Three experimental field campaigns (P1, P2 and P3) were conducted in two salmon farms exposing fish of different sizes to different sound frequencies and light intensities (see Table 1 for an overview). The tests were conducted in industry scale net pens at SINTEF ACE (SINTEF, 2023) October 19–20 2021 (P1), August 08–09 2022 (P2), and August 23–26 2022 (P3), under calm conditions to avoid the influence of strong currents and waves, and during normal daylight hours (09:00–17:00) to minimise variability in natural light affecting fish perception. Feeding and other treatments were paused to eliminate potential confounding effects. All tests featured a cylindrical aluminium structure of 30 cm diameter and 60 cm height (referred to as the “centre structure”) that housed the sensors and actuators needed to both induce tailored sound and light disturbances and observe the resulting fish responses (Figure 1). 360° mechanical scanning image sonars (Ping360, BlueRobotics Inc.) were mounted on the top and bottom panel of the centre structure to enable observing the distribution of the fish around the centre structure with redundancy (Figures 2, 3).

P1 and P2 were designed to study fish reactions to acoustic signals, and thus featured an underwater speaker (DRS-8 in P1 and DRS-8 MOD 2 in P2, Oceanears Inc.) installed inside the centre structure. The centre structure was then mounted inside a cylindrical yellow shell structure with dimension \diameter 60 × 60 cm that was suspended 8 m below the surface in both tests. This depth corresponds to a biologically active zone for fish during daylight hours (Føre et al., 2018), aligns with typical ROV operating depths in net pens, and falls within the waterproofing limitations of the equipment (Zhang et al., 2024). The underwater speaker was connected to a custom built amplifier placed topside that was controlled by mobile phone via Bluetooth, using the Android-application Tone Generator (TMSoft). This system was set to induce sounds at specific frequencies at set intervals using a high sound level (estimated at 180 dB re 1 µPa at 1 m) to maximize the chance that the fish would perceive the sound. Since farmed salmon hearing is more sensitive to low frequencies of approx. 100–580 Hz (Hawkins and Johnstone, 1978), and the acoustic noise measured in fish farms tends to be dominated by low frequencies (<1,000 Hz) (Radford and Slater, 2019; Barrett and Oppedal, 2025), the sound signals used in the trials spanned from 100 to 1,000 Hz. In P1, tones with centre frequencies 100 Hz, 200 Hz, 400 Hz, 600 Hz and 1,000 Hz were played in random order, each frequency occurring in total three times, while P2 featured only 200 and 600 Hz tones that were played sequentially six times. Each tone was maintained for 60 s, and there was a 10 min break between consecutive signals. The experiment was conducted in two different pens in both trials (Pens A and B in P1 and Pens C and D in P2), testing each sound frequency with 3 or 6 replicates across different pens.

In P3 the aim was to study fish’s reaction to light. Two SeaLED 300 LED lights (Imenco Inc.) were therefore installed on the centre structure (Figure 3), each of which could be set at three settings: 0 (=lights off), low, and high intensity. Combining the two lights resulted in four distinct light settings: low = 600 lx (one light off, one light on low), medium = 1,100 lx (both lights on low), high = 6,600 lx (one light on low, one light on high), very high = 14,500 lx (both lights on high). Since the LED lights needed to be visible for the fish, the centre structure was not clad in the cylindrical shell structure during P3. This experiment was also conducted in two pens to achieve fish group replication (Pens C and D). Additionally, since natural light strongly affects the light conditions in the pen, the trial was repeated at a deeper depth (12 m) in one pen (Pen D) to see if fish responses differed when natural light levels were lower. In the tests conducted at 8 m depth, each of the four light intensity levels was turned on six times at random, each treatment lasting 60 s, and with a 10 min break between samples (as for sound inP1 and P2). The 12 m depth trial was similarly set up, but with a total of six repetitions in one single pen divided between 2 days (four on August 23 + two on 26 Aug 2022).

This section describes the DL approach used to process the acoustic data and the statistical methods used to study fish behaviour changes.

