Integrating AIoT Technologies in Aquaculture: A Systematic Review

Abstract

The increasing global demand for seafood underscores the necessity for sustainable aquaculture practices. However, several challenges, including rising operational costs, variable environmental conditions, and the threat of disease outbreaks, impede progress in this field. This review explores the transformative role of the Artificial Intelligence of Things (AIoT) in mitigating these challenges. We analyse current research on AIoT applications in aquaculture, with a strong emphasis on the use of IoT sensors for real-time data collection and AI algorithms for effective data analysis. Our focus areas include monitoring water quality, implementing smart feeding strategies, detecting diseases, analysing fish behaviour, and employing automated counting techniques. Nevertheless, several research gaps remain, particularly regarding the integration of AI in broodstock management, the development of multimodal AI systems, and challenges regarding model generalization. Future advancements in AIoT should prioritise real-time adaptability, cost-effectiveness, and sustainability while emphasizing the importance of multimodal systems, advanced biosensing capabilities, and digital twin technologies. In conclusion, while AIoT presents substantial opportunities for enhancing aquaculture practices, successful implementation will depend on overcoming challenges related to scalability, cost, and technical expertise, improving models’ adaptability, and ensuring environmental sustainability.

1. Introduction

By 2050, it is expected that there will be nearly 10 billion people on the planet, putting tremendous strain on food systems to supply the growing demand for diets high in protein [1]. When addressing global food security, aquatic foods are an essential source of nutrition, especially in areas where they are the primary source of protein. The fastest-growing industry in food production, aquaculture has become a vital tool for balancing aquatic foods’ supply-and-demand gap, contributing over 50% of the world’s fish for human consumption [1]. However, the rapid expansion of aquaculture has also brought challenges, including rising production costs, environmental degradation, and resource inefficiencies, which threaten its long-term sustainability [2].

The rising costs of aquaculture production, driven by factors such as feed expenses, disease outbreaks, and labour shortages, have further exacerbated the industry’s challenges [3,4]. Traditional practices often rely on manual monitoring and decision-making, which are prone to inefficiencies and errors, leading to suboptimal yields and increased waste. In this context, there is an urgent need for innovative, sustainable solutions that can enhance productivity while minimising environmental and economic costs. AI, or artificial intelligence, has become a game-changing technology that can revolutionise aquaculture by enabling precision farming, optimising resource use, and improving decision-making processes [5]. By integrating AI technologies with the Internet of Things (IoT), the aquaculture industry can address critical challenges, reduce production costs, and sustainably contribute to global food security [6,7,8].

Aquaculture, a vital contributor to global food systems, encompasses diverse species tailored to meet varying market demands and ecological roles. Finfish, spearheaded by commercially significant species such as catfish, seabass, trout, carp, salmon, and tilapia, represent the most prominent group because of their high market value and ability to adapt to new aquatic environments [9,10]. Crustacean aquaculture, particularly shrimp farming, has emerged as a cornerstone of international trade, especially in regions with high consumer demand [11,12]. Molluscs, such as mussels, scallops, and oysters, are frequently farmed in marine habitats because of their nutritional and commercial value. Interestingly, the production of pearls from specific oyster species is one example of how aquaculture intersects with high-value luxury markets. Coral culturing is becoming increasingly important in decorative trade and ecological restoration initiatives, especially when restoring damaged coral reefs [13,14]. Additional species diversifying the aquaculture landscape include aquatic macroinvertebrates, bioindicators in environmental health assessment [15,16]. Phytoplankton, fundamental to aquatic food webs, are also cultured for their ecological and nutritional value [17,18]. Aquatic plants, such as algae and seaweed, expand aquaculture’s range, with applications ranging from direct human consumption to producing biofuels and cosmetics [1,19]. This wide range of farmed aquatic species highlights aquaculture’s exceptional versatility and crucial role in addressing several complex global issues, including environmental sustainability, food security, and economic development.

The integration of artificial intelligence (AI) and the Internet of Things (IoT) is transforming industries by enhancing sustainability, efficiency, and innovation [20,21,22]. AI enables the analysis of large, complex datasets, identifies patterns, and automates decision-making using deep learning (DL), machine learning (ML), computer vision (CV), and natural language processing (NLP). In aquaculture, these AI-driven solutions facilitate real-time monitoring, optimise operations, and contribute to sustainable practices [23,24,25]. IoT devices such as cameras, sensors, and automated monitoring systems continuously collect data on key environmental parameters, fish behaviour, and feeding patterns, which AI algorithms process to generate actionable insights [26,27,28]. These advancements reduce reliance on manual intervention, enhance decision-making, and improve resource efficiency [29,30].

The IoT framework in aquaculture consists of interconnected sensors, communication gateways, and cloud computing platforms, enabling real-time data collection, analysis, and transmission [31]. This infrastructure monitors critical water quality parameters, including temperature, pH, dissolved oxygen, and salinity, ensuring optimal conditions for aquatic organisms. Cloud computing enhances AI-driven solutions by providing scalable processing capabilities to handle large volumes of data. AIoT applications in aquaculture include automated breeding, disease prediction, biomass estimation, water quality management, intelligent feeding systems, and behaviour analysis [32,33]. AI-powered feeding systems utilise sensor data and underwater imaging to optimise feeding schedules, improving fish growth rates, reducing environmental impact, and minimizing feed waste [34,35,36].

A fundamental aspect of AIoT-driven aquaculture is the management of water quality. IoT sensors continuously monitor crucial water parameters such as temperature, pH, dissolved oxygen, and ammonia, while AI-based models analyse trends and suggest proactive interventions [37]. For example, deep learning models like long short-term memory (LSTM) networks process time-series data to predict oxygen depletion, allowing timely aeration adjustments [38,39]. Similarly, reinforcement learning (RL) algorithms optimise water management strategies, balancing environmental stability and energy efficiency. AIoT is also key in early disease detection, as outbreaks can cause significant financial losses. Deep learning models such as U-Net and YOLO analyse high-resolution images from IoT cameras, detecting lesions and abnormal swimming patterns indicative of disease. Additionally, biosensors integrated with AI models analyse biological markers, enabling early disease detection and mitigation before outbreaks escalate [40,41].

Beyond health monitoring, AIoT systems contribute to species identification, population tracking, and selective breeding optimization. Convolutional neural networks (CNNs), including VGG and ResNet, facilitate species segmentation and organism counts, ensuring compliance with regulations and biodiversity preservation [42,43]. AI-driven selective breeding programs leverage genetic and environmental data to predict desirable traits such as disease resistance and growth efficiency, improving sustainability and productivity. Robotics and automation are redefining aquaculture operations. Autonomous underwater vehicles (AUVs) are equipped with sensors and cameras that perform feed distribution, fish inspections, and tank cleaning with minimal human supervision. These systems are enhanced by reinforcement learning algorithms, enabling adaptive decision-making based on real-time environmental feedback.

Despite recent review papers on aquaculture [31,44,45,46,47,48], this review contributes a comprehensive synthesis of current AIoT applications in aquaculture, providing a structured overview of the diverse technologies and methodologies employed and systematically examining advancements in areas such as water quality management, automated counting, smart feeding, and disease detection. This paper identifies critical research gaps and outlines future directions for AIoT development. Specifically, it highlights the need for integrated multimodal systems, enhanced predictive analytics, and tailored solutions for crustacean aquaculture, offering a valuable resource for researchers, aquaculture experts, and practitioners seeking to advance sustainable and efficient aquaculture practices through AIoT integration.

Following this section, which provides a comprehensive overview of the role of AIoT in transforming aquaculture practices and highlights the challenges and opportunities in this field, Section 2 details systematic review approaches to gathering and synthesising relevant literature. Section 3 outlines the fundamental components of AIoT systems in aquaculture, focusing on IoT-enabled sensing devices and AI-driven analytics. Section 4 presents a detailed review of AIoT applications in aquaculture, covering various aspects such as water quality management, smart feeding, disease detection, and automated monitoring. Section 5 identifies the research gaps in current AIoT applications, emphasizing areas that require further investigation. Section 6 discusses the challenges to AIoT implementation in aquaculture, including initial costs, environmental variability, data privacy, and scalability. Section 7 explores future directions in AIoT for aquaculture, focusing on potential advancements and innovations. Finally, Section 8 concludes the review, summarizing the key findings and underscoring the potential of AIoT to drive a more efficient, sustainable, and resilient aquaculture sector.

2. Systematic Review Approaches

A structured and reproducible methodology was employed to conduct this systematic review, adhering to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines [49]. Literature searches were conducted across three major databases: Web of Science, Scopus, and PubMed. These were selected due to their comprehensive coverage of research in aquaculture, artificial intelligence (AI), and the Internet of Things (IoT).

