A case study exploring physical health and size stratifications of Atlantic salmon (Salmo salar) with depth in a commercial sea cage

High mortality rates in Atlantic salmon (Salmo salar) aquaculture remain a critical issue in Norway, with 62.8 million deaths registered during the sea phase in 2024. Improving fish health is therefore a priority for the industry, both for ethical and productivity reasons. Current health assessments rely on invasive sampling of a small number of individuals, which may not accurately reflect the overall population in sea cages. Automated sensor technologies offer a promising, non-invasive alternative, but their representativity may depend on sensor placement if fish are, as we hypothesized, vertically stratified by health status or size. In this case study, stereovision rigs and an automated camera sensor were deployed at four depths (1-14 m) in a commercial sea cage over a two-month period following a delousing operation. Fish were measured and assessed for six health indicators following the Laksvel protocol. While the overall physical health of the fish was good, findings suggest a long-term vertical stratification of salmon with chronic deviations and a temporary stratification of salmon displaying injury-based deviations. These results highlight the importance of sensor depth for representative health and size monitoring and call for further studies across seasons and under varying environmental conditions.

In recent years, Atlantic salmon (Salmo salar) production in Norway has faced high mortality rates during the sea phase. In 2024 alone, 62.8 million salmon died on Norwegian fish farms, corresponding to an estimated cumulative mortality of 15.4% of the standing stock (Moldal et al., 2025). The Norwegian government aims to reduce this number to 5% within the next decade (regjeringen.no). Improving fish health and lowering the mortality rate accordingly is also a strategic priority for the industry not least for productivity reasons. At the same time, it addresses ethical concerns from consumers and is likely to contribute to the welfare of fish throughout their life.

There are different animal welfare definitions and welfare is not easily evaluated in a comprehensive way, particularly in the case of fish that live in an environment so different from our own (Stien et al., 2020). Fish farmers currently assess how well their fish are doing mostly by monitoring some indicators that are relatively straightforward to measure on site (“operational welfare indicators”; Noble et al., 2018). Most of these current measures are associated with fish physical health (Noble et al., 2018), and they have been labeled as outcome-based, as they describe the result or consequences of how well the animal has been coping with its environment (Stien et al., 2020). At the group level, such indicators consist of, for example, growth rate, mortality, and behavior (i.e. feeding behavior, swimming speed, schooling, distribution and use of space in the sea cage, etc.). At the individual level, the appearance of injuries or deformities can be observed (Nilsson et al., 2022; Noble et al., 2018; Stien et al., 2020). Since these indicators focus on a selection of aspects of an individual’s physical state, the more specific term “health indicators” was chosen in the present study rather than the more commonly used “welfare indicators”, which may imply the additional consideration of e.g. natural behavior and mental well-being (Fraser, 2008).

Currently, fish health is occasionally, but not systematically, assessed on some farms during salmon lice counting (20 fish per cage) by scoring individuals manually for external deviations such as wounds, eye, fin, or skin damage, emaciation, deformities, etc. A sea cage can contain up to 200,000 individuals, and such limited sampling may not be representative of the overall population in a sea cage. Moreover, the handling involved in this process, which implies crowding, catching, sedation, and out-of-water manipulation, is highly stressful to fish and can potentially result in some of the physical injuries included in the scoring scheme (Sommerset et al., 2024). As such, there is a great need for non-invasive methods to assess fish health and behavior and improve representativeness. This has driven the development of new technologies, and automated sensors are now available to continuously monitor fish status, count lice, and measure growth using underwater images or videos (Føre et al., 2018; Barreto et al., 2022).

However, it remains unclear whether the outcome of health scoring and size estimation using automated sensors are influenced by the vertical or horizontal placement of the equipment within a sea cage. Indeed, a handful of studies have shown that salmon tend to be vertically stratified with size in a sea cage (Folkedal et al., 2012; Nilsson et al., 2013; Sauphar et al., 2024). Typically, smaller, leaner fish tend to stay in the upper layers of the water column, while larger, heavier fish are found deeper. It is also common to observe fish in poor health or with stunted growth displaying abnormal behavior and swimming slowly near the surface (Noble et al., 2018; Sauphar et al., 2024; also observed by farmers and veterinarians). It is possible that salmon actively select their vertical position in the water column based on their health or welfare status. For instance, a fish with a large wound might struggle to maintain homeostasis until the wound closes. In an environment with a strong salinity gradient, occupying a water layer with lower salinity could reduce osmotic pressure allowing energy expenditure to be directed to healing (Jensen et al., 2015; Ytrestøyl et al., 2020; Nilsson and Madaro, 2025). Choosing warmer temperatures within its optimal range will also result in faster healing (Jensen et al., 2015). Depth stratification may also arise when weaker fish are displaced from favorable environmental conditions, either because they cannot maintain their position within the main school or because they are actively excluded by healthier individuals. In all these cases, the vertical positioning of automated sensors may be crucial for obtaining accurate and representative fish status and size assessments.

In the present case study, the goal was to provide new insights into a potential depth-related stratification of Atlantic salmon held in a commercial sea cage based on their physical health and size. Cameras were mounted at four different depths between 1 to 14 m to measure salmon size and assess selected appearance-based health indicators following the scoring from the Laksvel protocol (Nilsson et al., 2022): spinal deformity, jaw deformity, emaciation, operculum damage, scale loss, and body wounds. Four camera pairs were deployed together on selected days at the edge of the sea cage, enabling simultaneous assessments at all four depths. In addition, one automatic camera sensor was deployed more centrally in the cage and systematically moved between the four depths on a daily basis, enabling continuous and long-term monitoring. The study was conducted before and after a delousing operation since delousing injuries are ranked among the most significant health issues in recent years (Overton et al., 2019; Moldal et al., 2025; Stien et al., 2025). Injuries were expected to occur specifically after the delousing treatment early in the present study. It was hypothesized that fish stratification in culture cages is correlated to health status, with injured or poor performing fish occupying distinctly different areas of the cage compared to their healthy and fast-growing counterparts.