The 50-s CFP data from P1 shows that salmon in Pens A and B reacted to 400 Hz sound by increasing the distance to the sound source on average by 28% and 12%, respectively. This indicates avoidance behaviour. Notably, when the sound stopped playing, distance between the fish group and the structure returned to the original distance (Frequency x Timing, F 8,89 = 2.131; p = 0.045, followed by pairwise comparisons with p < 0.05 for During > Before = After; Figure 7c). This redistribution pattern suggests a temporary displacement of fish away from the structure during sound exposure, followed by reaggregation once thedisturbance ended. Sounds played at 100, 200, 600 or 1,000 Hz did not cause any significant change in the distance the salmon kept to the structure. This suggests that the fish in this study were particularly sensitive to 400 Hz, possibly because of the overlap between 400 Hz and species-specific auditory sensitivity or structural resonance frequencies.

Similar results were observed in P2, with no effect of the onset of sound on the distance of salmon from the sound source measured for 200 or 600 Hz. Here, the only differences were found between the two test pens, where fish generally stayed further away from the structure in Pen D than in Pen C (Pen, F 1,71 = 30.27; p < 0.001; Figure 8). This spatial discrepancy likely reflects environmental or behavioural differences between pens rather than sound exposure effects.

When exposed to artificial light at a depth of 8 m, 50-s CFP data showed that the salmon swam closer to the light source when light was turned on (Timing, F 2,143 = 7.886; p < 0.001), demonstrating an attraction response. Moreover, there was a trend towards an increasing effect with increasing light intensity (Light Intensity, F 3,143 = 3.888; p = 0.011, Figure 9). That is, fish redistributed more tightly around the structure at higher intensities, suggesting a intensity-dependent behavioural response. This applied to both Pens C and D, though the salmon in Pen C generally kept a larger distance than fish in Pen D (Pen, F 1,143 = 78.753; p < 0.001).

Additional tests conducted at 12 m depth in Pen D showed a contrasting reaction where fish moved away from the light source, independent of the light intensity (Timing x Depth, F 2,143 = 21.899; p < 0.001). Moreover, fish generally kept a larger distance to the test structure in 12 m depth than in 8 m depth, suggesting a depth-related shift in their behavioural response to artificial light.

The salmon exhibited a noticeable reaction when exposed to sounds at 400 Hz, where they abruptly moved away from the sound source ¹ , demonstrating that the fish were able to sense the acoustic pulse, either auditory or through the lateral line organ. This behaviour was distinct and consistent across pens, suggesting a heightened sensitivity of fish to this particular frequency, which is consistent with previous studies reporting that Atlantic salmon have a relatively narrow hearing bandwidth up to approx. 300–500 Hz (Popper and Hawkins, 2019). The response resembles escape behaviour, which may indicate that the fish were startled by the sudden onset of the sound signal. Since the fish group structure settled back to the original distance after the sound stopped, it is likely that the response was directly linked to the acute sound stimulus. This resembled escape responses toward sound signals that have previously been observed in salmonids (Knudsen et al., 1997; Bui et al., 2013; Oppedal et al., 2025). While this implies that the findings in the study were in consensus with earlier findings, it should be noted that the previous studies used lower frequencies (1–150 Hz), and observed escape behaviours mostly when exposing the fish to signals in the infrasound range ( ≤ 12.5 Hz) (e.g., Oppedal et al., 2025).

The fish did not show a clear response when exposed to the acute emission of sounds at other frequencies (i.e., 100 Hz, 200 Hz, 600 Hz and 1,000 Hz). This lack of reaction may indicate that sounds at these frequencies were not perceived as an uncomfortable disturbance or as an indication of some increased risk by the fish. An alternate explanation for the lack of response for these frequencies can be that the sounds were outside the primary hearing range of the fish. While, according to Hawkins and Johnstone (1978), 100, 200 and possibly 600 Hz should be inside or close to the normal hearing range of salmon, the hearing loss caused by otolith deformities observed in farmed salmon (Reimer et al., 2016) could vary with frequency. If so, the fish may have been unable to detect the sounds with sufficiently clarity to distinguish them from the ambient environment, and hence not have any motivation to respond. This is supported by earlier studies that found that Chinook salmon with hearing loss in the 100–300 Hz range due to otolith deformities also showed a loss in hearing sensitivity for adjacent frequencies including 400 Hz (Oxman et al., 2007). Although Chinook salmon and Atlantic salmon are not the same species, this implies that hearing loss due to otolith deformation is likely to impact the whole aural range and not only a few selected frequencies.

