For thousands of years, aquatic resources have been a vital source of food for communities in coastal and marine areas. Several studies have highlighted the importance of this precious food in their daily lives. The first examples of humans exploiting marine resources, in various forms such as fishing, aquaponics and aquaculture occurred in the ancient civilizations of Hellas and China.

Production growth in aquaculture has been remarkable over the recent decades, with global production increasing from 2.6 million metric tons (mt) in 1970 to 87.5 million mt in 2020 (FAO, 2024). Intensive production of carnivorous finfish has developed into one of the most prominent sectors within aquaculture, largely because of the high nutritional value of the end product and its commercial appeal, and because recent advancements in farming methods have rendered this production form efficient and affordable.

The salmon industry has been considered a particularly successful segment in this sector, and has enjoyed rapid growth much due to significant research advancements and innovations within production forms and technology for salmon production in Norway (Afewerki et al., 2023).

There exist a plethora of different systems for the commercial production of Atlantic salmon. The most common practice by far is farming in marine fish farms featuring net pens suspended from floating structures that in turn are moored to the seabed. To extend the volume of the net pens, and hence the volume available for the fish, their lower ends are typically attached to weights and sinker tubes. The porosity of the net wall makes the net pens open to the surrounding water such that crucial cage environment variables such as oxygen and temperature are directly provided from the ambient environment (Klebert et al., 2013). This has stimulated the development of many new innovations within salmon farming practices and methodologies designed to counteract such challenges (Barrett et al., 2020; Føre et al., 2022) resulting in several sustainable approaches and concepts using emerging technologies (Araujo et al., 2022). One such approach is closed production where the fish are kept in units that are closed to the environment, thereby reducing the exposure to environmental features that may be negative for fish welfare, sustainability and production capacity (Barrett et al., 2020). Closed production can either be realized in marine farms or land-based systems, and specially designed feeding and sensors systems for such units can provide a technological foundation to ensure suitable conditions for the fish. Another concept that is gaining interest is the movement of production to more exposed locations further from shore (Bjelland et al., 2025). Exposed farming operations are often based on open cages much like conventional farming methods, and are based on the assumption that the waters further from shore feature fewer parasites and pathogens than coastal archipelagos and fjords. While the main motivation behind the development of these new concepts is to avoid many of the challenges in conventional farming, they are also industrially important when considering the development of the finfish market, as it is unlikely that one or a few production forms alone will be sufficient to meet future demands.

Irrespective of the production format, industrial salmon farms feature large volumes containing high fish densities, and day to day farm management usually entails regular operations such as feeding, cleaning and farm maintenance, and otherwise seeking to avoid disturbing the fish. It is possible to do some experiments in such venues to for example, monitor fish responses to environmental conditions (Johansson et al., 2006), assess growth over time (Føre et al., 2016) or assessing physical phenomena like hydrodynamics (Lader et al., 2008; Winthereig-Rasmussen et al., 2016; Klebert et al., 2015).

However, commercial facilities are often not suitable arenas for preliminary studies targeting more detailed or individual based responses or features, since this often requires a tighter control of the experimental conditions than what is possible to achieve in such systems. This has resulted in that most aquaculture related experiments are conducted in experimental facilities specifically designed to allow closer control of the conditions and continuous monitoring of the fish. To fulfill these requirements, these facilities are often designed and run in ways that differ greatly from the commercial facilities. In addition, most research infrastructures are realized with land based tank systems mainly because the infrastructure (i.e., power, internet, water supply) required for experiments is much easier to get on land than at sea. Research units are also often smaller than industrial tanks and cages, previous studies have found that the scale of production units have an impact on the fish, even affecting their growth (Espmark et al., 2017). Finally, the production environment in these type of tanks is usually different from those fish will encounter at farms, especially in crucial factors such as lighting, flow fields and temperatures.

