Evaluating data quality in the dissolved oxygen monitoring of Atlantic salmon sea cages

Abstract

Introduction:

Although dissolved oxygen (DO) is a limiting factor for animal welfare, growth and production performance in Atlantic salmon sea cage farming, the scope of issues relating to DO variation in the industry is largely undocumented, which may be attributed to a lack of data quality in the environmental monitoring systems.

Objectives:

In this study, we aim to understand the environmental monitoring systems of sea cage farms and investigate the data quality.

Methods:

This exploratory study evaluates the quality of environmental monitoring data from over one hundred sea cage farms across 18 Nordic companies, complemented by farmer interviews. Data quality is assessed using established frameworks across three key application areas relevant to both operational and administrative contexts.

Results:

The monitoring is conducted at one horizontal point across the full farm in 76% of the time series, usually in or near a cage and as a vertical profile. Although the sampling frequency is high, the period of monitoring is usually shorter than the average production cycle and missing values and noisy data pervade the time series. Across the full dataset, over 28% of the loggers contain no values or valid measurements, while 53% additional loggers have some degree of missing values and 42% contain invalid values.

Conclusions:

Although the farms can be considered to fulfil some best practices recommended in the literature, several discrepancies and substantial data quality issues are identified, with implications for the value-creation of the monitoring system. Improvements for the three application areas of the data are suggested.

1 Introduction

Dissolved oxygen (DO) is central to Atlantic salmon welfare and growth (Fry, 1971; Kramer, 1987; Noble et al., 2018; Nilsson et al., 2022). Low levels harm welfare (Remen et al., 2012; Remen et al., 2013; Remen et al., 2016; Noble et al., 2018; Nilsson et al., 2022), which may in worst case scenarios lead to mass-mortality events (Neis et al., 2023; Jyoti et al., 2024), while suboptimal levels reduce growth and exacerbate health issues (Remen et al., 2012; Remen et al., 2013; Remen et al., 2016; Moldal et al., 2023), ultimately impacting production and economic performance (Remen et al., 2014). As DO levels decline with climate change (Diaz and Rosenberg, 2011; Pitcher et al., 2021; Zahir et al., 2024), problems related to suboptimal DO are expected to increase in Atlantic salmon farming (Aksnes et al., 2019; Moldal et al., 2023; Jyoti et al., 2024; Zahir et al., 2024). Consequently, monitoring and managing DO levels in sea cages effectively grows increasingly important for farmers (Jyoti et al., 2024; Berntsson et al., 2025).

In the largest producer globally, Norway, statutory regulations require farmers to systematically monitor their farms to ensure that DO levels are sufficient for animal welfare (Akvakulturdriftsforskriften, norwegian ministry of trade, industry and fisheries: § 22, 2018), but what does a systematic approach to DO monitoring entail? DO is a highly varied and complex parameter to measure in the sea cage environment (Berntsson et al., 2025), yet how to effectively set up the environmental monitoring in a systematic manner is not clearly indicated in the literature. To our knowledge, little is documented regarding how the industry conducts the monitoring process, but wireless sensor networks (WSN) are often used, as they enable farmers to collect spatio-temporal information across the large spaces of their farms (Kattenborn et al., 2021; Bell et al., 2022), and best practices for using sensors to monitor DO variation are available in the literature and industry standards, such as the operational welfare indices (OWIs) (Noble et al., 2018), Laksvel (Nilsson et al., 2022), and the Aquaculture Stewardship Council (ASC) certification (Aquaculture Stewardship Council (ASC), 2025), however their recommendations differ. The OWIs and Laksvel state that DO should be monitored through sensors at the point of the farm where DO can be expected to be lowest, in a vertical profile, preferably with additional points of the farm to gain a better overview of the variation across the farm (Noble et al., 2018; Nilsson et al., 2022). According to the requirements for the ASC certification (Aquaculture Stewardship Council (ASC), 2025), a certificate of sustainable production that farmers can apply for, specific requirements to the monitoring process are expected, including daily monitoring of DO, sea temperature and salinity in the morning and evening at 5 m depth with GPS coordinates and altitudes recorded. Weekly averages and daily fluctuations should be calculated. Measurements immediately downstream from the farm should also be collected, and for sites exposed to tidal variation, the measurements should be conducted at high and low tide. Here, the ASC differs from the OWIs and Laksvel on whether to measure a vertical profile, the number of measurement points in the horizontal space of the farm, and what parameters to measure. The inconsistencies in monitoring recommendations may lead to a variation in how farms design and implement their environmental monitoring process, including differences in sensor deployment and spatial coverage. This may contribute to the lack of standardisation across the industry and reduced comparability of data across production cycles and between farms.

How to determine whether DO levels are suitable for Atlantic salmon also varies, as these can be considered by assessing environmental hypoxia or through the metabolic demands of the fish (Farrell and Richards, 2009; Remen et al., 2012; Remen et al., 2013; Remen et al., 2016; Berntsson et al., 2025). Although averages are often used and recommended as a metric for welfare, scholars advise against using averages (Page et al., 2005), as averages can miss impactful fluctuations. What increases the complexity is that the literature is not clear on when fluctuations become harmful to the animals in the sea cages in the commercial context, as the fish can swim in and out of harmful environments in the cage if the suboptimal variation is local, and most existing knowledge stems from controlled environments, such as experimental cages or tanks (Berntsson et al., 2025). In contrast to the animal welfare assessments and industry standards, mathematical models for analysing cage variations that also use the environmental monitoring data as input (Alver et al., 2022; Alver et al., 2024), the optimal placements of the sensors are outside of the influence from the farm, rather than inside the sea cages.

In conversations with both scholars and farm managers, we have anecdotally learned that both seem to find little value in the DO data collected through the monitoring process, indicating that there is a problem with how this process is currently conducted. The issue is described as two-fold: (i) there is a lack of trust in the data or monitoring system, and (ii) even if the monitoring system would be reliable, farmers are uncertain of how to mitigate negative effects in operations or see little operational value in the environmental data. The cause of (i) may be due to poor data quality (DQ), where scholars reject the data and farmers prefer to use other methods than the monitoring system to identify welfare problems in the population (Zhang and Su, 2025), such as assessing behaviour or appetite. According to the literature, mitigating negative effects and resolving parts of (ii) is theoretically possible (Quiñones et al., 2019; Alver et al., 2024). As summarised by a recent literature review (Quiñones et al., 2019; Berntsson et al., 2025), both farm configuration and daily routines could be adjusted to the limitations of the sea cage environment to minimise the negative effects on the animals and production. Yet DO is rarely used as an optimisation parameter in production. A key problem for learning about the problem and testing optimisations is the limited insights into the sea cage environment, reducing the opportunity for producers to act on reliable information. While evidence of poor DQ in the environmental monitoring system of Norwegian sea cage farms is reported in the literature (Falconer et al., 2025), we have not found reports of the current state of DO conditions in Norwegian sea cages, beyond statements that DO is becoming a larger issue in fish health reports (Moldal et al., 2023). Although DO is mentioned in these reports, DO is rarely included in data analyses or as a parameter in mortality, fish health or welfare studies (Oliveira et al., 2021; Østevik et al., 2022; Persson et al., 2022). Poor DQ and the lack of insight into the sea cage environment contribute to uncertainty of the problem scope, and thus, the potential and value of addressing DO-related issues. This may create a self-reinforcing cycle in which low confidence in the monitoring data limits its operational use, while the absence of perceived operational applications reduces incentives to invest in improving monitoring quality. Breaking this cycle requires both improving the reliability of environmental monitoring data and better understanding how such data can support operational decision-making. Furthermore, as the interest in digitalising aquaculture operations grows to apply precision fish farming technologies and advanced analytics to the large datasets collected in operations, poor DQ and a lack of structure hinder the development of data-driven decision-making (DDDM) (Føre et al., 2018; Pratiwy et al., 2022; Føre et al., 2024) and precision fish farming applications, automated feeding (Føre et al., 2018). As an initial step to understanding how to resolve the issue of limited insights and learn the scope of DO problems in the sea cage environment, we first need to understand how this process is executed in the industry today. To investigate how the environmental monitoring process is conducted in the sea cages and whether it can be considered systematic in relation to the existing literature on DO variation, we study the monitoring data from over a hundred sites from 18 Nordic companies. We assess the quality of data through the lens of the DQ literature, as presented in the following section, and present a unique current state map of the environmental monitoring of sea cage farms spread out across Norwegian and Faroe Islands production sites. Based on the results, areas of opportunity to improve the monitoring process are explored. In this study, the data from the production environments are studied in the format that it is collected by the farmers, and consequently, we have limited control over factors such as manual registration errors in the dataset.

