing shorter time periods for data aggregation, such as weekly instead of on a monthly basis. Furthermore, developing a model that takes hierarchical relationships into account could offer a promising approach.