Precision farming in aquaculture: assessing gill health in Atlantic salmon (Salmo salar) using a non-invasive, AI-driven behavioural monitoring approach in commercial farms

Abstract

As the aquaculture industry grows, more sophisticated technology is required to monitor farms and ensure good fish welfare, in line with the precision livestock farming concept. Using behaviour as a non-invasive monitoring tool, combined with artificial intelligence, enables greater control over farm management. This study aimed to assess temporal changes in farmed Atlantic salmon (Salmo salar) group behavioural profiles related to fish health and welfare. A machine vision algorithm applied to feed cameras on commercial farms was used to determine whether changes in gill health would induce visible group behavioural changes. Video cameras were deployed in all cages at two Scottish Atlantic salmon marine farms. One cage at each farm was also equipped with additional cameras (5 and 4 at sites A and B, respectively) to provide higher spatial coverage of fish behaviour and distribution. The algorithm processed video footage from these cameras and produced behavioural data termed 'activity' (%), which encompasses fish abundance, speed, and shoal cohesion. Additionally, gill health, Operational Welfare Indicators (OWI), mortality, and Specific Feeding Rate (SFR) were scored weekly at both sites. During summer 2023, gill health issues arose at both farms, leading to fish stress reflected in the behavioural data. For two months prior to the onset of poor gill health, the average (± standard deviation) activity levels of the fish across all cages were 25.6 ± 10.5% and 24.9 ± 7.0% for Farm A and B, respectively. After gill health was compromised, the activity rose significantly for two months in all cages with a mean of 43.6 ± 15.1% and 32.6 ± 9.6%, respectively. A generalised linear mixed model revealed that Proliferative Gill Disease (PGD) was the main driver of this increase in activity. This increase in activity coincided with fish migration to the centre of the cage, meaning tighter shoaling, which is a normal stress response often seen in relation to predators and other environmental or health stressors. The use of behaviour as a non-invasive welfare indicator and the potential to use artificial intelligence to automate the process of behavioural identification allows farmers to improve welfare conditions and ensure industry sustainability.

Supplementary information: The online version contains supplementary material available at 10.1186/s44365-025-00020-8.

Keywords: Aquaculture; Atlantic salmon; Fish behaviour; Gill health; Machine learning; Precision farming; Welfare.

Fig. 1

Overview of the study site. A Map of Scotland showing the location of the of the Atlantic salmon aquaculture farms on the Western Isles. Satellite image providing an aerial view of the Farm A (B) and Farm B (D; from Google Earth) with study cages outlined in red. C Layout illustrating the placement of cameras within the two study cages

Fig. 2

Comparative figure for Farm A (left) an Farm B (Right) with: depth of maximum activity where colours indicate the hourly-averaged amount of activity present at that depth (A,E), weekly gill health scores (0 = healthy gills, 5 = necrotic tissue), where red indicates Proliferative Gill Disease (PGD) and blue indicates Amoebic Gill Disease (AGD; B,F), daily mortality as a percent loss within the cage (C,G), and hourly temperature at 8 m depth (D,H)

Fig. 3

The model II linear regression for weekly activity and PGD scores for Farms A (orange) and B (blue) with confidence intervals (shaded region)

Fig. 4

The health data from each cage in Farm A (left) and Farm B (right) including: The weekly PGD scores (A. E), the daily activity of the fish (B, F), the daily specific feeding rate (SFR; C, G), and the daily mortality (D, H). The colours indicate average 2 months pre- (white) and post- (grey) increase in activity (July 7 [Farm A]; August 18 [Farm B]). Different letters denote significant differences within and among cages per farm