2 DQ and environmental monitoring

To set up an environmental monitoring system that creates value for both the farmers and administrators of the industry, understanding how to effectively manage DQ is important. Managing DQ is vital for all organisations that rely on information from operations and aim to achieve DDDM (Khong et al., 2023), as a decision made from quality data from operations often has positive impacts on production performance and a decision made from poor quality data risks the opposite effect (Haug et al., 2011). This section explores and summarises relevant literature on DQ management, and suggests applications in the context of monitoring DO variation.

2.1 DQ in the literature

2.1.1 Master data management

Master data management (MDM) is a central concept for succeeding in turning data into information in DDDM, as it enables organisations to harness their data across complex operations and multiple functions, while poor MDM contributes to ineffective data processing and reduced quality in subsequent analyses (Cichy and Rass, 2019; Pansara, 2021; Malik and Khan, 2024). MDM consists of metadata governance, DQ management, data integration and stewardship (Malik and Khan, 2024), where the first two components are particularly relevant to this study.

2.1.2 Metadata governance

Metadata contains additional information about the data that provides context to the user, such as the structure and capture of the data (Liaw et al., 2021). In the context of environmental monitoring, this could state what parameter is measured, how and where it was measured, for what purpose, details on maintenance of sensors and whether any deviations occurred during data capture, for instance, if the sensor is removed due to well boat operations. As DQ concerns whether the data meets the expectations for its intended use (Earley et al., 2017; Lighterness et al., 2024), specifications of these expectations in the metadata help guide the user on how to understand and interpret the data correctly. The documentation and specifications on what contextual information should be included in the metadata are termed metadata governance (Malik and Khan, 2024).

2.1.3 DQ management

Various frameworks for managing and measuring DQ exist in the literature (Cichy and Rass, 2019), where some focus on specific industries, including healthcare (Weiskopf and Weng, 2012; Lighterness et al., 2024) and building energy (Morewood, 2023). To our knowledge, none are oriented towards aquaculture, although some papers evaluate the parameters possible to include in the environmental monitoring setup (Su et al., 2020; Bell et al., 2022). In the literature on DQ frameworks, a variety of dimensions to evaluate the DQ exist, but according to a comprehensive review (Cichy and Rass, 2019), the most commonly used ones are listed below.

2.1.4 DQ dimensions

• Completeness, whether the expected data are present in the dataset, and if they are of sufficient breadth, depth, and scope for the task at hand.

• Accuracy, whether the data correct and reliable, and if the data represent the truth of the conditions they measure.

• Timeliness, whether the data are delivered to the user in a timely manner for their application, that is, if its age is appropriate for the task.

• Consistency, whether the data format is consistent and compatible with previous data.

• Accessibility, whether the data are easily available to the user.

Validity and uniqueness are dimensions used in some industrial frameworks (Morewood, 2023; Lighterness et al., 2024), where validity investigates if the data is within expected ranges and uniqueness if, the dataset contains duplicates of data points.

2.1.5 DQ metrics

To evaluate whether the dimensions are achieved, DQ metrics can be used as a tool to define and assess whether these dimensions are met contextually in an industry (Fu et al., 2024). These should be developed to consider the value of the data according to its application, not just as it is specified currently, but also potential future applications.

2.2 Applying DQ management to the environmental monitoring system

To secure DQ in the monitoring system, defining what the data will be used for, how the data will be collected for this use and specifying the context in the metadata is an important first step (Liaw et al., 2021; Malik and Khan, 2024). In this study, data is collected using WSN, and relevant DQ management is considered in this context.

2.2.1 Metadata governance of the monitoring system

As defining the use of data is key to DQ, three possible applications of the WSN data from the environmental monitoring are identified and described in Figure 1, including planning and forecasting of production, real-time information for operations and reports to study the environment post-production (Antonucci and Costa, 2020; Moldal et al., 2023; Føre et al., 2024). These applications are used to define how data should be collected in the environmental monitoring. Defining necessary metadata specifications and the subsequent development of the data collection strategy must consider what data is required for these applications and how the technology creates value accordingly. WSN can be a highly useful tool to farmers, as it enables remote insights into different conditions across the large physical spaces of a farm that would otherwise not be accessible to the farmer and can capture variation that occurs in space and time (Kattenborn et al., 2021; Bayih et al., 2022; Adewuyi et al., 2024). However, the choice of parameters and placement of the sensors collecting data influences the insights that can be derived from the data, and must be determined in accordance with the intended application: one sensor measuring DO variation at one physical point every five minutes will provide insights into this one point and the change that occurs over time in this resolution, while two or more sensors or measurements will provide insights into those specific points and the variation that occurs between those points in both time and space (Burke et al., 2021; Alver et al., 2024). As the DO variation in sea cages varies with many factors, in both short and long time periods, in the horizontal and vertical space within and between cages at the same site [reviewed by 16], these points must be considered carefully in accordance with the intended use, whether that is reporting responsible conditions, optimisations or achieving Precision Aquaculture applications.

Figure 1

As recommendations for monitoring DO in relation to reporting welfare conditions differ from DDDM and model applications (Noble et al., 2018; Nilsson et al., 2022; Alver et al., 2024; Aquaculture Stewardship Council (ASC), 2025), balancing what data collection is sufficient to satisfy the requirements for reporting, while generating the greatest value-creation of the system, can be analysed through developing specific DQ metrics, by comparing the costs of DQ management to its value-creation (Haug et al., 2011; Fu et al., 2024). One example of this in the environmental monitoring is the value of accurate DO data in relation to feed costs, for, the application of optimising operations in real-time (Figure 1). Suboptimal DO reduces maximum feed intake for a given sea temperature (Remen et al., 2014; Remen et al., 2016), and if data with low accuracy indicates to the operator at the feeding central that there are optimal DO conditions and appetite is high, when in reality the conditions are suboptimal, the cost of that feed waste could be used as a metric to determine the cost of accurate vs. inaccurate data. DQ issues related to WSN on farms are documented in the literature, caused by issues such as occluded sensors, degrading materials, disturbances from debris or production, maintenance demand and weak signal transmission (Selle et al., 2023; Jeong et al., 2024). This can cause irregularities such as outliers, repetitive measurements, systematic errors, temporal and spatial inconsistencies, datasets with zero standard deviation, i.e., constant value across single as well as multiple parameters, and datasets with exceptionally high or low values (Selle et al., 2023). Relating back to the importance of MDM, detailing information about known irregularities, such as maintenance, calibration or removal due to operations, is helpful in assessing such problems.

2.2.2 DQ management of the monitoring system

In Table 1, the DQ dimensions are applied in the context of the monitoring system, with simple metrics to assess to what extent the data fulfils the individual dimensions in relation to DO variation (Fu et al., 2024; Berntsson et al., 2025). The metrics are not quantified in terms of value-creation, as this study does not have access to the farms’ relevant financial or organisational data.

3 Materials and methods

3.1 Data sources

Multiple data sources are used in this study. The main data source is environmental monitoring data from multiple companies and sites. Access to the data was provided by a Norwegian environmental logger vendor after the companies agreed to share their data. The data is accessed by logging onto the vendor’s cloud platform using the login information from each company. To provide context to the monitoring data, we complement it with statements from qualitative interviews conducted with representatives of eight of the participating companies.

3.1.1 Technical specifications

The environmental monitoring of the companies consists of environmental loggers, equipped with sensors measuring specific parameters at one or more depths. The sensors measure one or multiple parameters, including DO, sea temperature (T), salinity (sal), sea current, turbidity and feeding barge.

The technical specifications of the equipment are available on the vendor’s website. These state that the standard measurement period is ten minutes and, as a standard, upload data every six hours to the cloud platform, where the measurements can be read and downloaded by the user. The sensors are described as optical and maintenance-free, with a battery life of twelve months, which can be shorter or longer, and with a charging period of 24-35 hours. Temperature tolerance spans 1-40°C, and the loggers have a measurement range of 10-200% DO. Accuracy is within 1% of the true values.