Site and fish information

This study was conducted in a commercial sea cage at Gudmundset fish farm in Møre og Romsdal county, Norway (62.45366, 6.60326). The cage had a diameter of 45 m, was 27 m deep (the upper 15 m were cylinder-shaped and the lower 12 m were conical), and equipped with a lice skirt mounted to 5 m depth. Post-smolts of Atlantic salmon (123,237 individuals, 670 g on average) were transferred from the land-based facility Hjelvik Settefisk to Gudmundset in two batches on the 12th and 14th of September 2023. The experiment was conducted from June to August 2024. Fish were deloused with thermal treatment on the 22 nd of June 2024. Before delousing, fish density in the cage was 16.1 kg/m 3. During delousing the fish were sorted and about 30% of the biomass of that cage was transferred to another cage leading to a new density of 11.5 kg/m 3. Fish were fed two meals a day: between 0700 and 1100 in the morning, and between 1600 and 1900 in the afternoon. A small amount of feed was also distributed 30 min before these times to initiate feeding. Cumulative mortality over the observational period was 2.1%.

Study design

Fish health indicators and size were monitored at four depths covering the cylindrical part of the sea cage: 1 m, 5 m, 9 m and 14 m. Four stereovision rigs mounted as a ladder with non-expandable ropes were placed against the cage wall facing toward the center of the sea cage. They were used to cover the four depths simultaneously and videos were recorded for an hour at a time. Each rig consisted of two GoPros HERO8 (GoPro, Inc., San Mateo, USA). The stereovision rigs were deployed on 11 days between the 13 th of June 2024 until the 18 th of July 2024, approximately twice a week (two days before delousing and nine after). During each of these days, the stereovision ladder was deployed twice to record for an hour, once during feeding hours (around 1000) and once during non-feeding hours (around 1400).

In addition, an automatic fish health and biomass camera sensor of the type CView Eye was used to record images continuously over time (CView Eye 2023-2024, Createview AS, Molde, Norway). The sensor consisted of a stereovision camera pair used for biomass monitoring, a camera with higher definition for fish health assessment, and two LED lights placed above and below the cameras. The sensor was placed between 5 to 10 m away from and facing toward the cage wall from the day after delousing onwards (23 rd June 2024 to 24 th August 2024). The sensor was mounted on an automatic winch, provided by the same company as the sensor, allowing the depth to be changed every 24h. The sensor was rotated daily between the above-mentioned depths, always in the same order, between 2130 and 2230. Data collected during this rotation period was excluded to ensure that each sampling day covered the same time interval. By combining the two systems, fish health indicators and size across complementary time and spatial scales were able to be recorded: the stereovision rigs provided short-term data close to the cage wall, while the CView Eye captured long-term data at a more distant position. A diagram with a timeline of the study design is provided in Figure A in the Supplementary Materials.

Environmental monitoring

To register dissolved oxygen, temperature, and depth, four aquaMeasure DOT(D) sensors (Innovasea, Boston MA, USA) were deployed in the sea cage 5 to 10 m from the cage wall close to the four intended depths (1.7 m, 7.9 m, 10.2 m, and 15.4 m on average). In addition, two salinity sensors aquaMeasure SAL (Innovasea) were fixed at 1.7 m and 10.2 m depths. One DOT(D) sensor and one SAL sensor were deployed temporarily at 5.5 m depth on a buoy for routine environmental monitoring of the fish farm until they were relocated into the cage following delousing on June 22 nd. In addition, illuminance was measured at depths of 1 m, 5 m, 9 m, and 14 m every 10 minutes with HOBO Pendant ® MX Temp/Light loggers (MX2202, Onset Computer Corporation, Bourne MA, USA) fixed close to the cage wall facing upwards. Sensors inside the cage were temporarily removed between 19 and 23 June to facilitate the delousing operation. An Aanderaa SeaGuardII Doppler Current Profiler (Xylem Inc., Washington DC, USA) was attached at 1 m depth facing down on one of the most exposed buoys of the fish farm. It was set on a 30 m range (30 cells of 1 m) to log current speed data in intervals of 20 minutes during the entire sampling period.

Health indicators scoring and size measurement

Stereovision dataset

The four stereovision rigs each worked with a designated pair of calibrated GoPro Hero 8 (for more details on the rigs see part 2.4. in Sauphar et al., 2024). Calibration was completed in a swimming pool using a 3D calibration cube (1 m × 1 m × 0.53 m) from SeaGIS, their associated software CAL and guidelines (V4.20–64 bit, SeaGIS Pty Ltd 2006–2024, Bacchus Marsh VIC, Australia). After the last deployment, calibrations of the cameras were controlled in a freshwater tank where objects of known sizes were measured. This confirmed the calibration at the start. All cameras were set to record using the wide lens, a resolution of 1080p and 60 FPS. Videos for each stereo-camera pair were synchronized using a flash of light as reference, and all rigs were time-synchronized using the GoPro app Quick (www.gopro.com) on each camera at the start of the day and the time at which all cameras were deployed. The first 50 minutes of recordings were used and divided into five parts of 10 minutes to spread the measurements across the sampling time (10 fish per 10 min intervals). Fish health indicators and size measurements (fork-length; FL) were manually computed by two observers (authors Sauphar and Hillmann) with EventMeasure (V6.43–64 bit, SeaGIS Pty Ltd 2006–2024, Bacchus Marsh VIC, Australia). To limit bias due to inter-observer differences, all deployments and depths of a given day were assessed by the same observer (see below for the assessment of agreement within and between observers). Days were assigned randomly (automatic randomization in excel) to one or the other observer, and each observer then conducted scoring in a random order (pulling blindly pieces of paper with numbers to decide which deployment to assess first, which camera rig, and which 10-minute interval).