The search was performed using a combination of keywords and Boolean operators to capture all relevant literature. The main search terms included:

• ‘Artificial intelligence’ OR ‘AI’ OR ‘machine learning’ OR ‘deep learning’ OR ‘Internet of Things’ OR ‘IoT’ AND ‘aquaculture’.
• ‘Aquaculture’ OR ‘fish farming’ OR ‘shrimp farming’ OR ‘aquatic farming’ OR ‘smart aquaculture’ AND ‘monitoring’ OR ‘automation’ OR ‘disease detection’ OR ‘water quality’ OR ‘feeding optimization’.

These searches retrieved both general and specific research articles related to AIoT in aquaculture. The literature was limited to English-language publications from 2015 to 2024. As shown in Figure 1, the screening process involved the following steps.

I.

Identification: An initial search across three databases—PubMed (320), Scopus (192), and Web of Science (129)—yielded a total of 641 records after duplicate removal.

II.

Screening: Titles and abstracts of all 641 articles were reviewed for relevance. At this stage, 346 articles were excluded based on predefined exclusion criteria such as domain mismatch, language restrictions, or lack of AI/IoT relevance in aquaculture.

III.

Eligibility Assessment: The remaining 295 articles were reviewed in full to assess their alignment with the study’s objectives and inclusion criteria. During this phase, 135 full-text articles were excluded due to methodological limitations, lack of practical implementation, or insufficient detail.

IV.

Inclusion: Finally, 150+ full-text articles were selected for detailed analysis in the systematic review, representing the most relevant and high-quality studies addressing AIoT applications in aquaculture.

2.1. Inclusion and Exclusion Criteria

To maintain a focused and high-quality review, studies explicitly discussed the application of AI, ML, or IoT in aquaculture, particularly research focusing on AI-based automation, disease detection, or smart monitoring in fish and shrimp farming. Studies also presented experimental results, simulations, or case studies relevant to AI applications in aquaculture or integrated AI with predictive modelling, real-time monitoring, or decision support systems for aquaculture management. Conversely, studies that focused on AI, ML, or IoT in non-aquaculture domains (e.g., agriculture, healthcare, robotics), articles mentioning aquaculture without significant AI, ML, or IoT components, non-English papers or conference abstracts without accessible full-text versions, and studies lacking experimental validation, practical implementation, or meaningful conclusions relevant to AI applications in aquaculture were excluded.

2.2. Data Extraction and Synthesis

A rigorous data extraction protocol was established to synthesise each selected study’s findings systematically. This comprehensive process involved meticulously recording essential bibliographic information: the study title, author(s), and publication year. The research objective and methodology were thoroughly documented to understand each investigation’s scope and intent. A key component of the extraction process was the detailed identification of the specific type of AI or IoT technology employed, allowing for a nuanced analysis of the technological landscape. Crucially, the precise aquaculture application targeted by each study, encompassing areas such as disease detection, water quality monitoring, and feeding optimisation, was carefully extracted to provide a clear overview of the practical applications of AI and IoT within the aquaculture sector.

This systematic review consolidates the understanding of AIoT applications in aquaculture, identifying current trends, technological gaps, and future research directions.

3. Components of AIoT in Aquaculture

The successful implementation of AIoT in aquaculture hinges on two fundamental components: IoT-enabled sensing devices for continuous data acquisition and AI-driven analytics for intelligent decision-making. IoT sensors are crucial for monitoring biological and environmental factors, including fish behaviour, health metrics, and water quality indicators such as pH, temperature, and dissolved oxygen. These sensors are integrated with advanced communication technologies like LoRa, NB-IoT, and 5G to facilitate seamless data transmission, enabling efficient and reliable connectivity between remote aquaculture sites and centralised management systems. This interconnected framework (Figure 2) allows real-time monitoring, predictive analysis, and automation, optimising resource utilisation and improving farm productivity.

3.1. IoT Sensors

IoT sensors are critical in modernising aquaculture by enabling real-time monitoring of critical environmental and biological factors, ensuring optimal conditions for aquatic life [29]. For instance, continuous tracking of dissolved oxygen levels helps mitigate risks associated with sudden drops, allowing farmers to activate aeration systems promptly and prevent hypoxic conditions that could harm stock [50]. Similarly, temperature sensors detect abrupt fluctuations, helping farmers maintain stable water conditions to minimise fish stress and reduce the risk of thermal shock [51]. Beyond environmental monitoring, IoT sensors also serve as early-warning systems, identifying potential threats such as disease outbreaks, inefficient feeding patterns, and water contamination, thereby enabling proactive interventions to enhance aquaculture sustainability and productivity [52].

3.1.1. Water Quality Monitoring Sensors

Water quality monitoring sensors are fundamental to ensuring aquaculture systems’ productivity and overall health, since they unremittingly track key parameters such as temperature, pH, salinity, dissolved oxygen, and ammonia. Studies [7,38,53] have observed the significant effect of these variables on the growth and physiological health of aquatic species, as even slight changes can cause stress, increase disease susceptibility, and in severe cases lead to mortality. Recent advancements in sensor technology have focused on improving precision, durability, and cost-effectiveness to meet the stringent monitoring requirements of aquaculture environments.

For pH monitoring, ion-sensitive field-effect transistor (ISFET) and glass electrode sensors are widely used because of their high sensitivity to changes in water acidity—an essential factor in fish health [54]. While glass electrodes are commonly employed, they require frequent calibration and are susceptible to biofouling, limiting their long-term reliability in aquaculture applications [55]. On the other hand, ISFET sensors have improved durability and increased resistance to drift, which makes them better suited for prolonged use in aquatic environments. One study [55] demonstrated their effectiveness in long-term aquaculture surveillance, reinforcing their potential for improved water quality management.

To maintain optimal water conditions, temperature sensors such as resistance temperature detectors (RTDs) and thermocouples are crucial, as fluctuations in temperature directly influence fish growth and metabolic rates. Similarly, dissolved oxygen (DO) sensors prevent hypoxic conditions, which can severely affect fish health. Traditional electrochemical DO sensors, including Clark electrodes, work via amperometric methods, but need regular preservation to prevent membrane fouling, a common problem in organically rich waters [37]. Optical dissolved oxygen (DO) sensors provide a more accurate and non-invasive solution by utilising the principle of fluorescence quenching caused by molecular oxygen [56]. However, as noted in [52], while optical sensors provide stable and accurate readings, they are more expensive and may degrade when exposed to specific waterborne chemicals. A comparative study in [57] revealed that even though conductive sensors provide reliable salinity measurements, inductive sensors are more resilient for long-term use in aquaculture systems, where biofouling poses a persistent problem.

Ammonia, a toxic metabolic by-product, must be carefully monitored to prevent harmful accumulations in aquaculture environments. The two primary methods for ammonia detection are ion-selective electrodes (ISEs) and colorimetric sensors. Though ISEs provide continuous and precise readings, they require frequent recalibration, since they are susceptible to sensor drift in fluctuating water conditions [58]. Colorimetric techniques are effective for ammonia detection, but have slower response times, making them less appropriate for real-time applications in high-density aquaculture settings [59].

Although sensor technologies are essential for managing water quality, several challenges persist, including biofouling, sensor drift, and significant maintenance requirements, all of which can impact data accuracy and the sensors’ durability [60]. To overcome these challenges, recent research has emphasised the creation of anti-biofouling coatings, the improvement of sensor materials, and the incorporation of multi-sensor systems. These advancements aim to improve data accuracy, minimise maintenance needs, and facilitate rapid detection of water quality changes.

3.1.2. Optical Sensors in Aquaculture

Optical sensors play an important role in monitoring water quality in aquaculture, providing real-time information on turbidity and water condition, two key indicators of aquatic systems’ environmental health. Turbidity, which results from suspended particles such as organic matter, sediments, and plankton, can negatively impact fish health by reducing light penetration, disrupting oxygen levels, and altering overall ecological situations [52]. Elevated turbidity is associated with phytoplankton blooms or pollutants, which are responsible for increasing fish stress and susceptibility to disease [55]. Optical turbidity sensors analyse the interaction of light with suspended particles, typically measuring either light scattering or absorption to assess water clarity.

Beyond turbidity monitoring, optical systems, particularly cameras integrated with computer vision algorithms, have gained prominence in aquaculture for behavioural analysis. These systems enable continuous observation of fish’s spatial distribution and movement inside the tanks, offering valuable data for assessing stress levels and early disease detection [61]. Research indicates that erratic movements, changes in swimming speed, or fish turning frequency are early indicators of environmental stresses or health problems, allowing for timely intervention [62]. Furthermore, computer vision-based monitoring has been instrumental in optimising feeding practices by correlating fish behaviour with feeding responses, thereby minimising feed waste and promoting sustainability [13].