3.1.2 Data extraction

The data was extracted twice. Initially, the data was manually extracted from the vendor’s cloud platform between March 15^th and April 2^nd, 2024, named Dataset A. This data is stored in spreadsheets for each individual company and subsequently detailed in one combined dataset in a separate spreadsheet, which was anonymised and used for analysis. This involved the manual extraction and recording of key metadata fields for each site and company, including logger type, position, depth, deployment period (start and end dates), and the number of loggers deployed per site. Only variables relevant for describing monitoring configurations and enabling cross-site comparisons were retained. Companies and sites without DO measurements were excluded from further analysis. To reduce the risk of human error during the extraction process, anonymised site and company identifiers were used to track intra-farm and inter-company relationships, including overlapping logger deployments. Furthermore, all extracted entries were cross-checked against screenshots of the platform captured at the time of extraction to verify consistency and account for any subsequent updates to the platform data. To ensure consistency in the interpretation and recording of metadata, all manual extraction and verification steps were conducted by the same team member using a standardised extraction procedure.

The data was subsequently automatically extracted on October 21^st, 2024, by scraping the HTML-code of the cloud platform for metadata, and APIs for the logger messages, i.e., the time series with recorded measurements by the sensors. This data is referred to as Dataset B, but when necessary to separate between the metadata and logger messages, “Metadata” and “Logger Time Series” are used as distinctions. Some discrepancies between the datasets were identified, and these are detailed in Section 3.1.3. Due to certain limitations in the datasets, they are suitable for different analyses, as detailed in Section 3.1.4.

3.1.3 Data description

As listed in Table 2, 18 companies share data from 106 sites across both datasets. The number of sites per company ranges from 1 to 21 (Dataset B) or 22 (Dataset A), with an average number of 6 sites per company and a median of 4, indicating a positive bias with some companies having a large number of sites registered. The count of sites per company in Dataset B is visualised in Figure 2, along with their geographical position, and Table 3 shows the frequency of the specific number of loggers registered per site. However, the number of sites per company does not reflect company size, only how many sites on which equipment from this vendor is used. The number of loggers registered in the datasets varies, where at the point of extraction for Dataset A, 278 loggers were registered at the sites, while at the point of extraction for Dataset B, this number increased to 280 in the Metadata, but, noteworthy, reduced to 265 in the Logger Time Series. This is because 15 loggers have no time series data registered in the platform at the point of extraction, although their metadata is registered. Two sites were identified as storage units, belonging to companies C5 and C8, and are excluded from site-related analyses, reducing the total number of sites to 104 for site-related analyses.

Figure 2

In Table 4, the logger types, specifying what parameters their sensors measure, are listed with corresponding counts. Most loggers measure DO and T, followed by the same parameters in combination with salinity. Only 20 sea current loggers are found across the full dataset, and five measure turbidity. The loggers have been active within the period 2019-2024, although the main bulk of data is from 2024 and 2023, and consequently, represent current practices.

3.1.4 Data analysis and processing

This study is structured as a case study, where multiple farm sites utilise the same vendor of monitoring equipment. The use of the equipment across different farm sites and companies is explored in this study, based on the literature on DO variation and DQ framework that is described in Section 2.

The monitoring is studied from two perspectives: how the system is set up in the physical environment and the DQ of the time series with records of logger measurements. The physical environment is analysed by assessing the number of loggers and combinations of measured parameters used at the same site, in the same period, the physical positions and time resolutions of the monitoring. The DQ of the time series is analysed by assessing the degree of missing values, validity of values and noisiness. A central challenge in DQ management for DO measurements is distinguishing anomalies that convey potentially important environmental information from measurements that are likely invalid or indicative of sensor malfunction. Because measurement errors often manifest as extreme values that resemble genuine environmental events, anomaly treatment is difficult.

Information regarding the physical placement of loggers and logger types is specified in some of the names of the loggers or on the cloud platform. The metadata in Dataset B did not contain this information beyond that which is registered in the logger names. Attempts to extract the data through keyword mapping or expected value intervals were ineffective, due to inconsistent structure and format of names and large deviations in recorded values. Consequently, manual analysis was necessary, and Dataset A was used for analyses on the number of loggers per site, the combinations of measured parameters and their physical position on the farm and was conducted in Excel spreadsheets. The time resolution and time series analyses were mainly performed on the Dataset B data, using Python scripts, due to more exact data on time periods.

On a final note, as the purpose of this study is to understand DQ issues, no further data treatment was conducted on either of the datasets beyond excluding sites with no loggers and no DO loggers for the full dataset, and excluding sites identified as storage sites from site-related analyses. Where these are excluded, it is specified.

3.1.5 Anonymisation

The vendor of the equipment and the participating companies are anonymised in the dataset, using keys C1-C18 for the different companies in the analyses. Individual sites and loggers are anonymised by IDs that cannot be traced back to the individual companies or their specific sites.

3.2 Limitations of the dataset

This study has some limitations, as listed below.

• Limited contextual information: Although we conducted qualitative interviews to supplement the available data, we do not have insights into contextual data, including whether biomass is present at the site for the period of the logger measurements.

• Manual recording of data points: Manual registration of data and a high degree of manual analysis provide an inherent risk of manual errors. However, this method choice is considered necessary due to the lack of standardisation and structure across the dataset.

• This study is limited to analysing DQ in the environmental monitoring data and does not include relevant data on parameters beyond the environment that may influence DO variation (Berntsson et al., 2025).

• All monitoring data originated from a single vendor and participation was voluntary. These conditions enabled the large scale of the study, but may introduce bias, as discussed in Section 5.

4 Results

4.1 Setup of the monitoring system in the physical environment

4.1.1 Logger setup

Most sites are registered with more than one logger across the space of the farm (59%) and range from one to 14 loggers per site, excluding the sites registered as storage units (Table 3). The average and median number of loggers registered per site is three and two, respectively, indicating a slightly positive bias due to some large outliers. This is consistent with the results of the qualitative interviews, where most informants detail that they have one to four loggers per site, and only one state that they place loggers in each cage of the site. Yet, when combining the time period and the number of loggers active within the same period, the majority of the sites do not have multiple loggers with overlapping time series (Figure 3). Here, the results show that 76% of the time series recorded by the individual loggers do not have additional loggers active at the same site for that time period. When adjusting for the time series overlap, both the average and median number of active loggers per site are reduced to one. These results are in contrast to the qualitative interviews, where only two informants stated that they use one logger per site, and another uses one to two loggers per site.

Figure 3

The two smaller pie charts in Figure 3 show the combinations of parameters monitored in the loggers, where one shows the distribution of one logger and the other shows multiple loggers with overlapping time series. The simplest monitoring process in sites with loggers with overlapping time series has two loggers with DO and T at one depth. The most complex processes are found on the sites that utilise most loggers, including multiple DO, T, sal, turbidity and current sensors at the same time.

4.1.2 Logger position

Figure 4 shows that the most common placement of the loggers is in or near a cage, but whether it is in or nearby is not specified. However, through the qualitative interviews, informants from the companies that place the loggers inside the cages all state that they place them between the centre and the edge of the cage. Furthermore, most informants state that they place the loggers in or near the cage they believe has the greatest risk of low DO throughout the production period, after assessing whether the monitoring equipment will be in conflict with other equipment or activities. The argument for the placement varies between the informants. Half of them state that the direction of the sea current is the determining factor of where they will find the lowest DO, and most of them believe that the poorest environment is found some distance into the farm in the direction of the current, except for one, who believes it will be in the furthest end of the farm in relation to the sea current. One informant alternates between considering biomass or current direction and continuously assesses where they believe the poorest environment is. Another informant places loggers at multiple points of the farm and, along with the company having loggers in all cages, specifies that they use multiple points to get an overview of the variation across the farm and consequently do not need to consider where the worst conditions would be. In the dataset, the location of 85 loggers was not specified, which is the largest category after the cage-position, followed by fleet, reference point and storage, despite excluding the sites identified as storage sites. Through the interviews, we find that reference loggers are most often placed on the fleet, but some have buoys upstream. The farms have 7-16 cages per site, and the most common placement of loggers is near or in cage three, which is confirmed both through interviews and positions specified in Dataset A.

Figure 4

Of the 257 loggers where the number of depths can be deduced, 22% are identified with one depth, 49% with two depths and 21% with three, meaning that 70% have some form of a vertical profile. 8% of the loggers cannot be deduced. Figure 4 shows the depths registered most frequently in the dataset, where 5 m and 15 m are most commonly used. These results are inconsistent with the findings from the qualitative interviews, where 50% monitor one depth, at 5m, and the remaining 50% monitor two depths, either at 1 m and 7 m or 5 m and 15 m.