For each assessment, 50 fish were measured and scored for health indicators. Although this is smaller than the census size of 160 fish used in other studies (Føre et al., 2016; Sauphar et al., 2024), a pilot evaluation comparing results from 50 and 160 fish, across two deployments on one randomly selected day post-delousing (all four depths), showed a strong correlation (Supplementary Materials Figure B). Data was pooled by scoring level, health indicator, depth, and deployment to obtain proportions, which were then compared graphically (Supplementary Materials Figure B). Based on these results, the first 10 fish per 10-minute interval were assessed (50 fish per depth and deployment, where possible), resulting in a total of 4,053 fish analyzed.

Only clearly visible fish that were <1.5 m away from the cameras (range value in EventMeasure), <1 m in case of poor water visibility, were selected. Fish were required to be relatively centered in both cameras’ field of view for the FL measurement. Scoring the same individual twice was avoided when possible. The FLs of the same fish used for health scoring were measured at a moment of the video when the fish was as perpendicular as possible to the line of sight.

Health indicators were scored from 0 to 3 (0 = no damage, 1 = minor damage, 2 = moderate to serious damage, and 3 = severe damage) following the Laksvel protocol (Nilsson et al., 2022) with images from the guide visible to the observers at all times. Spinal deformity, jaw deformity, emaciation, operculum shortening or damage (hereafter operculum damage), scale loss and body wound were selected because they could systematically be assessed on videos. Due to the difficulty of differentiating between two scores in some cases (bad visibility, suboptimal lighting, etc.), it was decided to systematically use the lower score if in doubt. If needed, fish movements were followed to optimize scoring (swimming behavior, operculum movement, angle of the jaw).

Since reliability is a central requirement for health assessment (reviewed by Meagher, 2009), the two observers involved (Sauphar, Hillmann) underwent assessor training followed by an intra- and inter-rater reliability check. First, Sauphar was trained by veterinarians and fish farmers to assess health indicators with real fish based on the Laksvel protocol (Nilsson et al., 2022). Sauphar then adapted the protocol to be suitable for assessment using video material and double-checked the adaptations with the help of a veterinarian. Afterwards, Sauphar served as “standard” and trained Hillmann in rating the selected health indicators.

For the training phase of the inter-rater agreement as well as for the intra-rater agreement, 75 fish were selected from eight 10-minute intervals from seven different days. The selection of these intervals was random (pulling blindly pieces of paper with numbers) with the condition that each depth, i.e. each camera rig, and both feeding- and non-feeding hours were included twice. Fish selection within a given interval was done following the criteria described above. For intra-observer agreement calculation, Sauphar and Hillmann assessed the selected same fish twice (Supplementary Materials Table A). Due to some low intra and inter-rater agreements, another session of discussion between the two observers was needed to adjust their assessments. A second inter-rater agreement with 80 new fish was then conducted (Supplementary Materials Table A). These fish were selected from two dates, with one deployment per day and one interval per depth. Intervals were again chosen randomly.

The weighted Cohen´s kappa (weight “quadratic” in package irr; Gamer et al., 2019), which is a measure of the agreement between two ordinally scaled variables, was used. It is used as an index of agreement between two raters or two assessments of the same rater (Gamer et al., 2019). The interpretation of kappa coefficients according to Landis and Koch (1977) is 0.0 to 0.20 slight, 0.21 to 0.40 fair, 0.41 to 0.60 moderate, 0.61 to 0.80 substantial, and 0.81 to 1.00 almost perfect. For the inter-observer agreement, all indicators reached a κ ≥ 0.69 except emaciation (k = 0.14; Table A in Supplementary Materials). Very few fish were assessed as being emaciated (Sauphar: 9, Hillmann: 7). This led to a very low probability for scores higher than 0 and, accordingly, false negative judgements were weighted very strongly in the Cohen’s kappa. It is the author’s opinion that this corrected agreement cannot reflect well the agreement in such a case. In addition, 57 out of 62 disagreements in scores (of totally 480 assessments) were between adjacent categories. During the actual scoring, Sauphar and Hillmann discussed their ratings during video analysis in case of uncertainty (approx. 100 out of 4,053 fish).