Recent advancements, such as self-cleaning optical surfaces and hydrophobic coatings, have been developed to mitigate biofouling and extend sensor lifespan in long-term aquaculture applications [52]. Integrating optical sensors with other monitoring systems, such as dissolved oxygen and pH sensors, provides a more comprehensive water quality assessment. This multi-sensor approach enhances data reliability and enables more effective environmental control in aquaculture [52]. Regardless of their advantages, optical sensors face challenges like biofouling, which may restrict optical pathways and degrade measurement accuracy over time [52]. Nevertheless, their ability to deliver precise, low-energy measurements makes them a cost-effective and practical solution for water quality management in aquaculture. By detecting slight variations in water clarity, these sensors facilitate proactive environmental monitoring, helping prevent adverse effects on fish health and ensuring the sustainability of aquaculture operations. Emerging technologies, including AI-driven image analysis and hyperspectral imaging, further enhance optical sensors’ capabilities, allowing for more precise and automated aquaculture monitoring [61].

3.1.3. Motion Sensors in Aquaculture

Accelerometers and other motion sensors are crucial for evaluating fish feeding habits, movement patterns, and stress reactions. These devices offer a non-intrusive way to monitor fish behaviour continuously, aiding in effectively managing health and productivity. Typically integrated into tags or attached externally, accelerometers record detailed movement data, including burst swimming, tail-beat frequency, and resting periods. Such information offers insight into feeding efficiency, activity levels, and metabolic rates.

Research has demonstrated that accelerometer-based monitoring can detect subtle behavioural changes that may not be easily observed, such as heightened activity caused by deteriorating water quality or reduced movement linked to health issues [63]. Moreover, integrating accelerometers with telemetry systems enables remote, real-time behavioural monitoring, allowing for rapid interventions when abnormalities are detected. This advancement enhances fish welfare management by facilitating timely responses to environmental or physiological stressors.

3.1.4. Deployment Strategies

Optimising the deployment of IoT sensors in aquatic systems is critical to ensuring accurate data collection, reliable monitoring, and seamless integration into various operational environments. One key consideration is selecting the most suitable wireless communication protocol: different technologies offer unique advantages based on specific aquaculture needs.

LoRa (long range) technology is extensively used in large-scale and remote aquaculture systems due to its energy efficiency and long-distance communication capabilities, which are particularly well suited for monitoring expansive or hard-to-reach locations [52]. LoRa’s ability to transmit data over long distances with minimal power consumption makes it ideal for applications where sensors are dispersed across a wide area and frequent battery changes are impractical. Aquaculture can include monitoring water quality parameters (e.g., temperature, dissolved oxygen) in large offshore fish farms or inland ponds located in remote regions. Also, 5G is increasingly adopted in settings with high data transfer rates and low latency, particularly for use cases such as real-time video-based behaviour analysis [63]. The high bandwidth and low latency of 5G enable the transmission of large volumes of data, such as high-resolution video and images, which are essential for advanced monitoring and analysis. In aquaculture, 5G can support applications like real-time monitoring of fish behaviour using underwater cameras, facilitating early detection of diseases or stress, and enabling automated feeding systems that respond to fish activity. Another effective alternative is narrowband IoT (NB-IoT), especially when long-range connectivity and power efficiency are more important than high-bandwidth transmission. It is especially effective for low-frequency sensor data, such as pH and dissolved oxygen measurements, which do not require the high-speed capabilities of 5G [64]. NB-IoT is designed for applications that require infrequent data transmission and long battery life. In aquaculture, this makes it suitable for deploying sensors that periodically measure and transmit basic water quality parameters, such as temperature, pH, dissolved oxygen, and salinity, from various locations within a farm.

While LoRa, 5G, and NB-IoT represent key wireless technologies in aquaculture, it is important to acknowledge that other protocols may also play a role, albeit often in more specific or limited scenarios. For instance, Zigbee is a short-range, low-power wireless communication technology suitable for communication within a contained tank system or a relatively small, closely monitored aquaculture facility. Zigbee excels in creating mesh networks, allowing devices to relay data to each other, which can be useful in environments where direct communication with a central hub is challenging. However, its limited range restricts its applicability in large-scale or open-water aquaculture operations. Wi-Fi is another widely used technology, offering higher data rates than Zigbee. It can be effective for aquaculture systems close to a base station or facility with internet access. Wi-Fi can support applications requiring more bandwidth, such as transmitting data from high-resolution sensors or enabling local monitoring via smartphones or tablets. However, its range and power consumption can be limiting factors, particularly in remote or off-grid aquaculture settings.

Selecting the most appropriate wireless communication protocol is a critical decision that depends on several interrelated factors. These include the following.

I.

Distance: The physical extent of the aquaculture operation significantly influences the choice. Long-range technologies like LoRa or NB-IoT are necessary for expansive farms, while short-range technologies may suffice for smaller, contained systems.

II.

Data rate: The volume and frequency of transmitting data are crucial considerations. High-bandwidth applications, such as real-time video monitoring, demand technologies like 5G, whereas LoRa or NB-IoT can adequately serve low-data-rate applications.

III.

Power consumption: The power needs of the sensors and the availability of power sources (e.g., grid power, batteries, solar) are essential factors. Low-power technologies are vital for remote deployments where frequent battery replacements are impractical.

IV.

Cost: Implementing and maintaining the communication infrastructure is a significant consideration, especially for small-scale aquaculture operations.

V.

Interference: The potential for interference from other wireless devices or environmental factors (e.g., water absorption) can affect communication reliability.

VI.

Environmental conditions: The harshness of the aquaculture environment, including humidity, salinity, and temperature variations, can influence the suitability of different wireless technologies.

Ultimately, the choice of protocol necessitates carefully evaluating the aquaculture operation’s specific requirements and constraints to balance performance, cost-effectiveness, and reliability. The comparisons in

Table 1. Comparative table of wireless communication protocols.

Feature	LoRa	NB-IoT	5G	Zigbee	Wi-Fi
Range	Long (Several km)	Long (Several km)	Short to Medium (Up to several km)	Short (10–100 m)	Short to Medium (Up to 100 m)
Data Rate	Low (0.3–50 kbps)	Low (20–250 kbps)	High (Gbps)	Low to Medium (20–250 kbps)	High (Up to Gbps)
Power Consumption	Very Low	Very Low	High	Low	Medium to High
Mobility	High	High	High	Low	High
Complexity	Low	Low	High	Medium	High
Cost	Low to Medium	Low to Medium	High	Low	Medium
Typical Use Cases in Aquaculture	Remote water quality monitoring, large-scale farms	Periodic data from distributed sensors	Real-time video monitoring, automated feeding	Localised sensor networks within tanks	Local monitoring, data transfer near base station
Strengths	Long range, low power, cost-effective	Long range, low power, good coverage	High data rate, low latency, high bandwidth	Low power, mesh networking	High data rate, widely available
Limitations	Low data rate	Low data rate	High power consumption, limited range	Short range	Power consumption, limited range

, summarizing the key features (range, data rate, power consumption, etc.) of different wireless protocols, should significantly enhance the reader’s understanding and aid in protocol selection.

In addition to selecting the appropriate communication technology, sensor placement is crucial for comprehensive environmental assessment [64]. The effectiveness of data collected by IoT sensors is directly influenced by where they are positioned within the aquaculture system. Both vertical and horizontal placement of sensors within cages, tanks, or ponds is significant. Variations in temperature, dissolved oxygen, and salinity can occur at different depths and locations. Sensors must be placed at strategic points to capture a comprehensive understanding of these variations. For instance, placing sensors at multiple depths can effectively monitor temperature stratification, where water layers exhibit distinct temperature differences. Similarly, positioning sensors near feeding zones can provide valuable insights into how feeding activities influence water quality parameters.

Additionally, deploying sensors at inlet and outlet points can help track the efficiency of water exchange processes within the system. However, sensor placement in aquaculture environments presents unique challenges, particularly in dynamic settings. Tidal influences and water currents can affect sensor stability and positioning, potentially leading to inaccurate or unreliable data. To mitigate these challenges, various solutions can be employed. Anchoring sensors securely can prevent them from drifting due to currents, while floating platforms can maintain sensors at specific depths in systems with fluctuating water levels.