4.1.3 Time series length and sampling frequency

Across the dataset, the time series of the loggers have a median of six months and an average of eight months, indicating a positive bias, as shown in Figure 5. The time series ranges from no full months of measurements to nearly 50 months. Not all loggers record measurements consecutively for the entire time the logger is active, but only for brief periods. The average time between measurements uploaded to the platform is 14 minutes, with a median of 10 minutes, indicating that the data is usually uploaded to the platform at the sampling frequency, rather than the six-hour standard in the technical specifications. 166 loggers have time series gaps beyond six hours across the dataset, of which 88% have gaps for more than one day, and 46% have gaps longer than 30 days. Both the length of the time series and the sampling frequency have a positive bias due to extreme outliers. In the sampling frequency, the maximum time between measurements is nearly three years, and the standard deviation is 22 hours, slightly more than one measurement per day. The companies with loggers having extreme gaps in the time series overlap with the companies that have loggers with extreme outliers in time series length.

Figure 5

4.2 DQ in the logger measurements

In this section, all analyses utilise the time series data with logger measurements in Dataset B. DQ issues here can stem from a variety of events, as detailed in Section 2.2, that may show up in the data as missing or invalid measurements. From the qualitative interviews, we gather that all participating farms conduct frequent maintenance on their loggers, ranging from every one to four weeks and clean them every one to two weeks during the high on-growing season. During the period the logger is not in the water, they will use a handheld or additional logger from storage or another cage. Furthermore, most of the informants stated that they move the logger during the production period, during specific activities or if the conditions for where they expect the lowest DO to be (if the farm considers biomass to have the highest impact on DO variation, they will move the loggers to the cage with the highest biomass, which may change during the production period). This section explores how missing values and invalid measurements appear in the quantitative time series data.

4.2.1 Missing values

NaNs and 0.0 values are flagged as missing values in this analysis. The missing values are distinct from gaps in time series, as the time series gaps are longer periods between two timestamps, while the missing values, NaN or 0.0, are missing measurements of specific timestamps. Out of the 265 loggers, 56 (21%) are identified with only missing values in their time series, and 59 (22%) are not flagged with any missing values. 141 (53%) of the loggers have some degree of missing values, spanning from < 10% to > 90% of all recorded measurements. Most have less than 10% missing values, although around 30% are identified with > 10% missing values, and 5% have > 75% missing. An example of how the missing values may present in the time series is included in Figure 6A.

Figure 6

4.2.2 Noisiness

Some degree of variation is expected in all parameters, so loggers with periods of no variation are unlikely to be valid measurements. Beyond the loggers with only 0.0 or NaN values identified as missing values, 18 loggers are identified with zero standard deviation across the full time series, indicating that no true values are captured by these loggers either. Example B in Figure 6 is an example of how such a time series can appear in the dataset.

In the dataset, DO saturation occasionally drops to 0% or rises to several hundred percent. Although supersaturation above 100% can occur in sea-cage environments and values above 150% may occur during periods of intense biological activity, such as algal blooms, extremely high values (600% saturation) are unlikely to be physically plausible. Likewise, prolonged periods of extreme supersaturation are difficult to reconcile with expected environmental conditions and may rather indicate poor data quality or sensor malfunction. Across the full dataset, 42% of loggers recorded DO saturation values above 150%, while 33% recorded values above 200%. The frequency of measurements above 150% varied substantially among individual time series, ranging from only a few observations to more than 25,000 observations, corresponding to approximately 0-100% of recorded measurements. In total, 12% of loggers exhibited recurring occurrences of values above 150%, while 5% recorded values above 150% for more than 15% of all measurements. Furthermore, loggers containing values above 150% frequently exhibited mean supersaturation levels between approximately 200% and 500% during these periods, suggesting the presence of a subset of devices characterised by persistently elevated or anomalous readings.

5 Discussion

Before discussing the implications of the results, the limitations of the dataset outlined in Section 3 should be considered. First, the analyses are based on data extracted from a single vendor, and some observed data quality issues may therefore be influenced by characteristics of the vendor’s equipment and data platform. Additionally, the logger data was shared voluntarily by companies agreeing to participate in the study, which may introduce self-selection bias, as participating farms may not be fully representative of the broader industry. The qualitative interviews also include a subset of these companies, which may further influence the perspectives captured. Second, limited contextual information was available for the monitored sites, and the analyses are restricted to the environmental monitoring data, rather than the broader set of biological and operational factors that may influence DO dynamics. Nevertheless, recent literature reports similar data quality problems (Falconer et al., 2025), suggesting that the findings are not solely attributable to the specific companies or vendor platform included here. As the dataset covers 106 sites from multiple companies, it provides a unique opportunity to investigate current environmental monitoring practices at a scale that has not, to our knowledge, previously been reported.

5.1 DQ in the environmental monitoring systems

In this section, the degree of DQ is evaluated and the implications of the results are discussed.

5.1.1 Evaluation of DQ in the environmental monitoring

The fulfilment of DQ dimensions by the results is evaluated using the DQ metrics developed from the literature on DQ (Table 1). Although Fu et al. (2024) recommend developing quantifiable DQ metrics, this evaluation is limited to an initial, overarching assessment of the relevant dimensions, due to limited contextual insights into the organisations that shared data. The achievement of the individual metrics for each dimension is assessed in the context of each of the three applications, reporting, real-time application and planning and forecasting (Figure 1).

5.1.2 DQ score

The matrix in Figure 7 summarises the performance of the monitoring system for each DQ dimension and metric, as discussed in the evaluation (Table 5). While some metrics are achieved for one or all applications, the majority are not and the DQ is consequently considered insufficient and likely to fail to create confidence in end-users (Fu et al., 2024), supporting the anecdotal statements (i) detailed in the introduction of this paper.

Figure 7

5.1.3 Implications of DQ performance

Overall, the DQ score shows unsatisfactory results, which has implications for the value-creation of the monitoring system for these farms and the administrators of the industry (Kattenborn et al., 2021; Fu et al., 2024). If farmers and administrators utilise this data for insights into the sea cage environment and subsequent decision-making, it is associated with certain risks (Haug et al., 2011). These are discussed below in the context of the sea cage environment.

5.1.3.1 Incomplete and inaccurate data fail to provide increased knowledge of the conditions in the sea cage environment

The failure to meet DQ dimensions related to completeness and accuracy indicates that the monitoring system does not sufficiently support an increased understanding of the sea cage environment. Critical environmental variations that influence DO dynamics remain largely undocumented. As a result, recurring patterns of suboptimal DO variation cannot be detected through data analysis (Ahmed and Shakoor, 2025; Berntsson et al., 2025). Although vertical profiling is common across the farms and supported by best practice (Noble et al., 2018; Nilsson et al., 2022), most farms only monitor a single horizontal point, leaving lateral spatial variability unknown (Burt et al., 2012; Solstorm et al., 2018; Burke et al., 2021). Similarly, as most time series are shorter than the average production period, temporal changes throughout the production cycle are often missed, preventing thorough time series analyses (Jeong et al., 2024). The presence of missing values, erroneous readings, and the inability to validate anomalies furthers the uncertainty of the monitoring system, and the loggers often fail to capture real environmental deviations or separate these from their potential causes. Because true anomalies contain valuable information about the environment, certainty regarding these values is particularly important. Although minor gaps or measurement errors can be treated through imputation techniques (Lien et al., 2023), longer periods cannot.

In essence, these errors contribute to a lack of knowledge of the conditions in the sea cage environment and reduce the ability to adapt production practices to actual environmental conditions (Berntsson et al., 2025; Zhang and Su, 2025). As the suboptimal DO levels negatively affect fish appetite, stress, and welfare, if environmental inferences based on poor-quality data are incorrect, they may lead to increased feed waste, increased stress from treatments (lice treatments), or missed opportunities to improve fish health and productivity (Forsberg, 1995; Remen et al., 2012; Remen et al., 2013; Remen et al., 2014; Remen et al., 2016; Berntsson et al., 2025). On the other hand, misjudging the presence of suboptimal conditions could result in unwarranted fasting or delays in production activities, driving up operational costs, if mitigation strategies are initiated inappropriately.

5.1.3.2 Lack of structure and availability increases the resources required and the ability to process the data

Beyond measurement quality, the lack of data structure and standardisation impedes timely access to relevant insights. This affects the DQ dimensions of timeliness, accessibility, and consistency. A common consequence of poor formatting and inconsistent structure is a substantial increase in the resources required to process the data (Haug et al., 2011). In many cases, processing may not even be possible without contextual metadata. For data to generate actionable value, it must be delivered in a usable format, on time, and with the necessary context (Fu et al., 2024). Currently, the monitoring systems fail to meet these expectations for most applications.