CView Eye datasets

Settings for best image quality were adapted daily by Createview. A total of 52 days, or 13 repetitions of the successive four depths of interest, were used for analysis. Due to some technical issues, the camera could not always be moved on consecutive days. Only video recordings acquired after camera changed vertical position were retained while footage obtained at times when camera did not change depth were discarded until the next depth change occurred. To maintain a consistent timeline, a variable named Days Since Start (1-62) was introduced. Furthermore, it was decided to divide the 23 hours (-1h between 2130–2230 when the camera was moved) of picture registration into six meaningful time periods where change in behavior would be expected by the fish, and therefore where their placement in the sea cage could differ and impact the results from the automatic sensor. The first period (P1), referred to as “Night” lasted from 2230 to 0500 (Central European Time). Although this geographic region does not experience complete darkness at this time of year, these were the hours with the lowest illuminance overall (< 1000 lux) registered at 1 m depth on the longest days of June. The second period (P2), from 0500 to 0659, corresponded to “Dawn”. The third (P3), fourth (P4), and fifth (P5) periods represented distinct feeding-related intervals: the first feeding period from 0700 to 1059, the interval between feedings from 1100 to 1559, and second feeding period from 1600 to 1859, respectively. The last period (P6), from 1900 to 2130, covers the time after feeding 2 and before the camera was moved to the next depth.

Fish health scoring

The physical health assessments from the camera sensor were conducted manually by Sauphar to ensure consistency with the scores from the stereovision rigs. The fish used for scoring were identified by the automatic sensor algorithm. Unfortunately, due to bad orientation, very high light reflection on the scales, or poor visibility, all six indicators could not always be scored on every fish. If a picture did not allow at least three indicators to be scored, it was not included in the assessment. A total of 31,676 images (1 m = 1,940, 5 m = 16,297, 9 m = 7,843, 14 m = 5,596) were kept by the algorithm of the sensor over the course of the study.

To spread measurements throughout the day and decrease manual labor, a maximum of 20 images per hour chosen randomly via the command random.sample using the Python programming language (Python Software Foundation, v3.8.10 - 2021) were analyzed. Some hours of the day also contained no registration. All pictures of all days were also randomly assigned a number to randomize the scoring order on EventMeasure. A total of 8,641 fish were scored for spinal deformity, jaw deformity, emaciation, operculum damage, scale loss and body wound where possible.

Weight measurements

The CView Eye sensor also provided biomass data from its integrated stereovision cameras. All measurements were used as available except those giving a weight lower than 100 g because they were considered as errors by looking at the associated images. Comparison with slaughter report at the end of the production suggested a reasonable size estimation from the sensor (average gutted weight from slaughterhouse 4.63 kg, average weight from sensor using a 15% weight decrease to compensate for gut 4.79 kg). Additionally, a subset of 2,360 stereovision images (all images from a randomly picked day) was manually reviewed to confirm the accuracy of the measurement points selected on the fish. Only 15 measurements were found to be incorrect, due to issues such as duplicate measurements on the same fish, the fish not being fully visible in the frame, or different fish being selected in each frame. A total of 57,616 fish images (1 m = 6,719, 5 m = 28,451, 9 m = 15,788, 14 m = 6,658) were obtained for weight estimation and kept for further analysis.

Statistical analyses

All analyses and figures were conducted and drawn, respectively, in R (version 4.4.0; R Core Team, 2024). For both datasets, depth (factor with four levels: 1 m, 5 m, 9 m, 14 m), phase, and their interaction were used as fixed effects in the models (Table 1). For the stereovision dataset, the phase consisted of the type of deployment (factor with two levels: feeding versus non-feeding); for the CView Eye dataset, it consisted of the period (factor with six levels: P1 = Night, P2 = Dawn, P3 = Feeding 1, P4 = Between feedings, P5 = Feeding 2, P6 = Post-feeding before camera relocation). For the CView Eye dataset, Days Since Start and its interactions with the other two explanatory variables as well as the three-way interaction were used as additional fixed effects (Table 1). Phase (deployment or period) was nested in day for the random effect of both models. Camera-pair was additionally nested in phase for the stereovision dataset.

To calculate p-values in a meaningful way in the presence of interactions, sum-contrasts for the categorical predictors (depth and phase) were used and the continuous predictor was centered at 0 (The variable Days Since Start was centered by subtracting its mean, 31, from the original values 1 to 62, resulting in values ranging from -30 to 31).

The datasets and the different scripts written for the statistical analyses are available in a Mendeley repository in a Mendeley repository (Sauphar et al., 2026), see part Data Availability Statement.

Health indicators analyses

A principal component analysis based on the health indicators showed that they were only weakly correlated, and each health indicator of the fish was used individually as an outcome variable in a generalized linear mixed model in both datasets (function gllvm in package gllvm; Niku et al., 2025). For the CView Eye health dataset, only periods P3, P4 and P5 were considered for analysis due to the low number of fish that were recorded during the nighttime periods. P-values for each fixed effect were calculated by performing single-term deletion tests. A bootstrap approach was used to calculate model estimates based on the simulate function provided in the package.

A total score per fish consisting of the sum of the scores from the six indicators together was used as an additional outcome variable (for all fish in the stereovision dataset and for those fish of the automatic sensor dataset where all indicators could be scored). It has been hypothesized that equivalent scores from two different indicators may not have the same impact on the fish’s welfare (Stien et al., 2013; Pettersen et al., 2014; Nilsson et al., 2022). For instance, an open wound scored 2 might be more deleterious for a fish than a scale loss scored 2 as it can lead to a higher risk of infection and of a disruption of osmoregulation. Therefore, a total weighted score was also calculated where injury-related indicators assessed in this study, operculum damage, scale loss and body wound, were weighted based on weighting scores suggested by Stien et al. (2025, Table 2).

The other indicators evaluated in this study, emaciation, spinal deformity and jaw deformity, are considered semi-permanent or permanent physical conditions that do not typically result in sudden changes to the fish’s health status and do not affect the fish as much as the injury-related indicators on the following days and weeks. Thus, they kept their initial scoring values in the total weighted score.