Power supply considerations are critical for deploying IoT sensors in aquaculture, especially in remote locations with limited access to consistent power sources. Traditional power methods may not be feasible in such scenarios, necessitating exploring alternative energy solutions. Batteries provide a portable power source, but their lifespan is limited, requiring periodic replacement. Solar power offers a renewable and sustainable alternative, harnessing energy from sunlight to power sensors. Other energy harvesting techniques, such as capturing energy from water flow or wave motion, are also being investigated to provide continuous and self-sufficient power for IoT devices in aquaculture. Power efficiency strategies are essential to prolong sensor lifespan and optimise power usage. Duty cycling involves putting sensors into low-power or sleep modes when not actively measuring, thereby conserving energy. Low-power modes reduce the sensor’s energy consumption during operation, extending the time between battery changes or reducing reliance on alternative power sources.

The process of transmitting collected sensor data is a crucial component of IoT systems in aquaculture. Data can be transmitted through various methods, including direct transmission to a cloud platform or local gateway. Cloud platforms offer scalable and accessible storage solutions, enabling data to be accessed and analysed anywhere with an internet connection. Local servers provide an alternative storage option, offering more control over data and potentially reducing latency in data access. The choice between cloud storage and local servers depends on data volume, accessibility requirements, security considerations, and available infrastructure. Both transmission and storage methods have significant implications for data accessibility and analysis, influencing the efficiency and effectiveness of AIoT applications in aquaculture.

Regular sensor maintenance and calibration are indispensable for ensuring data accuracy and the long-term reliability of IoT sensors in aquaculture systems. Aquatic environments pose several challenges to sensor integrity. Biofouling, accumulating microorganisms, algae, and other aquatic organisms on sensor surfaces, can impede sensor function and lead to inaccurate readings. Corrosion, caused by prolonged exposure to water and dissolved substances, can also damage sensor components. Additionally, physical damage from impacts or abrasion can compromise sensor performance. To address these challenges, regular maintenance schedules and best practices are essential. This includes regular cleaning of sensors to remove biofouling, calibrating sensors to maintain accuracy, and conducting periodic inspections to identify and address any signs of corrosion or physical damage. Adhering to these maintenance protocols ensures that sensors provide reliable data, enabling informed decision-making and effective management of aquaculture operations.

3.2. AI Algorithms in Aquaculture

AI algorithms are essential in converting raw data gathered from IoT sensors into actionable insights, facilitating more efficient and data-informed decision-making in aquaculture. AI improves critical farm management tasks through predictive modelling, automation, and process optimisation, including feeding strategies, water quality control, and overall operational efficiency.

The combination of computer vision, fuzzy logic, image processing, machine learning, and deep learning (DL) in aquaculture facilitates the analysis of vast and complicated datasets, each offering distinct advantages. ML and DL enable pattern recognition and predictive analytics, allowing for early detection of anomalies in fish behaviour or environmental conditions. Computer vision and image processing support real-time monitoring of fish health, movement, and feeding activity, reducing reliance on manual observation. Fuzzy logic provides a robust framework for handling uncertainties in aquaculture environments, improving decision-making processes in scenarios with variable or imprecise data. These AI-driven approaches enhance precision, efficiency, and sustainability in modern aquaculture management.

3.2.1. Machine Learning (ML) Approaches in Aquaculture

In predictive analytics and pattern recognition, machine learning (ML) [65,66] is essential, especially when processing structured data gathered from Internet of Things (IoT) sensors in aquaculture. ML algorithms are classified into three main types—supervised learning, unsupervised learning, and reinforcement learning—each providing distinct functionalities for analysing intricate datasets and enhancing farm management strategies [67].

Supervised Learning for Classification and Regression

In supervised learning [68,69,70], models are trained using labelled datasets to perform tasks such as regression and classification. Classification models are useful for organising data into predefined categories, supporting decision-making processes in areas like water quality evaluation and disease identification. Techniques such as logistic regression (LR), k-nearest neighbour (KNN), decision trees (DTs), random forests (RFs), and support vector machines (SVMs) are commonly employed to classify water conditions as either ‘safe’ or ‘unsafe’, facilitating prompt actions to address potential hazards. Likewise, classification methods are extensively applied in disease detection, where they analyse factors like fish behaviour or image data to enable early diagnosis and effective disease management. The performance of these models is assessed using metrics like accuracy, precision, and recall, which gauge their ability to differentiate between various conditions accurately.

In contrast, regression models [71,72] predict continuous values, making them suitable for forecasting environmental conditions such as water temperature or fish growth rates. Techniques like linear regression and neural networks analyse historical trends to provide predictive insights into key aquaculture parameters. Performance metrics (e.g., root mean square error (RMSE) and mean absolute error (MAE)) are commonly used to assess the accuracy of these models in predicting continuous variables like oxygen levels or pH fluctuations.

Unsupervised Learning for Pattern Discovery

Unsupervised learning [73,74,75] methods identify patterns and relationships in datasets by analysing unlabelled data. Clustering algorithms, such as hierarchical clustering, density-based spatial clustering of applications with noise (DBSCAN), and k-means are widely used in aquaculture to detect behavioural anomalies, segment fish populations based on feeding behaviour, and monitor changes in water quality. For instance, clustering techniques can identify groups of fish showing unusual movement patterns, potentially signalling stress, disease, or environmental disturbances. The effectiveness of clustering models is often measured using metrics like the Jaccard index, silhouette coefficient, Davies–Bouldin index (DBI), and normalised mutual information (NMI).

Reinforcement Learning for Adaptive Decision-Making

Reinforcement learning (RL) [76] is especially well suited for dynamic and adaptive decision-making processes in aquaculture. RL models operate through a trial-and-error learning mechanism, where an agent interacts with an environment and learns an optimal strategy based on rewards and penalties. This approach is highly applicable in optimising feeding schedules, aeration control, and water quality management [77].

Q-learning, a value-based RL algorithm, enables agents to determine optimal policies by estimating cumulative rewards for different actions. Q-learning can be applied in aquaculture to optimise aeration systems by adjusting oxygen levels based on real-time water quality data. By combining deep learning and Q-learning, Deep Q-Networks (DQNs) [78] extend these capabilities to complicated, high-dimensional environments. For instance, a DQN-based system can analyse real-time sensor and video feed data to model collective fish behaviour, improving monitoring and management decisions.

Policy gradient methods [79,80] offer another RL approach by directly optimising an agent’s decision-making policy, making them suitable for tasks requiring continuous adjustments, such as fine-tuning feeding routines. Value-based learning and policy gradient techniques are combined in actor–critic frameworks to improve aquaculture RL applications by balancing short-term operational efficiency with long-term sustainability. In such frameworks, the actor selects actions, while the critic evaluates their effectiveness, guiding improvements over time.

Furthermore, RL can be extended to multi-agent systems where several agents cooperate or compete within shared environments. In aquaculture, where a multi-tank setup is used, RL agents can coordinate feeding and aeration across multiple tanks to optimise resource allocation and ensure consistent fish growth. Advanced RL techniques such as proximal policy optimisation (PPO) and soft actor–critic (SAC) address stability and efficiency challenges, making them highly promising for real-world aquaculture applications.

3.2.2. Deep Learning (DL)

Deep learning (DL) [81,82,83], a branch of machine learning (ML), employs artificial neural networks (ANNs) with multiple layers to identify significant patterns from large and intricate datasets. Inspired by the neural architecture of the human brain, deep learning models are particularly effective for tasks such as classification, segmentation, and prediction in aquaculture [4]. These models can process structured and unstructured data, making them highly versatile for various applications [84].

At the heart of deep learning lies the concept of artificial neural networks (ANNs), which serve as the foundation for various architectures designed to handle specific data types and tasks. The most basic form of neural network is the feedforward neural network (FNN) [85], where data flow in a single direction—from the input layer, through hidden layers, to the output layer. FNNs are primarily utilised for tasks such as classification and regression, mainly when dealing with structured numerical data. In aquaculture, for instance, FNNs can process historical data to forecast essential water quality parameters like temperature and pH levels, enabling proactive farm management. Convolutional neural networks (CNNs) [86,87] are commonly used for tasks involving unstructured data, such as images. CNNs leverage convolutional layers to extract spatial features from images, making them highly effective for object detection, image classification, and segmentation applications. For instance, CNN architectures such as residual network (ResNet) and visual geometry group network (VGG) have demonstrated significant success in aquaculture. VGG is well known for its deep, yet uniform architecture and is commonly used to identify different fish species and detect abnormalities in aquatic environments. ResNet, which incorporates skip connections to address the vanishing gradient problem, allows for the development of deeper networks. This capability makes it particularly suitable for monitoring fish movements and assessing their health condition through high-resolution imagery [88,89,90].