5.2 Recommendations to improve DQ in the monitoring system

Achieving adequate DQ across all application types requires targeted improvements. To maximise long-term value creation with regard to future applications, the following two strategies are recommended:

• Simplify data extraction and analysis for users, tailored to their specific needs.

• Improve the quality of time-series measurements collected by loggers, focusing on accuracy, completeness, and reliability.

Addressing these areas will enhance the understanding of environmental conditions within the sea cages and reduce barriers to effective data utilisation. While such improvements may require increased investment in data infrastructure and quality assurance, these costs must be weighed against the potential consequences of poor-quality data, such as fish stress, reduced growth, feed waste and other losses (Haug et al., 2011; Remen et al., 2012; Remen et al., 2013; Remen et al., 2014; Remen et al., 2016; Fu et al., 2024).

5.2.1 Metadata governance recommendations

Based on the assessment above, specifications for metadata governance can be established for the different applications to improve DQ dimensions such as consistency and accessibility, in accordance with the literature (Liaw et al., 2021; Malik and Khan, 2024). For the environmental monitoring system, standardising metadata specifications with consistent structure and format on the following topics is recommended:

• Specifying the applications of data and requirements for contextual information.

o Defining what the data will be used for and potential future applications.

o Defining the additional contextual information necessary to truly understand the sea cage environment helps guide the data collection process and increase DQ by design (Fu et al., 2024).

• Specifying what parameters are measured.

o For all recorded measurements, specifying the parameters measured in a manner that is easy to extract enables efficient categorisation of the data variables and simplifies subsequent data cleaning and analyses.

• Specifying the logger position and farm configuration.

o Including the horizontal and vertical position on the farm, and whether it is inside or near a cage, on the fleet or at a reference point up- or downstream from the farm, improves the understanding of what type of data the user is handling. Additionally, if the farm configuration is easily available, the end-user can also investigate potential wake-effects and biases from the farm in the data (Alver et al., 2024).

o As recommended by the ASC (Moldal et al., 2023), recording the coordinates specific to the unique logger’s position enables the combination of additional data sources, such as meteorological or oceanographic data.

• Establish an event log.

o Log production event type, specifying if a production activity occurred, what activity was conducted and what part of the farm was affected. This logging helps distinguish between various production activities, with its potential effects on the variation in the environmental parameters, and maintenance of loggers or logger errors.

  • Log the type and period of the event, as well as where it was conducted on the farm.

  • If the logger is moved from its position, log the period where the logger is absent from its position and where it has been kept meanwhile, for instance, if it is moved up to land or another position on the farm.

o Log maintenance of the environmental loggers helps the end-user distinguish between variations in the parameters, logger errors and maintenance events, and understand if the quality is improved after the maintenance.

  • Log calibration and cleaning occurrences.

  • Log the period the logger is absent from the site.

This improvement in metadata governance enables the user to easily extract contextual information, process and analyse the data more effectively. It also helps improve the accuracy and completeness of deviations and the ability to distinguish between logger errors and maintenance or production activities. By evaluating what the end-user needs to effectively extract and understand the information, in addition to developing technical definitions of DQ in the environmental monitoring system, DQ can be improved by design (Fu et al., 2024).

5.2.2 Improving the score of DQ dimensions and metrics

For the system to create value for all applications and increase knowledge about the sea cage environment, further actions beyond improving metadata governance are required to capture a complete dataset with accurate measurements. Actions to improve these are suggested below.

5.2.2.1 Completeness

Including the relevant parameters and spatio-temporal saturation in measurement will improve completeness, along with actions to reduce missing values.

• Monitor the parameters that are required for the different applications: direct measurements of DO variation and influential parameters, including sea temperature, current and salinity. These measurements support advanced models (Alver et al., 2024), welfare standards (Noble et al., 2018; Nilsson et al., 2022; Moldal et al., 2023) and detection of patterns that recurringly cause suboptimal DO (Ahmed and Shakoor, 2025; Berntsson et al., 2025).

• Completeness in spatial resolution requires vertical profiles and lateral measurements (Burt et al., 2012; Solstorm et al., 2018; Burke et al., 2021). Multiple horizontal loggers for capturing lateral variation must be active within the same period. For maximum value-creation across applications, placing loggers at the positions that satisfy most requirements, in addition to capturing lateral variation, is likely most effective: modelling the farm requires upstream measurements where they are most unlikely affected by the farm (Alver et al., 2024), while animal welfare assessments require measurements inside the sea cages where lowest DO can be expected (Noble et al., 2018; Nilsson et al., 2022; Moldal et al., 2023).

• Ensure data is collected for the relevant periods, throughout production periods and across multiple production periods, to enable identification of recurring trends and suitable mitigation strategies of harmful variation [as reviewed by 16, 26].

• The sampling frequency is currently at a satisfactory level in most loggers and continuing with this, rather than an aggregated average, will likely capture fluctuations best (Page et al., 2005).

5.2.2.2 Accuracy

Measuring the environment with perfect precision in the large and dynamic environments of the sea cage farms is overall unlikely, but establishing plausibility thresholds and acceptable value ranges is useful to improve accuracy. Using automated warnings of anomalies and data measurement errors that require, for instance, manual controls and comments of causes in an event log in the data collection system can improve accuracy, as anomalies can indicate useful information about harmful variation or sensor issues. Techniques like sliding window analysis can help identify anomalies in real-time, flagging improbable values and prompting operational checks of the relevant loggers. Such automated validation techniques are recommended in environmental monitoring using WSNs (Dereszynski and Dietterich, 2011; de Souza et al., 2020; Selle et al., 2023) and can improve reliability. Utilising loggers with multiple parameters at multiple positions improves opportunities to verify measurements (Burke et al., 2021; Alver et al., 2024).

5.2.3 Practical considerations for implementation

While increased spatial and vertical sensor coverage can improve the resolution and reliability of environmental monitoring data (Burke et al., 2021), such configurations also introduce operational and economic trade-offs in commercial sea cage farming. These include higher equipment and maintenance costs, increased exposure to biofouling, and additional labour demands associated with calibration and sensor upkeep (de Souza et al., 2020). Dense sensor deployments may also interfere with routine production activities such as net cleaning, feeding operations, and farm logistics. As demonstrated in this paper, some farms operate with a high level of sensor instrumentation, deploying up to 14 loggers at one site, but maintaining high data quality becomes increasingly challenging with greater instrumentation. This highlights that additional sensor density must be carefully evaluated in relation to operational complexity and upkeep capacity. Consequently, optimisation of monitoring systems should consider both data quality gains and operational feasibility, rather than prioritising sensor density alone.

5.2.3.1 Implementation considerations for farmers and vendors

To translate the proposed DQ principles into operational practice, a simplified implementation framework can be outlined for farmers and system providers. This framework focuses on three key decision points in the monitoring workflow: (I) data collection, (II) data validation and structuring, and (III) operational use of processed data.

• When planning data collection processes, priority should be given to clearly defined monitoring objectives, ensuring that the spatiotemporal sampling design reflects the intended use of the data (operational management, welfare monitoring, modelling, validation). A recent study show that bias in modelling applications is minimised when on-farm loggers are placed at a fleet compared to a cage (Berntsson et al., 2026), although upstream buoys may be optimal (Alver et al., 2024), while welfare reporting require in-cage measurements (Noble et al., 2018). Considering what is operationally feasible to maintain, while collecting useful information towards the intended objective, may aid farmers in determining their best placements. This can reduce unnecessary instrumentation while maintaining necessary coverage. Developing tools that identify optimal logger placements for different applications, alongside complementary models that provide the required operational information (Alver et al., 2024; Berntsson et al., 2026), could help farmers obtain high-value data while minimising instrumentation and maintenance demands.

• To validate the data, basic automated quality control rules can be applied to incoming data streams. These may include plausibility checks for DO values relative to expected saturation ranges, as well as simple anomaly detection based on abrupt temporal changes that can trigger event log requirements and operational checks. Anomaly detection is usually provided by the vendors, including the platform used in this study. Rather than removing data, flagged observations should be retained with quality indicators to preserve traceability.

• At the operational stage, cleaned and contextualised data should be integrated into simple decision-support routines, such as threshold-based alerts for low DO conditions or dashboards that combine environmental data with relevant contextual metadata (operative recommendations, depth, location, and farm activity). This enables the monitoring system to support both short-term operational responses and longer-term management insights.