Based on the disparities between the conditions in commercial farms and those in a standard land based tank, it is apparent that researchers should develop setups that approximate the conditions on commercial farming sites as well as possible to achieve industrially relevant outputs. The most important measure towards ensuring this is to apply environmental enrichment. Enrichment in aquaculture is mainly used to provide the fish with stimuli that have positive impacts on fish welfare mainly to ensure improved animal welfare and more ethical production, but also because several studies have indicated the importance of fish welfare in providing optimal production environments and better practices (Aubin et al., 2009). Results from the introduction of enrichment measures are usually evaluated by interpreting fish behavior as observed by sensor systems in the production unit (Arechavala-Lopez et al., 2022). The enrichment process can also be used to provide a more realistic farming environment for the fish, thereby contributing to closing the gap between research facilities and farming sites.

A major topic within enrichment is the use of artificial illumination and several surveys have outlined the advantages of such practices regarding fish swimming behavior and growth. In a study conducted by one of the authors of this study and collaborators, artificial lights were used to control vertical fish movements and distribution patterns, offering insight into how fish can be steered through this stimuli (Føre et al., 2013; Wright et al., 2015; Herbert et al., 2011). In Wright et al. (2015), specially designed artificial lights were placed in aquaculture cages containing Atlantic salmon, the outcome of which was that the fish exhibited modified swimming patterns in response to lighting. Herbert et al. (2011) did a similar study, albeit in tanks, and found similar results. Together, these results imply that distribution and swimming depth may be possible to control using artificial lights, a feature that was further explored through mathematical simulations by Føre et al. (2013). In more recent studies, scientists have also investigated the swimming pattern and the growth maturation of the salmon in real aquaculture units, outlining the growth effect regarding external light sources (Hansen et al., 2017). Another enrichment measure relevant for fish welfare is to apply appropriate colors to the interior of the production unit, which is especially relevant for tank based trials. Earlier studies have explored the impacts of various color intensities on fish performance, and found different colors to elicit different responses (McLean et al., 2008). These results have later been reviewed, showcasing the importance of tank colors, and asserting that certain fish species perform better when subjected to specific colors (e.g., green, blue and black) than others (McLean, 2021). While many of the conditions at fish farms are hard to replicate in lab setups, the intelligent use of artificial lights and tank color can, in combination with camouflaging instruments and devices, render the visual tank environment perceived by the fish more similar to that in a commercial production unit. This could represent a first step towards achieving farm realistic conditions in experimental tank setups.

Since fish welfare is not easy to observe directly, research aimed at welfare assessment and monitoring often need to target other observable variables that are believed to be linked with the welfare of the animal. Based on Norecopa (Norway’s National Consensus Platform), welfare is defined in terms of animal’s perception (intrinsic experience) of itself. Since it is impossible to observe welfare in such terms directly, it can only be evaluated by proxies, underscoring the importance of current study. Variables derived from behavior are often easier to monitor than those linked with physiology since behavioral expressions usually entail spatial movements that are externally detectable. Furthermore, fish behavior has in itself intrigued the scientific community, as it can result in important information about the interactions between species, different environmental changes, and external stress factors. In light of recent developments toward replacing manual labor with automation and technologies such as precision fish farming (Føre et al., 2018), there has been a drive towards adapting new technologies for more detailed, continuous and robust observation of fish behaviors in aquaculture. Several studies have thus aspired to observe fish behaviors in fish farming systems using different sensing principles systems (Cui et al., 2024; Burke et al., 2025). Methods exploiting remote sensing principles including vision based sensors such as cameras and other optical instruments (Saberioon et al., 2017) and hydroacoustic devices (Li et al., 2024a) are particularly interesting due to their non-invasiveness and ability to gain information on both individual fish and groups.

Vision based sensors have become the most commonly used technology for monitoring fish in aquaculture, and have thus had a leading role in leveraging technological solutions for more accurate fish management. Several studies have used waterproof cameras to estimate fish morphometric parameters (e.g., Shi et al., 2020; Voskakis et al., 2021; Chuang et al., 2015) and behavioral patterns (e.g., Georgopoulou et al., 2021; Eguiraun and Martinez, 2023; Burke et al., 2025). Other surveys have applied cameras after retrieving the fish from the aquatic environment to accurately measure various fish parameters such as fish length (Hsieh et al., 2011; Monkman et al., 2020). The advancement of optical sensor technology has also resulted in different types of camera applications and the development of innovative processing methodologies providing new insights (Pinkiewicz et al., 2011; Zhou et al., 2019).