For the analysis of the health indicators, distribution families and assumptions were checked using a graphical analysis of the residuals (based on plots provided with the package DHARMa; Hartig, 2024). With the families that were chosen, no systematic deviations from the assumptions were apparent. For the stereovision dataset, a Poisson distribution was selected to fit the models for operculum damage, scale loss, and total score and a negative binomial distribution for spinal deformity, jaw deformity, emaciation, body wound and total weighted score (n = 4,053 fish for each). For the CView Eye health dataset, a Poisson distribution was used for the models of spinal deformity (n = 7,132), emaciation (n = 7,200), operculum damage (n = 5,863), scale loss (n = 5,427) and total score (n = 4,099), and a negative binomial distribution for the models of jaw deformity (n = 6,495) and total weighted score (n = 4,099). Due to the very low number of fish with body wound scores above 0 (n = 211 out of 5,376) in the CView Eye health dataset, the model could not be meaningfully estimated, and it was decided not to perform a statistical analysis for this indicator. Yet, some more detailed data is presented in Figure C in the Supplementary Materials section.

Fish size

Size variables (FL for the stereovision dataset and weight for the CView Eye dataset) were used as outcome variables in linear mixed-effects models (function lme in package nlme; Pinheiro et al., 2023). For these, we calculated marginal p-values with the anova function provided in the same package. In the stereovision dataset, fish FL was transformed using a power of four for the analysis to satisfy model assumptions as reflected in the distribution of residuals. Because a numerically non-positive variance-covariance matrix resulted in this model calculation, estimates were calculated based on a parametric bootstrap on a blmer model with the same structure (package blme; Chung et al., 2013; bootstrap provided in package lme4; Bates et al., 2015). In the CView Eye size dataset, weight was used untransformed as the outcome variable. The contrast function (package contrast; Kuhn et al., 2022) was used to calculate model estimates. Assumptions were again checked using a graphical analysis of the residuals. No major deviations from these assumptions were detected in these fish size models.

Environmental conditions

Temperatures during the first week of sea cage monitoring decreased with depth with similar temperatures in the upper layers around 12.5 °C and colder temperatures at 10.2 m and 15.4 m (10 °C and 8 °C respectively; Figure 1A). The temperature differences between depths never exceeded 2.5 °C apart from that first week of monitoring and the third week of July when temperatures dropped before they increased again from the 18 th of July. Water layers of 1.7 m and 7.9 m were well mixed and showed very similar temperatures around 16 °C and above from the last week of July, while temperatures fluctuated around 13.5 °C and 15 °C at 15.4 m and 10.2 m respectively. All depth layers were rather homogeneous from the second week of August until the end of the experiment.

Dissolved oxygen levels were above 80% across all depths with a few troughs down to 70% at 1.7 m and 7.9 m, without any consistent period to assign them to, and one trough at 60% at 7.9 m one night (Figure 1B). Outside that latter trough, dissolved oxygen was within the minimum levels ensuring normal activity in healthy, unstressed post-smolts for the temperatures registered (Nilsson et al., 2022).

Salinity values were below 30 ppt at the observed depths during the entire trial (Figure 1C). Levels were quite similar between 10.2 m and 15.4 m the last week of June. Afterwards, water layers became slightly more stratified with a difference in salinity up to 3 ppt.

Current speed was on average 5.03 ± 3.68 cm per second (Figure D in Supplementary Materials). Illuminance during the day (between 0500 and 2130) was 5,608 ± 8,467 lx at 1 m, 1,510 ± 2,557 lx at 5 m, 674 ± 969 lx at 9 m, 250 ± 371 lx at 14 m. Maximum hourly averaged illuminance registered per day are shown in Figure E in the Supplementary Materials.

Short-term simultaneous measurements (stereovision)

Salmon in the studied sea cage received severe scores of 3 very rarely (<1.3%). At least 84% of individuals were scored a 0 or 1 for all indicators, and 93% of individuals had a score 0 for spinal deformity, jaw deformity, emaciation and body wound. The highest proportion of scores 2 was found for operculum damage with 14.5% and scale loss with 8.6% (Figure 2).

On average, fish with higher scores for spinal deformity, emaciation, body wound, and total score, were found at 1 m depth, whereas in the case of scale loss, fish with higher scores were observed at 5 m depth (Figures 3, 4; Table 1). On the other hand, scores from jaw deformity, operculum damage, and weighted total score did not differ systematically between depths (Figures 3, 4; Table 1). In addition, fish with lower scores for emaciation, scale loss, and total score were observed during feeding compared to the non-feeding times (Figures 3, 4; Table 1). None of the depth-time interactions were obvious in the estimates nor could they be supported statistically.

In terms of fish FL, slightly longer fish stayed at 5 m and 14 m compared to the other depths (Figure 5; Table 1) independent of whether they were observed during feeding or non-feeding times.

Long-term monitoring (CView Eye)

The proportion of salmon that scored a 3 for an indicator was at most 0.4%. Most individuals (85%) received a score of 0 or 1 for the assessed indicators. The highest proportion of salmon that scored a 2 exhibited operculum damage (14.1%) and/or scale loss at (3.9%; Figure 2).

Fish with more serious damage were observed at 1 m depth for both of the indicators spinal deformity and emaciation (Figure 6; Table 1). No systematic pattern with depth was otherwise visible for jaw deformity, operculum damage, scale loss, total score and weighted total scores (Figures 6–8; Table 1). Lower scale loss scores were registered during feeding 1 and between feedings compared with feeding 2 (Figure 7; Table 1). No systematic difference was apparent between periods for any of the other indicators (Figures 6–8; Table 1). Average scores decreased with time for all indicators except spinal deformity and jaw deformity which showed an increase (Supplementary Materials Figures C, G–N). The models detected some interactions for spinal deformity and scale loss reflecting some minor variations in the slopes across time but no qualitative change in the observed patterns (Supplementary Materials Figures G, K, N).