Beyond traditional CNN models, more specialised deep learning architectures have been developed to address complex tasks in aquaculture. U-Net [91], a fully convolutional network, is particularly effective in photograph segmentation tasks, such as separating fish outlines from underwater video footage to facilitate automated fish population monitoring [29]. Another widely used model, You Only Look Once (YOLO), is an algorithm for detecting real-time objects that excel in identifying fish behaviour patterns and external threats, such as predatory species in aquaculture systems [34]. These advanced architectures enable real-time analysis of high-resolution images and videos, making them necessary for decision-making in aquaculture operations. In addition to image-based tasks, deep learning models are optimised for analysing temporal and sequential data. Long Short-Term Memory (LSTM) networks [92,93], an advanced form of Recurrent Neural Networks (RNNs), are especially well suited for time-series analysis. LSTMs are designed to handle long-term dependencies, making them highly effective in predicting water quality parameters and fish-feeding patterns by analysing historical data. For example, an LSTM model can analyse real-time sensor inputs to forecast dissolved oxygen levels for the upcoming 24 h, enabling farmers to optimise their aeration systems in advance.

Autoencoders [94,95] are widely used for feature extraction and dimensionality reduction for unsupervised learning tasks. In aquaculture, these models can analyse sensor data to detect anomalies in water quality, such as irregular pH levels or fluctuations in dissolved oxygen, which may indicate malfunctions of equipment or disturbances in the environment. Autoencoders perform by compressing large datasets into low-dimensional representations and then rebuilding them, enabling them to identify subtle deviations that might go unnoticed in traditional monitoring systems. A different category of deep learning models, known as multilayer perceptrons (MLPs), is composed of fully connected layers and is especially effective for tasks involving numerical prediction and classification. To categorise water quality conditions, MLPs can evaluate various environmental factors, including salinity, temperature, and dissolved oxygen levels. This information is crucial for enhancing fish health and improving the productivity of aquaculture operations [96,97].

Deep Learning offers various architectures that cater to aquaculture’s diverse needs, from image-based applications to sequential data analysis. Its ability to process complex datasets with high accuracy and efficiency makes DL an essential tool in modern aquaculture management, contributing to improved monitoring, predictive analytics, and sustainable farming practices.

3.2.3. Advanced Techniques in Visual Data Processing

Artificial intelligence (AI) is effectively integrated into aquaculture through computer vision and image processing, which facilitate interpreting and analysing visual data collected from video feeds. These technologies facilitate the assessment of fish behaviour, the detection of abnormal patterns, and the identification of signs of stress or disease. Advanced techniques such as edge detection [98], image segmentation [99], and object tracking allow researchers and farmers to extract meaningful insights from video footage, including estimation of fish biomass and monitoring the movements of each fish. Studies such as those conducted by Huang et al. [2] have demonstrated how computer vision enhances understanding fish health and behaviour, providing a foundation for developing responsive and data-driven aquaculture management strategies.

3.2.4. Application of Fuzzy Logic in Aquaculture

Fuzzy logic is a computational method that emulates human reasoning, making it particularly useful for managing uncertainty and imprecision in complex and dynamic systems like aquaculture. In contrast to conventional binary logic, which operates on strict true or false conditions, fuzzy logic enables more flexible decision-making by incorporating varying degrees of truth. This flexibility makes it particularly useful for optimising aquaculture operations. For example, a study in [29] combines U-Net with a neuro-fuzzy system optimised through particle swarm optimisation (PSO) to accurately detect fish diseases in underwater photos, facilitating timely intervention and disease management. Similarly, research in [53] employs fuzzy inference systems in combination with FFAUNet segmentation to optimise feeding operations in aquaculture, minimising waste while increasing fish growth. A study referenced in [54] employs a fuzzy neural network integrated into an expert system to identify fish diseases. This approach combines real-time data with expert insights to facilitate early disease detection and improve treatment planning. Such implementations demonstrate the versatility and efficacy of fuzzy logic in boosting operational efficiency, sustainability, and productivity within the aquaculture industry.

4. AIoT Applications in Aquaculture

Incorporating Artificial Intelligence of Things (AIoT) into aquaculture is revolutionising conventional farm management practices by merging real-time data from IoT sensors with sophisticated AI-powered analytics. This collaboration (Figure 3) improves sustainability, operational efficiency, and productivity in various aquaculture areas. AIoT applications range from monitoring water quality and smart feeding to detecting fish disease and autonomous robotics. Each of these applications contributes to optimising resource utilisation, improving yield prediction, and ensuring the well-being of aquatic organisms.

4.1. Smart Feeding Systems

Smart feeding, which utilises AIoT applications to track fish feeding activity and improve feeding schedules, is one of the most significant AIoT applications in aquaculture. Techniques such as deep learning models (e.g., ResNet [100]) and gradient boosting machines (GBM) [101] analyse real-time data from computer vision and acoustic sensors to assess fish appetite appropriately and adjust the amount of food accordingly. These systems reduce food waste and improve feed conversion ratios (FCRs) as well as the growth rates of fish. However, their effectiveness varies across species and environmental conditions, requiring further refinement. Research by [102] employs bioenergetic and protein flux models to optimise gilthead seabream’s feeding schedules, demonstrating that the fish’s weight increased by 46.85% by adjusting feeding frequency and composition. Additionally, real-time acoustic monitoring, as explored in [103], detects variations in feeding intensity by analysing sound patterns, offering a non-intrusive way to assess hunger levels.

4.2. Water Quality Management

Water quality management is crucial in sustainable aquaculture, directly influencing aquatic species’ health, growth, and welfare. Deviations in key parameters such as temperature, pH, dissolved oxygen (DO), ammonia, and turbidity can lead to significant health issues, impacting growth rates and increasing susceptibility to diseases [104]. Integrating IoT sensors and AI-based predictive analytics has become crucial for real-time monitoring and early intervention in modern aquaculture. Several studies have shown that proactive monitoring systems using deep learning techniques like convolutional neural networks (CNNs) and recurrent neural networks (RNNs) can accurately predict critical hydrological conditions, such as degradation of water quality and oxygen depletion [105]. AI-based approaches, including hybrid CNN-LSTM-GRU models and TD-LSTM models, have been utilised for accurate temperature predictions, improving monitoring across various aquatic environments [44].

pH levels are another crucial factor in aquaculture, influencing biochemical processes, metabolic rates, and fish health [7,8]. AI models such as random forests and simple recurrent units (SRUs) have been applied to predict pH fluctuations, ensuring optimal conditions and preventing stress-induced health issues [106]. Similarly, salinity control is vital for maintaining osmotic balance and physiological stability in aquatic species. Advances in IoT-enabled salinity sensors have enhanced measurement accuracy, with novel transducer designs and conductivity meters providing real-time monitoring capabilities [56,107]. Effective management of DO is also critical, as hypoxia can severely impact aquatic organisms. Predictive models integrating LightGBM, BiSRU, and attention mechanisms have improved DO forecasting, allowing for timely adjustments in high-density aquaculture systems [108]. Other approaches [109], such as genetic algorithms combined with XGBoost and CatBoost, have demonstrated strong predictive performance, though model recalibration is necessary for different environmental conditions [37].

Turbidity affects water clarity, oxygen distribution, and fish-feeding behaviour, necessitating efficient monitoring systems. Neural networks embedded in IoT sensor nodes have been used to classify pollutants. At the same time, decision tree and support vector machine (SVM)-based models have been applied for fault diagnosis in turbidity data [110,111]. Chlorophyll-a levels indicate phytoplankton biomass, with excessive growth leading to harmful algal blooms. AI techniques such as CNNs and LSTMs have been employed for chlorophyll monitoring, supporting early intervention strategies to prevent ecological imbalances [112,113]. Similarly, high concentrations of ammonia, a toxic byproduct of fish metabolism, can be hazardous. Ammonia levels must be regularly monitored to avoid harmful accumulations, as their toxicity increases with higher pH. AI-driven predictive models, such as general regression neural networks (GRNN) and LSTMs, have been used to predict ammonia fluctuations in high-density shrimp aquaculture [58]. Advanced approaches like XAdaBoost combined with LSTMs have also achieved lower error rates than the conventional method [18].

4.3. Disease Detection and Classification

The health and productivity of fish and shrimp populations in aquaculture rely heavily on effective disease detection and classification. Disease outbreaks threaten aquatic species and cause significant economic losses. Integrating artificial intelligence (AI), sensor technologies, and advanced data analysis techniques has led to significant progress in early disease detection and management. These innovations enhance the accuracy of disease identification, reduce reliance on manual inspections, and allow for real-time monitoring in aquaculture systems.