While hydroacoustic technologies have mainly been used for underwater vehicle navigation, inspection and guidance (Skaldebø et al., 2024; Zerr et al., 2005; Hansen et al., 2003), emerging studies have highlighted the utilization of such technology in aquaculture settings (Li et al., 2024a). The most common application of sonars (single beam) or echo sounders has been to study the spatial distribution of fish (Tao et al., 2010; Boswell et al., 2007) often providing insights into specific patterns such as vertical distribution (Oppedal et al., 2007). Others have sought to study individual fish swimming behavior (Schwarz, 1985; Plimpton et al., 1997; Arrhenius et al., 2000) or individual sizes (Knudsen et al., 2004) using split-beam sonars placed in sea cages. More recent studies have sought to explore the potential of even more complex active hydroacoustic systems such as multibeam sonars in aquaculture, demonstrating the possibility of obtaining 3D spatial data (Kristmundsson et al., 2023).

As emerging technological advancements are rapidly transforming many aspects of research and daily life, it is imperative to preserve the concept of natural development processes of living organisms. The present study introduces the cyber-enhanced tank (CET) which is a conceptual framework that bridges the gap between experimental tanks and commercial production facilities. To demonstrate the CET concept, we have expanded an existing physical tank setup with technological means including a novel adjustable submerged lighting system capable of simulating the conditions in an aquaculture farm unit and an innovative sensor suite for fish monitoring. This is an updated and expanded version of a previous research outcome (Voskakis et al., 2024) and offers an arena for conducting controlled experiments in a more realistic environment and which offers multiple non-invasive modes for observing fish behavior. Increased realism is sought achieved by designing the lighting system to emit a light spectrum resembling that of natural sunlight and to follow the natural photoperiod and that is not perceived as a point source by the fish. A further enrichment measure towards this end is to choose the tank wall color from a selection of colors previously identified as having positive impacts on fish welfare. The sensor suite is specifically designed for behavioral monitoring and targets several features of the fish behavior by simultaneously using four different sensors (i.e., a surveillance camera, stereo vision, an event camera and a scanning sonar). In addition to providing multi-modal observation, the sensor suite combines technologies that have seen little or no use as research tools in tank based experiments with more conventional observation methods, thereby enabling new methods for monitoring fish. These sensor systems are either camouflaged or mounted outside the production volume to provide a “clean” tank for the fish with few disturbances. While this study outlines the design and features of a specific system developed for fish experiments, the overall objective of this work is to introduce the idea to the research community that technological enhancements of research facilities can help achieve results that are more industrially relevant also under controlled laboratory conditions.

The cyber-enchanced tank was built around a conventional indoor square fish tank (dark green colored) with dimensions 2 × 2 × 1 m, with a volume of approximately 4 m 3 . Figure 1 shows the 3D design of the tank, demonstrating the placement of the lights and instruments comprising the sensor suite.

The sensor suite featured: 1) a scanning sonar (Ping 360, BlueRobotics Inc.); 2) a event camera (DAVIS 346, iniVation AG, Switzerland); 3) a stereo video setup (Alvium 1500 C-510 NIR, Allied Vision GmbH); and 4) a surveillance camera mounted in the ceiling above the tank (Reolink RLC 823a, Reolink). This provided a multi-modal observation method that enables a more detailed and deeper insight into fish behavior during tank trials (see Appendix Table A1 for sensor suite costs).

To ensure steady operation during data capture, a mini computer built around a Jetson AGX Orin (NVIDIA Corp.) with high performance skills (Table 6) was used. The main role of this system was to manage the two novel sensor technologies, (i.e., the event camera and the sonar), and to avoid synchronization issues the operating code was written using multithreading python.

For each recording, the operative system first set the initial parameters and recording time and then started the multithreading data collection process. This prompted the event camera to collect the event and frame acquisition datasets with timestamps, while the sonar system did likewise with sonar scans. The operative system then stored both datasets in a designated folder. Since the systems were operated by the Jetson, the resulting datasets were synchronized.