A depth-related size distribution was observed: smaller fish tended to remain at 1 m (264 g below the overall average), slightly larger individuals at 5 m (18 g below average), and the largest fish were predominantly found at depths of 9 m and 14 m (96 g and 186 g above average respectively; Figure 9; Table 1).

Fish recorded during the period between feedings were larger on average (135 g average) than during the other periods, followed by the ones registered during feedings 1 and 2 (both 73 g above average), then during the period post-feeding 2 (15 g below average), and finally during dawn (27 g below average) and night (238 g below average; Figure 9; Table 1).

This vertical size stratification appears more pronounced during nighttime compared to daytime periods and also later in the observation period even though these effects were only marginally (or not) statistically supported (Figure 9; Table 1).

In this case study, several health indicators were evaluated, and salmon size was measured at four different depths within a commercial sea cage to investigate vertical stratification. Two types of equipment were used reflecting short-term and long-term measurements. Based on this, the hypotheses that salmon exhibit a vertical stratification due to their health status and size were tested. This is highly relevant for the placement of automatic sensors that register fish health indicators and estimate size to ensure the most representative assessment of the entire cage population.

Physical health stratification

The absolute average differences that were observed in health scores, e.g. between different depths, were small. However, this should not be interpreted as evidence of negligible biological relevance. Even small differences in mean scores can reflect meaningful changes in the proportion of affected fish (Nilsson et al., 2022; Stien et al., 2025). This is why to help interpret average scores accordingly, the distribution of the different scores were presented as proportions (Figure 2), allowing subtle but relevant shifts to be detected. Still, considering the overall good status of the fish in this study and the very few instances of moderate to severe damage, the results should be carefully interpreted.

A consistent pattern was observed across both datasets from the stereovision rigs and the CView Eye sensor where individuals with higher scores for spinal deformity and emaciation were more frequently observed at 1 m compared to the other depths. Spinal deformity can impair swimming ability in the most severe cases and increases metabolic cost (Powell et al., 2009). Affected salmon may struggle to swim efficiently, potentially leading them to be excluded from schools and forced into suboptimal depths, such as near the surface where they are more vulnerable to predation and UV exposure (Fernö et al., 1995; Juell, 1995). This depth stratification could also be influenced by an impaired swim-bladder function that has been linked to abnormal skeletal development in other species (Chatain, 1994; Canepa et al., 2023). Although a direct relationship between swim-bladder morphology or function and spinal deformities has not yet been demonstrated in salmon, severe vertebral deformities could compromise normal swim-bladder function and reduce buoyancy control, limiting affected salmon to shallower water layers.

The pattern observed for emaciation aligns with previous studies, which found that emaciated or growth stunted fish tend to occupy the upper layer in the water column (Stien et al., 2013; Vindas et al., 2016; Noble et al., 2018; Sauphar et al., 2024). Emaciation is well correlated with fish size, and similar findings with leaner fish found at 1 m depth were observed in the present study regarding size stratification (see part 4.2). Folkedal et al. (2012) hypothesized that the lower lipid content in emaciated fish makes them less buoyant. As a result, they must rely solely on their swim bladder to regulate buoyancy. However, the mechanism might not be sufficient under increased pressure at greater depths, restricting them to shallower depths. Although swimming faster can help overcome negative buoyancy (Dempster et al., 2008; Sievers et al., 2018), the energetic cost may be too high for emaciated fish or individuals with spinal deformity to swim at deeper depths for long periods of time, further confining them to shallow waters. On the other hand, one could argue in favor of a deliberate choice made by weaker fish to swim at shallow depths to benefit from reduced flow speed caused by the lice skirt (reviewed in Jónsdóttir et al., 2023). As an additional hypothesis, swimming in the upper layers may also reduce the energy expenditure associated with foraging, since feeding occurs at the surface and weak fish might increase their chances of capturing pellets by staying in these water layers. Nevertheless, these considerations likely do not apply to severely emaciated salmon, often referred to as “loser fish”, which may already be in terminal decline (Vindas et al., 2016; Noble et al., 2018). Whether this stratification is driven by buoyancy or swimming constraints and/or by an active energy-saving behavior from the fish remains an open question as data from the present study alone does not allow the establishment of a clear cause-effect relationship. Nevertheless, this consistent pattern between short-term and long-term monitoring is interesting since both spinal deformity and emaciation are chronic indicators and no consistency was found in terms of injury-related indicators. This suggests a long-term stratification that should be investigated further across different seasons to see whether it would be impacted by more drastic changes in environmental conditions.

Also consistent between the collection methods is the pattern for the indicators jaw deformity, the last chronic indicator, and operculum damage for both of which no depth stratification was apparent. This suggests that a fish with jaw deformity or operculum damage may not actively select one depth over another based on these characteristics alone, at least in the environmental conditions and indicator severity encountered over the course of this experiment. Finally, no stratification pattern was observed based on the comprehensive metric of all health indicators (weighted total score). As a closer proxy to fish health, and to some extent welfare, this result suggests that there was no stratification based on the fish overall condition. Whether a fish received a high score from one indicator or another cannot be determined here, but it could be said that fish experiencing similarly challenging health conditions are not grouped at a certain depth. On the other hand, the patterns were similar with the total unweighted scores. However, with weighting, the spread between values became much larger, which inflated the standard errors and reduced the statistical power as most injuries were scored as absent or minor. Consequently, the total weighted score might hide the changes in a population touched primarily by minor damage, as was the case in this study. In addition, while this score may serve as the best available proxy for fish welfare in this context, it lacks further information necessary for a comprehensive overview of fish status such as other health indicators not registered in this study and broader welfare aspects beyond fish physical integrity.