In this context the lack of a clear reaction to sounds at 200 Hz was unexpected as previous research had indicated higher sensitivity for 100–300 Hz than for 400 Hz (Oxman et al., 2007). This discrepancy might reflect interspecies variation, differences in experimental design, or variability in hearing ability due to aquaculture-related otolith deformities. Another element to consider in this discussion is that the acoustic power level used in this study (estimated at 180 dB re 1 µPa at 1 m) was higher than that observed in studies exploring soundscapes at fish farms (which typically range from 98 to 112 dB re 1 µPa at 1 m, Radford and Slater, 2019; Barrett and Oppedal, 2025). This means that the sound signals played during the trials were louder than the ambient noise level at the farm and implies that the fish should have been able to perceive these signals clearly, yet a response was only observed at 400 Hz. More specific research into the aural properties of Atlantic salmon is needed to conclude whether the absence of response toward sounds with frequencies other than 400 Hz is due to hearing impairments or the fish making a conscious decision of not responding.

Further studies into the hearing of farmed salmon could help shed light on these aspects.

The salmon consistently stayed closer to the light when it was placed at 8 m, an effect that was modified by light intensity as the fish moved closer to the source when the intensity was increased. This suggests that the fish were attracted to the light at 8 m depth but appeared to avoid it at 12 m depth. Both these patterns were emphasised when light intensity was increased (see Supplementary Figure S3, Supplementary Material). These findings reflect observations by Juell et al. (2003) and Juell and Fosseidengen (2004) where fish tended to cluster more tightly around lights on the surface or in shallow water but became more dispersed when lights were placed deeper. The impact of artificial lights on the spatial distribution and behaviour of fish within the water column is thus a more complex combination effect of light intensity and depth rather than solely depending on light intensity.

Closer to the surface, the impact of natural light on the pen environment is stronger, meaning that the introduction of artificial lights at shallow depths may not affect lighting conditions as much as it will deeper in the pen. Moreover, the fish population in the pen may be stratified, with those seeking surface light, e.g., to feed, being present in the upper meters and showing positive phototaxis. In contrast, fish in lower depths may be actively avoiding the daylight on the surface, and thus may also show negative phototaxis and avoidance to the artificial light, as seen in the experiment at 12 m depth. Such variation in phototaxis has been described in other studies (Oppedal et al., 2011) and may vary throughout the population in the pen. For example, Wright et al. (2015) used light at night to attract fish to a specific depth but did not find the reaction to be consistent for the entire population in the pen, with some fish following the light while others showed no reaction.

This study utilised a hierarchical replication approach: (a) within-pen replication, involving repeated testing of each impact factor under consistent conditions within the same pen to isolate factor effects; and (b) between-pen replication, involving the repetition of identical trials across different pens at different time to capture natural environmental variability typical of aquaculture systems. This dual-scale approach enabled distinguishing the tested impact factor effects from potential confounding factors associated with spatio-temporal environmental variations, thus enhancing the robustness and validity of our findings. Previous research has shown that strong currents and large waves can influence fish swimming performance and behaviour, particularly in smaller individuals with lower swimming capacity. However, such effects are generally associated with prolonged exposure. Short-term experiments (on the scale of hours), as in our study, are unlikely to be significantly affected by variations in fish swimming behaviour due to environmental conditions (Hvas et al., 2021; Zhang et al., 2024).

Each 60 s stimulation in the trials was followed by a 10 min break period that allowed the fish to return to baseline behaviour, and also effectively provided a “no-disturbance” reference condition within the same environment. To control for potential behavioural drift or disturbance effects, we assessed the behaviour of the fish across three distinct phases: before, during, and after the stimulus. There was no statistical difference between before and after the fish were exposed to sound or light. As the entire school of fish tends to move in a circular pattern, the fish being exposed to a stimuli at a certain time and potentially reacting to it are unlikely to stay in the area after responding. Thus, fish in the observation area after the stimuli was switched off, were probably unstimulated fish that had not responded to the original treatment but are entering the area for the first time. Moreover, the similarity in avoidance distance before and after exposure suggests that the pen environment remained stable during the experiments and the experiment only affected fish in immediate vicinity to the structure. In sum, this indicates that the responses exhibited by the fish were primarily elicited by the intended sound/light stimuli and not due to other factors (e.g., long-term habituation, environmental variability, or procedural disturbances such as external disruption clearing entire pen areas), supporting the validity of our experimental design and data collection.