Computer vision techniques have been widely used to identify visual symptoms of diseases in fish and shrimp. Machine learning algorithms, especially convolutional neural networks (CNNs), have accurately detected infections such as White Spot Syndrome Virus (WSSV). For example, an artificial neural network (ANN) trained with Canny edge detection and grey-level co-occurrence matrix (GLCM) features demonstrated a classification accuracy of 94.71% for distinguishing infected from healthy shrimp [12,67,114]. Additionally, support vector machines (SVMs) and random forest classifiers, combined with K-means clustering for segmentation, have successfully classified diseases like Epizootic Ulcerative Syndrome (EUS) and Tail and Fin Rot, with reported accuracies exceeding 88% [114]. These approaches highlight the potential of AI-driven models in disease diagnostics. However, challenges such as variations in image quality due to environmental conditions and the need for disease-specific models still require further research.

Deep learning architectures incorporating attention mechanisms have improved disease classification by focusing on disease-specific features. ResNet50 combined with the convolutional block attention module (CBAM), has improved classification accuracy, achieving 89.9% in detecting fish diseases [115]. Similarly, models combining CNNs with online sequential extreme learning machines (OSELM) have improved feature extraction, demonstrating 94.28% accuracy in identifying infections even under variable water conditions [116]. Transfer learning approaches using pre-trained deep learning models, such as InceptionV3 and VGG16, have also been successfully applied to fish disease classification, with accuracy levels above 91% [117].

Water quality monitoring plays a crucial role in predicting disease risks in aquaculture. Deviations in parameters such as pH, dissolved oxygen, and ammonia concentrations often indicate stress conditions that increase susceptibility to infections. Gradient boosting models (GBMs) have been employed to predict disease outbreaks based on water quality variations, achieving an accuracy of 92% [118,119]. Additionally, integrating water quality sensors with deep learning models, such as MobileNetV2, has enabled real-time disease prediction, with reported accuracy rates of 91.5% [119]. These models provide valuable insights into environmental factors contributing to disease outbreaks, allowing farmers to take preventive measures before infections spread.

Cross-modal and zero-shot learning techniques have emerged as innovative approaches for disease detection, particularly when labelled image datasets are limited. These methods utilise textual descriptions from the scientific literature to map disease characteristics onto image features, facilitating the recognition of previously unseen diseases [120]. Similarly, transfer learning models trained on diverse datasets have been adapted for fish disease detection, demonstrating classification accuracies above 91% when using architectures such as VGG-16 and AlexNet [121]. These techniques help overcome data scarcity challenges and expand the applicability of AI-driven disease detection systems.

Ensemble and hybrid models have improved disease classification accuracy by integrating several machine-learning approaches. For instance, the performance metric-infused weighted ensemble (PMIWE) model, which integrates DenseNet-121, ResNet-50, and EfficientNetB3, has achieved a classification accuracy of 97.53% in identifying freshwater fish diseases [122]. Additionally, hybrid models combining CNNs with feature extraction techniques, such as histogram of oriented gradients (HOGs) and thresholding, have improved disease classification robustness [123].

Adaptive neural fuzzy systems have also been developed to enhance diagnostic precision in aquaculture. An improved U-Net segmentation model combined with multi-head attention has been proposed to precisely identify disease-affected areas in fish photos [59]. Similarly, fuzzy neural network-based expert systems have been implemented for fish health management, showing high accuracy in detecting common diseases in species like grass carp [124]. These intelligent systems contribute to automated disease detection and decision-making in aquaculture.

Mobile devices and IoT-based disease monitoring systems offer scalable solutions for real-time disease detection in aquaculture. Convolutional Neural Network (CNN) models, specifically optimised for mobile applications, have been created to deliver disease alerts and treatment suggestions via cloud platforms. This makes AI-powered disease diagnostics accessible to small-scale fish farmers [125]. Furthermore, disease detection models utilising YOLOv5, integrated with intelligent water quality monitoring systems, have achieved accuracy rates of over 97% [126].

Biosensor technology is crucial in pathogen detection, enabling rapid and precise identification of infectious agents. Gold electrode-based biosensors utilising electrochemical impedance spectroscopy (EIS) have been optimised for aquaculture applications. Findings suggest that low surface roughness in gold electrodes enhances impedance response quality, improving pathogen detection sensitivity [127]. Integrating biosensors with AI-driven analytical models offers new opportunities for real-time disease prevention and control in aquaculture environments.

These advancements highlight the potential of AI, computer vision, and biosensor technologies in transforming disease detection and management into aquaculture. Future studies should focus on improving the generalizability of AI models across different aquaculture species, enhancing feature extraction techniques, and integrating multimodal data sources to develop more comprehensive disease prediction systems.

4.4. Fish Behaviour Detection

The continuous monitoring of fish behaviour is essential for maintaining fish health, welfare, and overall aquaculture productivity. Behavioural changes in fish often serve as early signs of environmental stress, disease onset, and feeding responses, enabling timely interventions to prevent potential losses. Recent advancements in artificial intelligence, computer vision, acoustic monitoring, and IoT-based sensor networks have significantly improved the accuracy and reliability of behaviour analysis. These technologies facilitate real-time and non-invasive monitoring in complex aquaculture systems, reducing the need for manual observation while enhancing decision-making processes.

Deep learning models have been widely applied to detect abnormalities in fish behaviours, such as lethargy, rapid movement changes, and erratic swimming, which are often associated with a disease or suboptimal water conditions. A ResNeXt 3×1D convolutional network was implemented in one study, achieving 95.3% accuracy in identifying abnormal fish behaviours [128]. The model improves computational efficiency and monitoring effectiveness in large-scale aquaculture systems. Another study [61] used Faster R-CNN for object detection, dynamic time warping for behaviour analysis, and directed cycle graphs for posture classification, obtaining a 92.8% accuracy in detecting abnormal fish behaviours. These hybrid models provide a strong foundation for monitoring behaviour in real time, though they can be affected by external factors like lighting and water clarity. Additionally, a dual-flow deep network combined with a multi-instance learning framework has been employed to study swarm behaviour in densely populated fish groups, offering important insights into how these groups react to stress and disease [129].

Analysis of locomotion and swimming speed are also crucial indicators of fish health and metabolic activity. Acoustic telemetry has been employed to measure swimming speeds in marine cages, demonstrating a root mean square error of 7.85 cm/s across different velocity ranges. This non-invasive technique allows for continuous monitoring of fish movement; however, signal interference from environmental noise and reflections can affect accuracy [130]. In another approach, computer vision-based deformable models have been applied to analyse swimming patterns, capturing species-specific movement characteristics. Using convolutional neural networks (CNNs) to analyse fish locomotion has further improved classification accuracy, with studies reporting up to 96% accuracy in predicting activity patterns based on movement data [131].

Behavioural analysis has also proven valuable for early disease detection, as abnormal swimming patterns and reduced activity levels often precede visible symptoms. A transformer-based model, EchoBERT, has been introduced to analyse echograms for detecting early-stage pancreatic disease in Atlantic salmon, achieving a Matthews correlation coefficient of 0.694. This model demonstrated superior performance compared to traditional long short-term memory (LSTM)-based methods, allowing for disease detection up to one month earlier [132]. The potential for AI-driven behavioural monitoring to enhance disease management in aquaculture is significant, with ongoing research focusing on improving generalisation across different species and environments.

Shoaling and social behaviour modelling provide additional insights into fish group dynamics and welfare. CNN-based models have been developed to classify different shoaling states, such as feeding activity and stress responses, with an accuracy of 82.5% using spatiotemporal fusion images [133]. By integrating spatial and optical flow data, these models improve behavioural classification but may still be affected by variations in environmental lighting and water conditions. Reinforcement learning approaches, such as Deep Networks, have also been explored to simulate schooling behaviour, allowing for studying group interactions without predefined rule-based models. These simulations successfully reproduce ordered formations within fish populations and hold potential for real-world applications in optimising aquaculture stocking densities and welfare assessments [62].

Acoustic and echogram-based monitoring methods offer an alternative to vision-based systems, particularly in environments with poor visibility. EchoBERT, initially developed for disease detection, has also been applied for behaviour analysis, utilising echogram data to identify stress-induced behavioural changes [132]. Similarly, acoustic imaging has been used to monitor fish swarm activity, achieving high accuracy in detecting abnormal behaviours [129]. These methods provide a non-intrusive solution for continuous welfare monitoring, though further improvements in signal processing and environmental adaptability are needed to enhance their robustness across diverse aquaculture settings.

By integrating AI-driven models, acoustic sensors, and IoT-enabled monitoring systems, fish behaviour detection is becoming increasingly precise and scalable. Future advancements will likely focus on improving model interpretability, reducing computational complexity for edge computing applications, and enhancing system adaptability to varying aquaculture environments. The continuous refinement of these technologies will contribute to more efficient and sustainable fish farming practices, ensuring early intervention and improved stock management.