Together, these steps provide a pragmatic pathway for implementing data quality improvements within existing aquaculture monitoring infrastructures, balancing the need for improved data reliability with the operational constraints of commercial production systems.

5.3 Further research

5.3.1 Standardising the environmental monitoring for the industry

While efforts to improve DQ at individual farms and among technology vendors will improve the systems’ value for operations, without broader industry standardisation, the impact will be limited for scholars and administrators (Falconer et al., 2025). Both regulations and the relevant literature are relatively arbitrary in terms of what a systematic approach to environmental monitoring entails (Akvakulturdriftsforskriften, norwegian ministry of trade, industry and fisheries: § 22, 2018; Noble et al., 2018; Nilsson et al., 2022; Moldal et al., 2023), causing a variation in how it is conducted in practice, as evidenced by the results of this study. This inconsistency hinders comparability, reduces value for large-scale analyses, and ultimately limits large-scale studies and regulatory improvements. Furthermore, while this study focuses on Atlantic salmon farming particularly, other species are also farmed regionally and globally, with specific demands to their environment. The data quality framework could be transferred to other species and environmental parameters as well, but needs to be adapted to the specific dynamics and physiological requirements of those species (OWIs for rainbow trout (Noble et al., 2020)). Broader standardisation could also further enhance the potential value for larger-scale environmental studies, and the operational value of water quality and welfare analyses in aquaculture in general.

As farms tend to increase in size despite the higher risk of DO challenges in large cages (McIntosh et al., 2022; Alver et al., 2023), the importance of understanding environmental carrying capacity becomes even more critical (Aksnes et al., 2019; Quiñones et al., 2019). Developing monitoring standards in the regulatory requirements (Akvakulturdriftsforskriften, norwegian ministry of trade, industry and fisheries: § 22, 2018) that are suitable for both scientific analysis and operational decision-making is essential for sustainable industry growth. While site-specific factors influence DO variability (Johansson et al., 2007; Burt et al., 2012; Burke et al., 2021; Berntsson et al., 2025), consistent standards would allow for more meaningful interpretation across these contexts. Standardised monitoring could support both data-driven farm-level optimisation and broader research into animal welfare and environmental change, contributing to evidence-based regulatory frameworks (Moldal et al., 2023; Falconer et al., 2025).

5.3.2 Demand for increased knowledge on harmful DO variation in the sea cage environments

The literature has multiple definitions for hypoxia tolerance of Atlantic salmon (Farrell and Richards, 2009; Remen et al., 2013; Remen et al., 2016; Noble et al., 2018; Jeong et al., 2024). Most studies to date have relied on controlled environments, leaving a knowledge gap in understanding how DO fluctuations affect fish under commercial farming conditions, where multiple stressors interact (Afonso, 2020; Berntsson et al., 2025).

To maximise the value of environmental monitoring systems, better definitions of harmful exposure thresholds are needed. This includes determining:

• The spatial and temporal extent of suboptimal or harmful DO levels that impact fish welfare and growth. Although various experiments have observed effects of hypoxia on physiology and appetite for fluctuating and continuous exposure (Remen et al., 2012; Remen et al., 2013; Remen et al., 2016; Oldham et al., 2020; Bowden et al., 2022; Rojas et al., 2025), a greater understanding of what volume of the sea cage and the period of exposure cause harm to welfare and growth is necessary.

• How these thresholds vary with fish size and health status, as fish size and various diseases can impact the tolerance limits in different ways (Oldham et al., 2020; Bowden et al., 2022; Berntsson et al., 2025).

• The influence of environmental factors such as sea temperature on fish behaviour, whether the fish are likely to escape warmer, low-DO zones or stay in warmer layers despite low DO (Oppedal et al., 2011; Remen et al., 2012).

• How to align stressful activities (treatments, handling) to environmental constraints (Quiñones et al., 2019; Berntsson et al., 2025).

Further research should also evaluate optimal sensor placement to reduce measurement bias and improve site-wide DO representation (Noble et al., 2018; Nilsson et al., 2022; Moldal et al., 2023; Alver et al., 2024). Broader studies could combine environmental data with other relevant operational data, such as biomass, production activities, and costs, to better understand the drivers of DO variation. Investigating DQ at sites using different sensor technologies could also help validate and expand the findings of this study.

6 Conclusions

This study has explored the various aspects of DQ in the environmental monitoring data of over 100 sites from 18 companies operating in the Nordic region. Our findings document and support the anecdotal statements regarding the DQ issues described in (i) in the introduction of this paper, that the monitoring system fails to satisfy various dimensions of data quality and consequently does not build trust in the data for the end-users. To build trust in the data collected through the monitoring system in the sea cage environment and enable optimisations based on DO variation, the application of the data must be considered when installing the system. Both the system and routines to maintain it must be sufficiently robust to achieve and sustain high DQ in the dynamic sea cage environment, to create value for farmers, scholars and administrators of the industry.

References

• 1

  AbdullahA. F.ManH. C.MohammedA.MurniM. A. K.SuleimanNurshahidaU. Y. A. B. M. J. (2024). Charting the aquaculture internet of things impact: Key applications, challenges and future trend. Aquacult Rep.39, 102358. doi: 10.1016/j.aqrep.2024.102358

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  AdewuyiA. Y.AnyibamaB.AdebayoK. B.KalinziJ. M.AdeniiyiS. A.WadaI.et al. (2024). Precision agriculture: Leveraging data science for sustainable farming. Int. J. Sci. Res. Arch.12, 1122–1129. doi: 10.30574/ijsra.2024.12.2.1371

  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  AfonsoL. O. B. (2020). “ Chapter 5 – Identifying and managing maladaptive physiological responses to aquaculture stressors,” in Fish Physiology, vol. 38 . Eds. BenfeyT. J.FarrellA. P.BraunerC. J. ( Academic Press, Cambridge, MA), 163–191. doi: 10.1016/bs.fp.2020.10.002

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  AhmedN.ShakoorN. (2025). Advancing agriculture through IoT, Big Data, and AI: A review of smart technologies enabling sustainability. Smart Agric. Technol.10, 100848. doi: 10.1016/j.atech.2025.100848

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  AksnesD. L.AureJ.JohansenP. O.JohnsenG. H. (2019). Multi-decadal warming of Atlantic water and associated decline of dissolved oxygen in a deep fjord. Estuar. Coast. Shelf Sci.228, 106392. doi: 10.1016/j.ecss.2019.106392

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  AlverM. O.FøreM.AlfredsenJ. A. (2022). Predicting oxygen levels in Atlantic salmon (Salmo salar) sea cages. Aquaculture548, 737720. doi: 10.1016/j.aquaculture.2021.737720

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  AlverM. O.FøreM.AlfredsenJ. A. (2023). Effect of cage size on oxygen levels in Atlantic salmon sea cages: A model study. Aquaculture562, 738831. doi: 10.1016/j.aquaculture.2022.738831

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  AlverM. O.FøreM.UrkeH. A.AlfredsenJ. A. (2024). Mathematical modelling of dissolved oxygen levels in a multi-cage salmon farm. Aquaculture593, 741291. doi: 10.1016/j.aquaculture.2024.741291

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  AntonucciF.CostaC. (2020). Precision aquaculture: A short review on engineering innovations. Aquacult Int.28, 41–57. doi: 10.1007/s10499-019-00443-w

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  Aquaculture Stewardship Council (ASC) (2025). “ ASC farm standard, version 1.0,” ( Aquaculture Stewardship Council, Utrecht, Netherlands). Available online at: https://programme-centre.asc-aqua.org/farm-v1/farm-standard/ (Accessed June 02, 2025).

  • Google Scholar
  • View reference in article

• 11

  (2018). Akvakulturdriftsforskriften, norwegian ministry of trade, industry and fisheries: § 22. Available online at: https://lovdata.no/forskrift/2008-06-17-822/%C2%A722 (Acccessed August 8, 2023).