The main advantage of the LED based illumination system in experiments lies in its ability to provide light that resembles that in a fish farm both in terms of intensity and color, and how it disperses in the tank volume. This will be important in trials seeking to study how farmed fish respond to specific events or situations arising under production as it will reduce the chance that responses to the illumination confuses the outcomes of the trial. Moreover, the ability to active control the light level with a high accuracy allows for replication of natural photoperiods. Since the cyber-enhanced tank is shielded from the impact of other artificial light sources, this allows for actively steering the light levels in the tank to follow any chosen photoperiod pattern. If further combined with active control of the temperature of the inlet water, the tank could thus be set up to actively mimic the conditions in a fish farm at any season irrespective of the actual time of year. Considering that seasonal variations may have direct impacts on farmed fish (e.g., Bowden et al., 2007; Versteeg et al., 2021) and hence how they respond to other stimuli, this could open for a more efficient experimentation where seasonal variation can be taken into account without requiring the experiments to span across actual seasons.

The most apparent use of the Ping360 sonar in describing fish behaviors and distributions is to use the data to visualize the spatial distribution of the fish. While this can give some insight into the spatial distribution of the fish, this system is generally not able to give insights into individual fish behaviors due to the relatively long scan time. However, recent studies using the same sonar type have demonstrated how more advanced processing of the data could be used to get added value from the Ping360 data, and several of these could also have potential uses in the cyber-enhanced tank concept. In a recent example from aquaculture, a deep learning based method was developed and applied to observe the distance farmed salmon preferred to keep from an intrusive object with various physical features (Zhang et al., 2024). Although this study was conducted in a full-scale sea-cage and hence at a much larger spatial scale, a similar approach could be adapted for tank use to provide measures of the distances to and sizes of fish shoals observed with the sonar. Other recent studies have sought to develop AI-based methods for more accurate object detection using the Ping360 sonar. This has for example, entailed the use of a U-net based segmentation model (Hasan et al., 2024) and applying the promptable Segment Anything Model (SAM) (Tolie et al., 2024) to segment and detect various objects observed by the sonar. While neither of these two latter cases targeted fish, these or similar methods are likely transferrable to the cyber-enhanced tank application as long as they can be adjusted to take the short observation range and low power settings into account. If properly trained on relevant data, segmentation methods could ultimately be used to count and even size the individual fish observed by the sonar.

Since event cameras have never been applied in previous fish studies, there exist no examples of processing methods designed for deriving metrics relevant for analyzing fish behaviors. However, the direct analysis of the raw event output may in itself be valuable for describing the dynamics in fish shoals. Simple time series plots of the number of events in the image could as such be a good enough indicator for detecting the onset and scale of the behavioral response of fish when subjected to both acute and chronic factors. This approach could potentially be expanded to provide spatial indications on where in the image frame the activity occurs by dividing the image frame into a grid of cells and then counting the events within each cell. Such an approach could, for instance, shed light on variations in spatial distribution dynamics when the device is in topside position. The use of event cameras in other application areas led to the development of more advanced processing methods for further refining the outputs from event cameras, some of which may be relevant for describing fish. Examples of such methods include self-supervised frameworks that estimate optical flow based on event streams (Zhu et al., 2019), methods for image reconstruction based on event data (Scheerlinck et al., 2020) and real-time event-based stereo-visual odometry (Zhou et al., 2021), and approaches for fine grained object detection using event streams (Kim et al., 2021). These are just a few examples of methods emerging from the ongoing surge of studies that seek to use this technology, and several of these and other approaches may be adaptable to a fish experimental situation.

Unlike the event camera technology, the conventional cameras used for surveillance and stereo video have been a staple within fish monitoring for decades. Consequently, there exist several methods designed to derive information on fish behaviors by processing camera footage. For instance, subjecting the video stream from the surveillance camera to methods such as optical flow (Beauchemin and Barron, 1995), surface activity using Convolutional Neural Networks (Ubina et al., 2021) or entropy and fractal techniques (Eguiraun and Martinez, 2023; Eguiraun et al., 2018; Eguiraun et al., 2014) can result in data describing the dynamics in the entire fish group. Likewise, stereo camera images can be analyzed using state-of-the-art stereo processing methods, many of which have been adapted to fish (Li et al., 2024b). Such adaptations have enabled the detection and tracking of, e.g., individual movements and behavior (Saad et al., 2024), wounds and defects (Nissen et al., 2024), and fish size (Silva et al., 2024). These are but a few examples from the expanding toolbox for analyzing video footage. The inclusion of these and similar methods in the processing of the output from the sensor suite will contribute to making the cyber-enhanced tank more diverse and robust in terms of observation modes. Moreover, since these established methods are usually validated, they can be used to generate ground truth data for validating new processing methods for processing the data generated by the event cameras and the sonar.