Finally, the observed fish distribution patterns are also consistent with the hypothesis that fish respond to stationary stimuli placed within the pens by spatially distributing in a circular pattern centred around the source of the disturbance. Such circular avoidance patterns likely reflect a natural schooling response to internal localised perturbations. In contrast, external disturbances, such as predators or maritime traffic, would likely elicit different spatial responses, potentially in the form of directional rather than symmetrical avoidance patterns. Future studies could explore how external disturbance directionality influences the shape and extent of avoidance behaviour, providing deeper insight into the complexity of fish responses under varying impact factors.

In this study, we applied a DL-based semantic segmentation approach to sonar data to analyse fish avoidance distance in response to sound and light. Conventional sonar processing techniques, such as threshold-based segmentation, usually rely on raw signal intensity to detect objects present in the data. These methods often capture not only fish but also the reflections from other objects including central structures, the sonar itself or other devices present in the water volume. Because of this, estimating true fish avoidance distances using such approaches will require additional post-processing steps, which in turn reduces the autonomy of the method and may introduce new errors and inaccuracies in the outcomes. The DL-based method in contrast leverages the capacity of convolutional neural networks (CNNs) to learn complex spatial patterns and contextual features embedded in the sonar data, and directly identifies the specific regions where fish avoid the sound/light. This enables a more direct, robust, and targeted quantification of avoidance behaviour, while minimising the influence of background noise and structural artifacts. Although the original method as presented by Zhang et al. (2024) was found to provide reliable results, the introduction of the second class of invalid data in the present study further improved the prediction accuracy of the method.

Our results provide valuable information on fish reactions to sound and light in net pens. This data can be used to design underwater vehicles, machinery, and other sound-emitting human systems/activities to operate at frequencies that have less chance of disturbing farmed fish. For instance, Cai and Bingham (2011) found that the motors of a standard ROV emitted noise at 400–500 Hz, which overlaps with the frequency found to elicit an avoidance response in the salmon in the present study. Designing new vehicles equipped with motors that emit sounds with frequencies outside this range could thus reduce the risk of startling or displacing fish populations during ROV operations. However, care must then be taken to also avoid the lower frequencies that have previously been identified to have adverse effects on farmed fish ( ≤ 100 Hz, Knudsen et al., 1997; Bui et al., 2013). Such design measures could also help minimising the environmental impact of vehicle activities on aquatic ecosystems in general, as it is likely that wild fish have similar sensitivities to sounds as farmed salmon.

At the same time, this apparent sensitivity to 400 Hz could also be exploited in developing more effective fish management strategies. Emitting sounds that repel the fish could, for example, be employed to guide fish movement in aquaculture systems. Underwater speakers could thus be used to deter fish away from areas where they may be in danger, e.g., near inlets/outlets or other equipment. Alternatively, this approach could also be used to steer the fish into specific areas to facilitate easier capture or handling, possibly reducing the need for handling and the use of crowding nets. In this context, the potential of the fish habituating to an originally repellent sound frequency may need to be assessed.

The findings on responses toward artificial lights could likewise be used in the development of new technological tools. Developing “fish-friendly” light sources that are less likely to have negative impacts on fish behaviour, physiology, and ecosystems is a particularly promising direction in this area. Underwater vehicles or exploration systems could, for example, be set up to use dimmer or more diffuse lighting in deep waters than when closer to the surface to minimise their impact on the natural behaviour of fish. Moreover, turning up lights gradually when a mission commences may be more beneficial than simply turning all lights on as it may prevent the fish being startled. While such a reduction in light intensity would also reduce visibility for the cameras on the ROV, this can be countered by using more advanced cameras that can operate on lower light levels.

The expanded sonar based method for identifying fish distribution patterns developed in this study could also be used in other studies aiming to scope the responses of fish toward external stressors. While the present study focused on the effects of sound and light, there are several other acute or chronic factors known to have impacts on fish behaviour and distribution patterns (Oppedal et al., 2011). In cases where such factors are localised (i.e., can broadly be considered point sources), the same method based on sonar data and DL could shed light on their impacts on the fish. The method could also be further developed as a tool for continuous assessment of noise/light pollution at farming sites since it will only require the permanent deployment of a sonar attached to a topside computer for processing. This would be a simpler technical solution than the conventional approach of using stationary echo sounders as one Ping360 unit could cover a volume that would otherwise require several echo sounders. The resulting data could allow the correlation between sound/light levels and behavioural changes in the fish (e.g., feeding patterns, stress responses, and swimming activity). Such a tool could help the farmer adjust operations to reduce the elicited response, which may in turn lead to improved fish welfare during production.