  • Google Scholar
  • View reference in article

• 12

  BayihA. Z.MoralesJ.AssabieY.de ByR. A. (2022). Utilization of internet of things and wireless sensor networks for sustainable smallholder aquaculture. Sensors22, 3273. doi: 10.3390/s22093273

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  BellJ. L.MandelR.BrainardA. S.AltschuldJ.WenningR. J. (2022). Environmental monitoring tools and strategies in salmon net-pen aquaculture. Integr. Environ. Assess Manage.18, 950–963. doi: 10.1002/ieam.4622

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  BerntssonE. V. C.AlverM. O.LilandK. H.StevikT. K. (2026). Modelling cage-level dissolved oxygen variation within salmon farms. Front. Aquacult5. doi: 10.3389/faquc.2026.1813350

  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  BerntssonE. V. C.StevikT. K.BergheimA.PerssonD.StormoenM.LilandK. H.et al. (2025). Managing the dissolved oxygen variation of open Atlantic salmon sea cages: a narrative review. Rev. Aquacult17, e12992. doi: 10.1111/raq.12992

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  BowdenA. J.AdamsM. B.rewarthaS. J.ElliottN. G.FrappellP. B.ClarkT. D.et al. (2022). Amoebic gill disease increases energy requirements and decreases hypoxia tolerance in Atlantic salmon (Salmo salar) smolts. Comp. Biochem. Physiol. Part A Mol. Integr. Physiol.265, 111128. doi: 10.1016/j.cbpa.2021.111128

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  BurkeM.GrantJ.FilgueiraR.StoneT. (2021). Oceanographic processes control dissolved oxygen variability at a commercial Atlantic salmon farm: Application of a real-time sensor network. Aquaculture533, 736143. doi: 10.1016/j.aquaculture.2020.736143

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  BurtK.HamouteneD.MabroukG.LangC.PuestowT.DroverD.et al. (2012). Environmental conditions and occurrence of hypoxia within production cages of Atlantic salmon on the south coast of Newfoundland. Aquacult Res.43, 607–620. doi: 10.1111/j.1365-2109.2011.02867.x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  CichyC.RassS. (2019). An overview of data quality frameworks. IEEE Access7, 24634–24648. doi: 10.1109/ACCESS.2019.2899751

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  DereszynskiE. W.DietterichT. G. (2011). Spatiotemporal models for data-anomaly detection in dynamic environmental monitoring campaigns. ACM Trans. Sens Netw.8, 1–36. doi: 10.1145/1993042.1993045

  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  deSouzaP. S. S.RubinF. P.HohembergerR. (2020). Detecting abnormal sensors via machine learning: An IoT farming WSN-based architecture case study. Measurement164, 108042. doi: 10.1016/j.measurement.2020.108042

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  DiazR.RosenbergR. (2011). Introduction to environmental and economic consequences of hypoxia. Int. J. Water Resour. Dev.27, 71–82. doi: 10.1080/07900627.2010.531379

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  EarleyS.HendersonD.Sebastian-ColemanL. (2017). “ The DAMA guide to the data management body of knowledge (DAMA-DMBOK),” ( Technics Publications LLC, Sedona, AZ, USA).

  • Google Scholar
  • View reference in article

• 24

  FalconerL.HalstensenS.RinøS. F.NobleC.DaleT. (2025). Marine aquaculture sites have huge potential as data providers for climate change assessments. Aquaculture595, 741519. doi: 10.1016/j.aquaculture.2024.741519

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  FarrellA. P.RichardsJ. G. (2009). “ Chapter 11: Defining hypoxia: an integrative synthesis of the responses of fish to hypoxia,” in Fish Physiology, vol. 27 . Eds. RichardsJ. G.FarrellA. P.BraunerC. J. ( Academic Press, London), 487–503. doi: 10.1016/S1546-5098(08)00011-3

  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  FøreM.AlverM. O.AlfredsenJ. A.RasheedA.HukkelåsT.BjellandH. V.et al. (2024). Digital twins in intensive aquaculture – challenges, opportunities and future prospects. Comput. Electron. Agric.218, 108676. doi: 10.1016/j.compag.2024.108676

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  FøreM.FrankK.NortonT.SvendsenE.AlfredsenJ. A.DempsterT.et al. (2018). Precision fish farming: A new framework to improve production in aquaculture. Biosyst. Eng73, 176–193. doi: 10.1016/j.biosystemseng.2017.10.014

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  ForsbergO. I. (1995). Oxygen consumption of post-smolt Atlantic salmon during crowding and handling stress. Aquacult Int.3, 55–59. doi: 10.1007/BF00240921

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  ForsbergO. I. (1997). The impact of varying feeding regimes on oxygen consumption and excretion of carbon dioxide and nitrogen in post-smolt Atlantic salmon Salmo salar L. Aquacult Res.28, 29–41. doi: 10.1046/j.1365-2109.1997.00826.x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  FriskM.HøylandM.ZhangL.VindasM. A.ØverliØ.JohansenI. B.et al. (2020). Intensive smolt production is associated with deviating cardiac morphology in Atlantic salmon (Salmo salar L.). Aquaculture529, 735615. doi: 10.1016/j.aquaculture.2020.735615

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 31

  FryF. E. (1971). The effect of environmental factors on the physiology of fish. Fish Physiol.6, 1–98. doi: 10.1016/S1546-5098(08)60146-6

  • CrossRef
  • Google Scholar
  • View reference in article

• 32

  FuQ.NicholsonG. L.EastonJ. M. (2024). Understanding data quality in a data-driven industry context: Insights from the fundamentals. J. Ind. Inf. Integr.42, 100729. doi: 10.1016/j.jii.2024.100729

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 33

  HansenT. (2019). “ Tema: Laks i oppdrett,” in Havforskningsinstittutet. Bergen, Norway. Available online at: https://www.hi.no/hi/temasider/arter/laks/laks-i-oppdrett (Accessed May 7, 2026).

  • Google Scholar
  • View reference in article

• 34

  HaugA.ZachariassenF.van LiempdD. (2011). The costs of poor data quality. J. Ind. Eng. Manage.4, 168–193. doi: 10.3926/jiem.2011.v4n2.p168-193

  • CrossRef
  • Google Scholar
  • View reference in article

• 35

  JeongJ.AwosileB.ThakurK. K.StryhnH.BoyceB.VanderstichelR.et al. (2024). Longitudinal dissolved oxygen patterns in Atlantic salmon aquaculture sites in British Columbia, Canada. Front. Mar. Sci.10, 1289375. doi: 10.3389/fmars.2023.1289375

  • CrossRef
  • Google Scholar
  • View reference in article

• 36

  JohanssonD.JuellJ. E.OppedalF.StiansenJ. E.RuohonenK. (2007). The influence of the pycnocline and cage resistance on current flow, oxygen flux and swimming behaviour of Atlantic salmon (Salmo salar L.) in production cages. Aquaculture265, 271–287. doi: 10.1016/j.aquaculture.2006.12.047

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 37

  JyotiS.JiaB.SaksidaS.StryhnH.PriceD.RevieC. W.et al. (2024). Spatiotemporal patterns of mortality events in farmed Atlantic salmon in British Columbia, Canada, using publicly available data. Sci. Rep.14, 32122. doi: 10.1038/s41598-024-83876-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 38

  KattenbornT.LeitloffJ.SchieferF.HinzS. (2021). Review on convolutional neural networks (CNN) in vegetation remote sensing. ISPRS J. Photogramm Remote Sens173, 24–49. doi: 10.1016/j.isprsjprs.2020.12.010

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 39

  KhongI.YusufN. A.NurimanA.YadilaA. B. (2023). Exploring the impact of data quality on decision-making processes in information intensive organizations. APTISI Trans. Manage.7, 253–260. doi: 10.33050/atm.v7i3.2138

  • CrossRef
  • Google Scholar
  • View reference in article

• 40

  KramerD. L. (1987). Dissolved oxygen and fish behavior. Environ. Biol. Fish18, 81–92. doi: 10.1007/BF00002597

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 41

  LiawS. T.GuoJ. G. N.AnsariS.JonnagaddalaJ.GodinhoM. A.BorelliA. J.et al. (2021). Quality assessment of real-world data repositories across the data life cycle: A literature review. J. Am. Med. Inf. Assoc.28, 1591–1599. doi: 10.1093/jamia/ocaa340

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 42

  LienP. L.DoT. T.NguyenT. (2023). “ Data imputation for multivariate time-series data”, in: 15th International Conference on Knowledge and Systems Engineering (KSE) (Hanoi, Vietnam: IEEE), 1–6. doi: 10.1109/KSE59128.2023.10299484

  • CrossRef
  • Google Scholar
  • View reference in article

• 43

  LighternessA.AdcockM.ScanlonL. A.PriceG. (2024). Data quality-driven improvement in health care: Systematic literature review. J. Med. Internet Res.26, e57615. doi: 10.2196/57615

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 44

  MalikA.KhanS. (2024). Master data management: Building a foundation for data-driven decision-making. MZ Comp J.5.