4.5. AI-Driven Approaches for Automated Counting in Aquaculture

Accurate counting of aquatic organisms is crucial for efficient stock management, health assessment, and optimising feeding practices. Conventional approaches, which typically involve manual counting or invasive sampling, are time-consuming and susceptible to inaccuracies, especially in high-density aquaculture environments. Recent progress in artificial intelligence, computer vision, and sensor technologies has facilitated the development of automated and precise counting techniques for various species, such as fish, shrimp, and holothurians. Numerous studies have highlighted the success of deep learning models in this field. For example, ShrimpseedNet, a smartphone-based model, has been created to count shrimp seeds, offering a convenient and effective tool for hatchery operators. A deep learning approach based on density estimation has been used to enhance the accuracy of shrimp larvae population assessments. In controlled environments, computer vision methods have been applied for fish counting, and echo-sounder-based algorithms have been developed to estimate fish populations within farming nets. Advances in technology include an enhanced YOLOv5 model for detecting underwater fish and multi-scale context-aware convolutional neural networks (CNNs) for estimating fish density. Additionally, attention-based deep learning techniques have been implemented to count fry in high-density cultures, and non-contact systems have been created to evaluate holothurian populations in aquaculture settings.

Counting shrimp larvae and seeds is crucial in hatcheries to ensure accurate stock estimation and maintain optimal stocking densities. AI-based models such as ShrimpCountNet, which uses density estimation, have achieved over 98% accuracy in counting shrimp larvae. Despite its high accuracy, the model faces challenges when applied to highly dense populations, where overlapping individuals can hinder precise detection. Similarly, ShrimpSeedNet has been optimised for smartphone deployment, making it a practical tool for aquaculture workers. However, lighting conditions and image quality variations can impact its performance, highlighting the need for further refinement. Future research aims to improve model robustness under diverse environmental conditions and extend applications to additional aquatic species.

Fish counting in controlled and natural environments remains complex due to occlusion, movement, and varying environmental factors. Advanced deep learning models such as YOLOv5 have been employed to detect and count fish while analysing their movement for stress detection. A multi-scale context-enhanced CNN has been developed to improve fish density estimation, addressing occlusion and perspective distortion challenges. This model has demonstrated superior accuracy in crowded environments, though its performance decreases in low-density populations. An alternative approach involves using a super-resolution GAN-based network for fry counting in high-density settings, achieving a 97.57% accuracy rate. While these models show promise, limitations remain in adapting to real-time video processing and integrating multimodal data sources for enhanced precision. Research has also investigated the use of deep learning in marine ecology and fisheries management, such as employing YOLOv2 for detecting scallops in low-contrast underwater imagery and utilising generative models to estimate jellyfish swarm populations. Nonetheless, these methods encounter difficulties adapting to varied environmental conditions, highlighting the need for advancements in data augmentation and model adaptation strategies.

Echo-sounder technology provides a non-invasive method for assessing fish populations, especially in aquaculture nets where visual enumeration is difficult. Using acoustic signal processing techniques, fish populations can be estimated with an accuracy margin of under 10%, effectively accounting for disruptions caused by net structures. However, accuracy declines in sparse populations due to reduced signal reflections, emphasising the need for advanced signal processing techniques to enhance detection capabilities in low-density settings. Expanding these methods to various aquaculture environments could improve their reliability and scalability for large-scale fish farming operations.

Non-contact counting methods have also been applied to holothurian (sea cucumber) populations to minimise disturbances in aquaculture settings. AI-powered detection systems incorporating YOLOv3 and Faster R-CNN, combined with tracking algorithms, have achieved robust counting results even in turbid water. Nevertheless, underwater visibility and occlusion pose significant challenges, affecting model accuracy. Researchers are working to enhance these models by expanding underwater datasets and integrating more sophisticated tracking algorithms to improve counting performance across diverse conditions.

Automated counting of phenotypic features, such as lateral line scales, is also gaining attention in fisheries management and breeding programs. AI-driven models like TRH-YOLOv5, which incorporate transformer mechanisms and small-target detection modules, have been developed to improve precision in scale counting, achieving high accuracy levels. These methods support phenotypic analysis for selective breeding and species identification. However, environmental factors such as underwater lighting variations and motion blur remain obstacles. Future improvements will focus on refining image preprocessing techniques and extending automated phenotypic feature detection to a broader range of aquaculture species.

AI-driven solutions are revolutionising organism (highlighted in

Table 2. Applications of AI and machine learning in aquaculture: techniques, data sources, and methods.

Aquaculture Domain	Specific Application	AI/Machine Learning Techniques	Data Sources and Methods	References
Fish health and behaviour	Disease detection and monitoring	Computer vision, support vector machines (SVMs), convolutional neural networks (CNNs), attention mechanisms, online sequential extreme learning machine (OSELM), knowledge distillation, YOLOv7	Imaging systems, sonar, video, annotated datasets, feature extraction, model training, and classification	[67,134]
	Feeding behaviour and growth analysis	Gaussian mixture models (GMMs), k-nearest neighbour (KNN), CNNs	Sonar imaging, length and weight estimation, distribution analysis	[135]
	Stress state recognition	Knowledge distillation, GhostNet, ResNeXt101	Video analysis, lightweight network training, transfer learning	[136]
	Mortality detection	Computer vision, machine learning, sensor integration, YOLOv7	Image data, annotated datasets, and real-time monitoring	[137]
	Behaviour analysis (general)	Acoustic technology, computer vision, deep learning, FlowNet2, 3D CNN	Real-time monitoring systems, RGB and optical flow data, and data fusion	[138,139]
Aquaculture monitoring (general)	Zooplankton monitoring	Optical imaging, computer vision, SVM	Imaging systems, feature extraction, classification	[140]
	Aquaculture site extent and location	Machine learning (random forest, SVM), GIS, CNN	Satellite imagery, geospatial data, model training, visualisation	[141,142]
	Seaweed growth and health	Image processing, computer vision, DeepLabV3+	Image/video capture, annotation, segmentation, hyperparameter tuning	[143]
	Underwater species monitoring	Computer vision, YOLOv5, transfer learning	Underwater imagery, model training, feature extraction	[144]
	Phytoplankton detection	CNN, Fast R-CNN	Microscopic images, dataset curation, image analysis, classification	[145]
	Shrimp monitoring	Machine learning, ultrasonic imaging, YOLOv4	Ultrasound images, image augmentation, model training	[146]
	Water turbidity prediction	Random forest	Remote sensing data, modelling	[147]
	Vessel activity and infrastructure monitoring	Satellite imagery, remote sensing, convolutional networks (ConvNet)	Multi-band satellite images, model training	[148]
	Red tide detection	Computer vision, deep learning, U-Net	High-resolution satellite imagery analysis	[149]
	Aquaculture site surveillance	Computer vision, deep learning (Mask R-CNN)	Drone-captured visual data, cloud-based processing, semantic segmentation	[135,150]
	Aquaculture raft monitoring	Multi-scaled attention U-Net, dilated convolution, offset convolution	SAR image segmentation, deep neural network modelling	[151]
	Water quality and turbidity forecasting	Dynamic network surgery–deep neural networks (DNS–DNNs)	UAV multispectral imagery, ground sensor data, spectral analysis, and model training	[152]
	Marine aquaculture characterization	Incremental double unsupervised deep learning (IDUDL)	SAR image segmentation, unsupervised Learning	[75]
	Mussel float monitoring	Computer vision, multi-object tracking	Image processing, machine learning, tracking pipeline	[153]
Aquaculture management and production	Feeding management and status	AlexNet, ResNet34, convolutional block attention module (CBAM), YOLOv4-Tiny-ECA	Underwater imagery, feature representation, object detection	[100,154]
	Water quality identification	Computer vision	Water colour analysis	[155,156]
	Water-wheel tail measurement	Computer vision, YOLOv8	Image calibration, projective transformation	[156]
	Aquaculture area identification	Supervised semantic segmentation, CNN	Multisource remote sensing images, feature extraction, and classification	[157]
	Scallop recognition	Deep learning (semantic segmentation)	Time-lapse images and automated identification	[158]
	Fish response analysis	CNN, U-Net++	Sonar data, image segmentation, fish distribution pattern analysis	[159]
	Raft and cage aquaculture area identification	Computer vision, deep learning (CNN)	High-resolution remote sensing images	[160]
	Wetland classification and monitoring	ResU-Net	Multi-temporal satellite images, digital elevation models	[161]
	Marine organism classification	U-Net	Multifrequency echograms, semantic segmentation	[162]
	Wild fish monitoring	Deep learning (YOLOv4)	Underwater video, object detection	[163]
	Underwater fish detection	Mask R-CNN, Gaussian mixture models (GMMs)	Underwater wireless sensor network data, background subtraction	[84]
	Crustacean larvae counting	Computer vision, YOLOv5s	Image processing, object detection	[164]
	Fish counting	Improved YOLOv5, transformer module	Fish scale detection, model enhancement	[165]
	Fish biomass estimation	Deep learning (YOLOv5n), stereo vision	Real-time image detection and object extraction	[166]
	Tuna egg quality prediction	Vision-based CNN, R-CNN	Pre-trained images and object detection	[167]
	Shrimp seed counting	Modified CSRNet	Smartphone images, crowded scene recognition	[11]
	Fish weight estimation	Mask R-CNN, regression models	Image dimension extraction, regression learning	[72]
	Slaughter trait prediction	Image analysis, statistical modelling	Fish images, feature analysis	[168]
	Decision support systems	Geospatial techniques, cloud computing, GIS	Landsat imagery, database management, web-based applications	[169]
	Aquaculture area extraction	SAMALNet, RADNet, DeepLabV3+, dense residual U-Net, U-Net, R3Det, SegNet, U-Net++, marker-controlled watershed segmentation	Satellite image segmentation, feature fusion, model training, and accuracy assessment	[170]
	Cage segmentation	Swin transformer, ensemble learning	Remote sensing images, feature extraction, segmentation	[171]
	Object detection	Mask R-CNN, recursive feature pyramid (RFP), DCN	Computer vision, object detection	[172]
	Seaweed farm classification	Deep learning (DCGAN, U-Net, DeepLabv3, SegNet)	Time series dataset, feature enhancement, sample amplification	[53]
	Pond extraction	U2-Net	Remote sensing images, model training	[173]