In the short-term monitoring conducted in the present study, a higher prevalence of salmon with elevated body wound scores was observed at 1 m depth and scale loss at 5 m. Since faster metabolic rate is associated with higher temperature (Hvas et al., 2017), salmon exhibiting wounds or scale loss may seek water layers within the higher range of their thermal preference, where conditions can facilitate healing (Anderson and Roberts, 1975; Jensen et al., 2015). In our case, warmer temperatures were found at 7.9 m and above, suggesting temperature may have played a role in the placement of these fish. Surface water also equates to lower salinities. As osmoregulation may be compromised if a fish has a large wound (Noble et al., 2018), moving to a water layer with lower salinity will decrease osmotic pressure. In the week after delousing, there was only a 2 ppt difference between 10.2 m and 1.7 m. It is difficult to state whether this would be enough for the fish to make an active choice based predominantly on salinity in this case. Regarding scale loss, experimental work on smolt found no effect on osmoregulation (Nilsson and Madaro, 2025), suggesting limited benefit from seeking lower salinity. This seems even more likely for adult salmon, such as the individuals followed in our case study, that have a high tolerance to changes in salinity (Bakke et al., 1991). However, the results of this smolt study were obtained under controlled, near-optimal conditions without additional stressors and only examined standardized dorsal scale loss. Since injuries in practice may occur elsewhere, potential location effects warrant further study (Nilsson and Madaro, 2025). Dissolved oxygen concentrations were similar on average at 7.9 m and 1.7 m for the month following the delousing and cannot explain these stratifications. Finally, the hypothesis that injured individuals are being excluded by other fish or cannot be kept up in the main school remains. Fish density at 5 m was quite high, indicated by the amount of fish observations that were substantially higher at this depth compared to the other depths from the CView dataset, indicating that salmon with scale loss are not excluded by congeners and can remain in a school. This was likely due to compromising between preferred temperatures and the theorized avoidance of surface UV (Fernö et al., 1995; Oppedal et al., 2011) for which their epidermis might have become more sensitive. However, this hypothesis in the case of salmon with body wounds cannot be refuted as these fish might not only select their swimming depth based on more favorable environmental conditions for healing but might also be forced out of the main group. Further investigations are nevertheless required to unravel such causality.

Finally, higher scores for total score were registered at one meter in comparison to other depths in the short-term dataset. This seems logical since it was the case for half of the assessed indicators. As mentioned in the methodology, such total scores do not exactly reflect the status of a given fish since neither the levels nor the indicators are weighted against one another. Since a low occurrence of scores above 1 was observed, this total score still reflects quite well whether an individual was accumulating physical deviations or not which appeared to be the case for fish at 1 m.

Long-term monitoring revealed an effect of time on all indicators. Scores for permanent spinal and jaw deformities tended to increase over time, while scores for emaciation, total scores and injury related indicators decreased (although particularly small estimates for operculum damage). This suggests visible healing of the latter indicators over time or the loss of these animals. This might explain why a stratification of an injury-related indicator such as scale loss was only visible in the short-term dataset since measurements stopped 25 days after delousing. An additional month of data collection occurred in the long-term monitoring and could have obscured these differences between depths, although the estimates are extremely small in the statistically supported interaction depth x time for scale loss. It is difficult to unravel meaningful trends and such low p-values might have happened due to the large sample size. No statistics could be conducted on the indicator body wound for the long-term dataset unfortunately and cannot be considered in support or refutation of this hypothesis. Still, plotting body wound scores against time (Supplementary Materials Figure C) showed less occurrence of wounds with a score above 1 after some days, indicating that healing is also visible for that indicator. In addition, a large number of healed wounds were observed during the scoring but were not included in the analysis as they were not assumed to affect the fish’s health further. Another possible explanation is that collecting measurements over several continuous hours (0700 to 1900 in the present analysis), rather than during a short 1-hour window, allows salmon to shift their vertical and horizontal distribution throughout the day. This extended sampling period increases the likelihood of capturing multiple groups of fish, resulting in a more homogeneous representation of the population compared to shorter sampling intervals. Finally, the horizontal position of the stereovision rigs versus the position of the automatic sensor could be responsible for the differences between datasets. Nevertheless, since a stratification for two chronic health indicators, and a depth x time interaction statistically supported for scale loss were documented, results tend to favor a possible temporary stratification. Further studies are, however, required to validate this hypothesis and should include additional health indicators.

Period of the day for scale loss and some interaction terms for spinal deformity and scale loss were statistically supported without any clear patterns to attribute these changes to due to the very small differences in estimates. These statistically supported variations likely reflect an overly large sample size, leading to statistically significant but biologically negligible results.

Size stratification

According to the stereovision dataset, the maximum differences between depths related to size were only 1 cm during feeding deployments and 0.5 cm during non-feeding deployments. These differences are considered negligible, particularly given that the study focused on adult salmon, and similar small variations were not statistically supported in previous research (Sauphar et al., 2024). The presence of a low p-value in the present analysis is likely a statistical artifact resulting from the large sample size. Most measurements in the stereovision dataset were conducted shortly after delousing and size sorting, which may explain the absence of clear size-based vertical stratification. In contrast, Sauphar et al. (2024) observed a more pronounced vertical size stratification, likely because that fish population was never sorted after their transfer in a sea cage until the first slaughter event.