  • Google Scholar
  • View reference in article

• 45

  McIntoshP.BarrettL. T.Warren-MyersF.CoatesA.MacaulayG.SzeteyA.et al. (2022). Supersizing salmon farms in the coastal zone: a global analysis of changes in farm technology and location from 2005 to 2020. Aquaculture553, 738046. doi: 10.1016/j.aquaculture.2022.738046

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 46

  MoldalT.Wiik-NielsenJ.OliveiraV. H. S.SvendsenJ. C.SommersetI. (2025). “ Norwegian fish health report 2024,” in Report 1a/2025 ( Norwegian Veterinary Institute, Ås, Norway). Available online at: https://www.vetinst.no/rapporter-og-publikasjoner/rapporter/2025/norwegian-fish-health-report-2024 (Accessed May 5, 2026).

  • Google Scholar
  • View reference in article

• 47

  MorewoodJ. (2023). Building energy performance monitoring through the lens of data quality: A review. Energy Build279, 112701. doi: 10.1016/j.enbuild.2022.112701

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 48

  NeisB.GaoW.CavalliL.ThorvalsenT.HolmenI. M.JeebhayM. F.et al. (2023). Mass mortality events in marine salmon aquaculture and their influence on occupational health and safety hazards and risk of injury. Aquaculture566, 739225. doi: 10.1016/j.aquaculture.2022.739225

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 49

  NilssonJ.GismervikK.NielsenK. V.IversenM. H.NobleC.FrotjoldH.et al. (2022). Laksvel. Standardisert operasjonell velferdsovervåking for laks i matfiskanlegg. Rep. Norwegian Ins Mar. Res. (Bergen, Norway), 2022–2014.

  • Google Scholar
  • View reference in article

• 50

  NobleC.GismervikK.IversenM. H.KolarevicJ.NilssonJ.StienL. H.et al. (2018). “ Welfare indicators for farmed Atlantic salmon: tools for assessing fish welfare,” ( Nofima, Bodø, Norway), 351.

  • Google Scholar
  • View reference in article

• 51

  NobleC.GismervikK.IversenM. H.KolarevicJ.NilssonJ.StienL. H.et al. (2020). Welfare Indicators for Farmed Rainbow Trout: Tools for Assessing Fish Welfare (Tromsø, Norway: Nofima).

  • Google Scholar
  • View reference in article

• 52

  OldhamT.DempsterT.CrosbieP.AdamsM.NowakB. (2020). Cyclic hypoxia exposure accelerates the progression of amoebic gill disease. Pathogens9, 597. doi: 10.3390/pathogens9080597

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 53

  OliveiraV. H. S.DeanK. R.QvillerL.KirkebyC.JensenB. B. (2021). Factors associated with baseline mortality in Norwegian Atlantic salmon farming. Sci. Rep.11, 14702. doi: 10.1038/s41598-021-93874-6

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 54

  OppedalF.DempsterT.StienL. (2011). Environmental drivers of Atlantic salmon behaviour in sea-cages: A review. Aquaculture311, 1–18. doi: 10.1016/j.aquaculture.2010.11.020

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 55

  ØstevikL.StormoenM.EvensenØ.XuC.LieK. I.NødtvedtA.et al. (2022). Effects of thermal and mechanical delousing on gill health of farmed Atlantic salmon (Salmo salar l.). Aquaculture552, 738019. doi: 10.1016/j.aquaculture.2022.738019

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 56

  PageF. H.LosierR.McCurdyP.GreenbergD.ChaffeyJ.ChangB.et al. (2005). “ Dissolved oxygen and salmon cage culture in the southwestern New Brunswick portion of the Bay of Fundy,” in Environmental Effects of Marine Finfish Aquaculture. Ed. HargraveB. T. ( Springer-Verlag, Berlin/Heidelberg), 1–28. doi: 10.1007/b136002

  • CrossRef
  • Google Scholar
  • View reference in article

• 57

  PansaraR. (2021). Master data governance best practices. Int. J. Comput. Sci. Mobile Comp10, 1–3. doi: 10.47760/ijcsmc.2021.v10i11.001

  • CrossRef
  • Google Scholar
  • View reference in article

• 58

  PerssonD.NødtvedtA.AunsmoA.StormoenM. (2022). Analysing mortality patterns in salmon farming using daily cage registrations. J. Fish Dis.45, 335–347. doi: 10.1111/jfd.13560

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 59

  PitcherG. C.Aguirre-VelardeA.BreitburgD.CardichJ.CarstensenJ.ConleyD. J.et al. (2021). System controls of coastal and open ocean oxygen depletion. Prog. Oceanogr197, 102613. doi: 10.1016/j.pocean.2021.102613

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 60

  PratiwyM.CahyaM. D.AndrianiY. (2022). Digitization of aquaculture: A review. Int. J. Fish Aquat. Stud.10, 18–22. doi: 10.22271/fish.2022.v10.i1a.2623

  • CrossRef
  • Google Scholar
  • View reference in article

• 61

  QuiñonesR. A.FuentesM.MontesR. M.SotoD.León-MuñozJ. (2019). Environmental issues in Chilean salmon farming: A review. Rev. Aquacult11, 375–402. doi: 10.1111/raq.12337

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 62

  RemenM.AasT. S.VågsethT.TorgersenT.OlsenR. E.ImslandA.et al. (2014). Production performance of Atlantic salmon (Salmo salar l.) postsmolts in cyclic hypoxia, and following compensatory growth. Aquacult Res.45, 1355–1366. doi: 10.1111/are.12082

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 63

  RemenM.OppedalF.ImslandA. K.OlsenR. E.TorgersenT. (2013). Hypoxia tolerance thresholds for post-smolt Atlantic salmon: dependency of temperature and hypoxia acclimation. Aquaculture416–417, 41–47. doi: 10.1016/j.aquaculture.2013.08.024

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 64

  RemenM.OppedalF.TorgersenT.ImslandA. K.OlsenR. E. (2012). Effects of cyclic environmental hypoxia on physiology and feed intake of post-smolt Atlantic salmon: initial responses and acclimation. Aquaculture326-329, 148–155. doi: 10.1016/j.aquaculture.2011.11.036

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 65

  RemenM.SieversM.TorgersenT.OppedalF. (2016). The oxygen threshold for maximal feed intake of Atlantic salmon post-smolts is highly temperature dependent. Aquaculture464, 582–592. doi: 10.1016/j.aquaculture.2016.07.037

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 66

  RojasI.de MelloM. M. M.ZanuzzoF. S.SandrelliR. M.PeroniE. F. C.HallJ. R.et al. (2025). Chronic hypoxia has differential effects on constitutive and antigen-stimulated immune function in Atlantic salmon (Salmo salar). Front. Immunol.16, 1545754. doi: 10.3389/fimmu.2025.1545754

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 67

  SelleM.SpiessF.VisscherC.RautenschleinS.JungA.AuerbachM.et al. (2023). Real-time monitoring of animals and environment in broiler precision farming – How robust is the data quality? Sustainability15, 15527. doi: 10.3390/su152115527

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 68

  SolstormD.OldhamT.SolstormF.KlebertP.StienL. H.VågsethT.et al. (2018). Dissolved oxygen variability in a commercial sea-cage exposes farmed Atlantic salmon to growth limiting conditions. Aquaculture486, 122–129. doi: 10.1016/j.aquaculture.2017.12.008

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 69

  SuX.SutarlieL.LohX. J. (2020). Sensors, biosensors and analytical technologies for aquaculture water quality. Research9, 1–15. doi: 10.34133/2020/8272705

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 70

  SyedA. B.HasanR.ChowdhuryN. I.RahmanH.AhmedI. (2025). A systematic review of time series algorithms and analytics in predictive maintenance. Decision Analytics J.15, 100573. doi: 10.1016/j.dajour.2025.100573

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 71

  WeiskopfN. G.WengC. (2012). Methods and dimensions of electronic health record data quality assessment: enabling reuse for clinical research. J. Am. Med. Inf. Assoc.20, 144–151. doi: 10.1136/amiajnl-2011-000681

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 72

  ZahirM.SuY.ShahzadM. I.AyubG.RahmanS. U.IjazJ.et al. (2024). A review on monitoring, forecasting and early warning of harmful algal bloom. Aquaculture593, 741351. doi: 10.1016/j.aquaculture.2024.741351

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 73

  ZhangQ.SuB. (2025). A hybrid approach towards real-time monitoring of fish distributions in aquaculture net cage. Aquacult Eng110, 102527. doi: 10.1016/j.aquaeng.2025.102527

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article