The main motivation for equipping the cyber-enhanced tank with a diverse set of sensors was to enable multi-modal observation. While separately analyzing the data collected by each instrument is valuable in itself, the value can be multiplied if the data were to be merged into one holistic dataset describing the biological situation in the tank.

A first step on this pathway could be to look into combining the outputs from the vision based methods. Since the three methods used here describe different features of the observed fish, a combined view could potentially exploit the advantages of all methods. For instance, the surveillance camera will provide good inputs on the horizontal fish distribution and general shoal movements. Moreover, the fast sampling rate of the event camera compared with the other optical methods renders it much less sensitive to motion blur and other disturbances, meaning that the inclusion of event data in an analysis could contribute to better detection of details.

A drawback of all the vision based methods included in the sensor suite is that they require light to capture objects and events. This means that neither of these systems are likely to provide data when there is low natural or artificial light. Since the cyber-enhanced tank is intended to simulate the conditions in a fish farm including the natural photoperiod, the fish will at times be kept in darkness, especially when simulating night time conditions during winter. At such times, the optical methods will struggle to provide data on what the fish are doing. While this could, to a certain extent, be compensated with additional artificial lights. The inclusion of such lights may compromise the aim of keeping conditions as similar to a production situation as possible (unless the simulated case features artificial lights). In the cyber-enhanced tank concept, monitoring in darker periods is therefore intended to be facilitated by the Ping360 sonar. Although the sonar by itself can provide insight into behaviors during darkness, added value can be obtained by fusing the sonar output with the visual data. While such an integration may not be as seamless as when merging the three optical methods, the co-analysis of sonar data with video would have a dual purpose. Firstly, it would make the sensor suite more robust in providing yet another mode of observation. Secondly, the data collected during brighter periods when all systems are running can be used to validate the sonar output and develop methods for analyzing the resulting data to gain deeper insights into fish responses.

Finally, a long term aim will be to integrate digital twin technology (Rasheed et al., 2020) in the cyber-enhanced tank concept. This will require integrating the sensor suite with mathematical models describing the system dynamics, either in the form of Knowledge Based Models (KBM) synthesizing existing system knowledge into mathematics or Data Driven Models (DDM) that can predict response patterns based on a set of input variables (Føre et al., 2024).

The data used to demonstrate the systems were collected during preliminary trials before an upcoming stress experiment using the CET that was in compliance with the Norwegian animal welfare act under approval by the Norwegian Animal Research Authority (permit no. 30968). The full trials lasted from January to March 2025, during which the CET contatined a maximum of 150 Atlantic salmon (Salmo salar L.) post-smolts. The average length and weight of the fish were 25.56 ± 1.46 cm and 164.80 ± 25.50 gr, respectively.

Figure 6 presents an image obtained by the sonar during preliminary trials demonstrating the dynamics of a fish group in the tank volume and the tank wall are visualized. The Ping360 registers the strength of the reflected signal from each object appearing in the beam. These images were obtained using Python code developed specifically for processing the sonar data collected from the cyber-enhanced tank that mapped intensity values to values between 0 (black = low intensity) and 1 (white = high intensity) in the figure. The location of the sonar is represented as a white glowing half-circle shape at the lower edge of the image, while the tank wall, which reflects high intensity, is rendered as a white contour. White patches (Figure 6a) within the tank boundaries are objects in the tank volume, in this case, a fish shoal that is clustered at the side of the tank opposite to the sonar. Conversely, Figure 6b depicts a situation where the shoal had been startled by a sudden event that caused them to move away from the observation area.