The 450 g difference on average observed between fish at 1 m versus 14 m on CView Eye dataset is coherent with previous studies where on average leaner fish were found at 1 m compared to deeper depths and fish of average size was present at all depths (Folkedal et al., 2012; Nilsson et al., 2013; Sauphar et al., 2024). The weight estimate obtained at 5 m depth showed the smallest deviation from the overall average, which is likely attributable to this layer being the most consistently occupied by fish. Furthermore, during periods post-feeding 2 and dawn, behavioral transitions between day and night may have enhanced mixing between groups, thereby improving representativity at this depth. It is difficult to explain why 135 g heavier salmon were registered on average during the period in between feedings and 238 g lighter during night. The vertical size stratification appeared more pronounced during nighttime compared to daytime periods suggesting a different group dynamic between day and night. This mostly means that it is important to conduct measurements continuously throughout the day and night to cover the size variability and that short-term monitoring might lead to bias in sampling either leading to over estimation or under estimation of the average weight.

Use of automatic sensor considerations

Although subtle, results of the present study suggest that vertical stratification of injured salmon in cages is likely taking place, especially on farming operations involving handling or during winter ulcer outbreaks where a large number of severe wounds could be expected. Scoring fish manually as an alternative to the automatic sensor revealed several limitations that should be taken into consideration. Ideally, more injury-related health indicators should be included in the assessment, e.g. eyes or snout injuries, since they are most likely the ones inflicting noticeable changes in fish health, but this requires the automatic sensor to register very systematic images with fish in a particular position. In any case, some indicators cannot be combined for simultaneous assessment in a single image. For instance, snout damage and operculum damage cannot be combined, unless severe and very obvious injuries are present, since the fish would need to have its head facing toward the camera for a good assessment of snout damage, while it would need to be seen from the side and with its operculum closed to assess operculum damage. Similarly, an image with a fish seen at an angle from the back does not offer the possibility to assess the status of the eye, the snout or the jaw. Currently, automatic sensors rely on scoring systems created to assess fish taken out of the sea cage, with a possibility to observe them from different angles. This makes automatic sensor assessments prone to false negatives, even assuming perfect measurement precision and image quality, since the system registers an absence of deviation rather than the impossibility to assess one. To address this issue, algorithms should first consider the fish’s position in the image to select which indicators are possible to score. Indicators that cannot be assessed due to positioning should be assigned ‘NA’ values, while only those with sufficient visibility should be scored.

Substantially fewer measurements were also obtained during nighttime, and the quality of the images at night was also lower. Night measurements rely on the use of artificial lighting (also present during the day though its effect is considered negligible in summer daylight) which has been shown to attract salmon in commercial sea cages (reviewed in Oppedal et al., 2011). If light did attract some individuals, the sensor might still have failed to capture images of sufficient quality for analysis. Additionally, salmon are believed to break their schooling pattern at night (Oppedal et al., 2011). It is possible that fish drawn to the light approached the sensor voluntarily, resulting in suboptimal angles in the captured images, unlike during the day, when fish typically pass in front of the sensor while swimming in a coordinated school.

The placement of the sensor had a clear impact on the number of images registered. Almost twice as many images were retained at 5 m compared to 9 m, followed by 14 m and finally 1 m having the least quantity of images, probably reflecting the density of fish at these depths over the course of the experiment. Fish at 5 m reached a good compromise in environmental conditions by staying at warmer temperatures while avoiding high light intensities. During winter when upper water layers become colder than deeper waters, it would be beneficial to lower the sensor to increased depths. Using environmental cues to place the sensor is a good alternative to the use of echo sounder if none are present on site. In case of health issues becoming increasingly more visible, positioning the sensor at a depth where environmental conditions are better for healing purposes (i.e. aiming for the warmest temperatures within the optimal range for salmon even if it means higher light intensities) could be relevant to target injured individuals and follow the healing process. In addition, it seems reasonable to advise a change in the position of the automatic sensor close to the surface once or twice a week, both to assess the evolution of health indicators such as emaciation and spinal deformity, and to capture the full-size variation present in the sea cage.

Fish in this study displayed a fairly good physical health status overall considering they had just been subjected to a delousing treatment. In addition, environmental conditions were in favor of an optimal recovery. Still, results suggest both a long-term vertical stratification of salmon with chronic deviations and a temporary stratification of salmon displaying injury-based deviations. Future studies should be conducted in more cages and across different seasons with more stratified water columns and specifically during periods when salmon are more at risk, e.g. during jellyfish blooms or outbreaks of winter ulcers. However, it would be necessary to rely on automatic measurements to make the repetition of such study realistic and ideally using several camera sensors simultaneously. Nevertheless, several aspects need to be considered to improve the precision of the measurements. Placing the sensor based on the environmental conditions seems to be reasonable in order to target the most fish (if no echosounder on site). If injury-related indicators increase, it might be relevant to place the sensor where environmental conditions are optimal for healing purposes (i.e. warmest temperatures within the optimal range for salmon, to monitor injured individuals and follow their recovery). Finally, positioning of the automatic sensor close to the surface once or twice a week appears necessary, both to assess the evolution of health indicators such as emaciation and spinal deformity, and to capture the full-size variation present in the sea cage.