In being a completely novel way of observing fish, the outputs from the event camera can best be described when compared with the outputs from conventional cameras. Since the event camera used in the present version of the cyber-enhanced tank can provide regular monochrome images in addition to events, it is also reasonable to compare monochrome image outputs from a motive with the corresponding events output. Two example outputs obtained during feeding events in the preliminary trials are presented in Figure 7, where conventional monochrome image frames are placed on the left (a and c) and corresponding event images are provided to the right (b and d). The image pair on the top of the figure was collected with the event camera placed topside (i.e., next to the surveillance camera), while the lower image pair was collected when the device was mounted at the bottom. Blue and black dots in the event images (b and d) indicate a decrease or an increase of the light intensity in a pixel as registered by the event camera. These examples clearly show how static details and features that are present in the regular images are not picked up by the events, while dynamic details such as movement are clearly detected as events. The two camera placements offer different insights into how the event camera can be used in that the topside position gives data on the entire tank while the bottom position in not enabling a full view of the tank, gives data on fewer fish but at a much closer distance.

Moreover, the images also show how the events are mostly concentrated around the fish, which are the most mobile objects in the image. Another detail that is apparent from these images is that event outputs from the device are practically insensitive to motion blur which can often be a major challenge in computer vision applications with conventional cameras. This is particularly clearly seen in the images taken from the bottom position in that these capture a fast moving fish at a close distance, both of which are factors that tend to increase the challenge of motion blur. While the outline of the fish is blurred in the monochrome image (left), it is sharp and clear in the event image (right). This illustrates the potential of this technology in visualizing fish even in situations where they are indistinguishable in regular images.

To provide a more concrete example of quantitative use of the events, Figure 8 shows two time series of the summed up events (i.e., number of black and blue dots in the event images) during feeding trials. The feeding time is marked by the red dotted line. While it is clear that both camera positions imply the system was able to detect the onset of feeding, bottom deployment (Figure 8a) gave generally higher number of events but more variations as would be expected since it is more sensitive to individual fish entering and leaving the observation volume. Conversely, when the system was mounted topside (Figure 8b), the number of events was more homogeneous before and after the onset of feeding, likely because it captured the entire shoal and thus was less sensitive to individual fish variations. The lower total number of events is because the fish were further away from the camera and thus represented fewer pixels.

To illustrate the raw outputs from the surveillance and stereo cameras mounted in the tank, Figure 9 presents an overview image of the experimental tank from perspective of the surveillance camera (a), and a side view of the fish collected by one of the stereo cameras (b). Both images were collected during preliminary trials in the CET and show a group of fish in a non-agitated state where the fish exploit much of the tank volume.

The cyber-enhanced tank concept outlined in this study aspires to provide an environment that simulates the light conditions that prevail in fish farming units, while offering the fish a “clean” environment where sensors and other components are concealed. In combining this with the collection of comprehensive and varied data sets describing the fish responses, the setup is well suited for analyzing fish responses to tank management. To our knowledge, this is one of the first tank based system seeking to imitate conditions in an aquaculture unit while also retaining the main advantages of a small and highly controllable experimental setup.

In conclusion, this study has introduced a new concept for how future experiments on farmed fish can be conducted in a highly controlled tank environment while still featuring conditions similar to those experienced in commercial farms. The resulting cyber-enhanced tank concept features a multi-modal sensor suite can provide deeper insight into fish responses. This will be practical when environmental conditions are challenging for automated observation due to factors such as low light intensities, high turbulence and high turbidity. The next steps in this research are to use this first version of the cyber-enhanced tank setup in experiments where detailed data is collected from the different sensors and processed to detect fish movement patterns when they are subjected to various external effects such as stressors. This will both demonstrate the utility of this concept for future use in fish experiments that aim to provide industry relevant knowledge while still being controllable, and provide necessary experiences and inputs on how the concept can be refined further to fulfill its intended role in the scientific toolbox. Being this the first study using event camera technology to observe fish responses, that will also aspire to validate and explore the utility of event cameras in detecting stress responses in fish. The resulting data will also provide a foundation for developing new methods for processing data from the sensors both separately and in combined multi-modal analyses.