) counting in aquaculture by improving accuracy, reducing labour costs, and enabling real-time monitoring. While significant progress has been made, environmental variability, occlusion, and data scarcity still need to be addressed. Ongoing research aims to refine these models by incorporating multimodal sensing, enhancing dataset diversity, and optimising algorithms for deployment in real-world aquaculture settings.

5. Research Gaps

Despite the extensive integration of AI applications in aquaculture, existing studies primarily focus on fish health monitoring, behaviour analysis, aquaculture area monitoring, and feeding management. Various AI techniques, including computer vision, deep learning, and machine learning, have been applied for disease detection, fish behaviour analysis, satellite image processing, and feeding behaviour identification. However, notable gaps remain.

(i)

Limited AI Integration for Broodstock Management: While much attention has been given to monitoring fish health, growth, and environmental conditions during the grow-out phases of aquaculture production, there is a notable lack of research on AI-driven broodstock management. Broodstock management, the critical process of selecting and caring for parent fish for breeding, significantly influences aquaculture operations’ overall success and sustainability. Effective broodstock selection, breeding optimization, and tracking physiological health are essential for improving aquaculture productivity. AI has the potential to revolutionise this area. For instance, AI could analyse extensive genetic datasets from broodstock to predict optimal breeding pairs, thereby accelerating selective breeding programs and enhancing desirable traits in offspring. Furthermore, AI can contribute to monitoring broodstock’s physiological health by analysing sensor data related to hormone levels and activity patterns, enabling more precise timing of breeding interventions and improving reproductive success rates. The limited exploration of AI in broodstock management represents a significant opportunity to enhance aquaculture practices.

(ii)

Lack of Real-Time, Multimodal AI Systems: Many existing AI solutions in aquaculture operate as isolated applications, focusing on data from single sources. For example, CNNs might be used for disease detection based on images, or YOLO models might be applied for biomass estimation. However, there is a lack of real-time, multimodal AI systems that integrate information from multiple sensor inputs to provide a more holistic understanding of the aquaculture environment. Such systems would combine data from sources such as water quality sensors (measuring parameters like temperature, pH, and dissolved oxygen), video cameras (for observing fish behaviour), acoustic sensors (for detecting fish sounds related to stress or disease), and even genetic profiling. Integrating these diverse data streams through AI analysis would allow for identifying subtle correlations and patterns that single-source analysis might miss. This capability would enable more accurate and timely decision-making, for example, by predicting disease outbreaks early through the correlation of water quality changes and behavioural anomalies, or by optimizing feeding strategies based on genetic information and growth rates. The development of these multimodal AI systems is crucial for advancing aquaculture management.

(iii)

Scarcity of AI Solutions for Crustacean Aquaculture: While numerous studies have explored AI applications for fish species, AI solutions are relatively scarce and tailored to crustacean aquaculture, particularly for species like mud crabs. Crustacean aquaculture presents unique challenges that necessitate specialised AI approaches. For instance, the complex life cycle of crustaceans, with its multiple larval stages, demands specific monitoring techniques. AI offers the potential to develop automated systems for monitoring crustacean broodstock, tracking reproductive cycles, and assessing health. It can also be applied to analyse microscopic images of larvae to evaluate their health and predict survival rates, enabling early interventions to improve production yields. Moreover, AI can contribute to optimizing environmental conditions for crustacean growth and survival by predicting the effects of temperature and salinity. Developing AI solutions designed explicitly for crustacean aquaculture is essential for enhancing this important sector’s sustainability and economic viability.

(iv)

Limited AI-Driven Predictive Analytics for Sustainability: Many current AI applications in aquaculture focus on classification and detection tasks, such as identifying diseased fish. However, there is a need to move beyond these applications and develop more advanced AI-driven predictive analytics to enhance the sustainability of aquaculture practices. AI has the potential to create forecasting models that predict future events, enabling proactive management strategies. Examples of such applications include the early prediction of disease outbreaks, which would allow farmers to implement preventive measures and minimise losses, and the automated adjustment of breeding conditions based on predicted environmental changes, optimising reproductive success and reducing resource waste. Furthermore, AI can forecast yields, assisting farmers in production planning and marketing. Developing and implementing AI-driven predictive analytics is crucial for promoting environmentally and economically sustainable aquaculture.

(v)

Challenges in AI Model Generalization and Transferability: AI models’ limited generalisation and transferability is a significant challenge in applying AI to aquaculture. Many models are developed for specific environments or datasets, which restricts their applicability to diverse aquaculture settings. Aquaculture environments vary significantly regarding species, geographic location, environmental conditions, and management practices. Therefore, there is a pressing need for research on adaptive AI models that can generalise across these different contexts. This includes developing robust AI models that are less sensitive to variations in input data and employing techniques such as transfer learning, which enables models trained on one dataset to be adapted for use with another. Creating adaptive learning systems that continuously adjust their parameters based on new data is essential for achieving generalization across diverse aquaculture environments. Overcoming these limitations in model generalization and transferability is vital for the widespread and effective adoption of AI in aquaculture.

These research gaps highlight the need for AI-driven broodstock management systems that integrate real-time monitoring, predictive analytics, and adaptive learning for sustainable aquaculture development.

6. Challenges to AIoT Implementation in Aquaculture

Artificial Intelligence of Things (AIoT) integration is undeniably transforming aquaculture practices, yet its widespread adoption faces significant hurdles. A primary obstacle is the substantial initial investment required. Deploying comprehensive AIoT systems encompassing sensor networks, robust data infrastructures, and powerful computational resources often presents prohibitive upfront costs, particularly for small to medium-sized aquaculture farms. This initial outlay includes purchasing and installing diverse sensors for monitoring parameters like water quality, temperature, and salinity, as well as data processing and storage hardware. Beyond the initial setup, ongoing sensor maintenance, software updates, and data storage expenses compound these financial challenges, demanding significant capital and specialised technical knowledge.

Furthermore, the inherent variability of aquatic environments poses a considerable challenge to the effectiveness of the AI model. Aquaculture operations are constantly influenced by fluctuating water quality, temperature, and salinity. These dynamic conditions can lead to AI models that struggle to generalise across aquatic settings or species. Consequently, frequent retraining and adaptation become necessary, consuming substantial resources and time, especially for smaller operations. The interconnected nature of AIoT systems also raises critical concerns regarding data privacy and security. The continuous collection of sensitive operational data, including fish health and environmental conditions, creates hacking and unauthorised access vulnerabilities. Ensuring the integrity and confidentiality of these data is paramount to prevent economic losses, protect proprietary farming techniques, and maintain consumer trust. Therefore, stringent cybersecurity measures, such as encryption, secure network architectures, and robust access controls, are essential to mitigate these risks.

Scalability presents another significant barrier to AIoT adoption in aquaculture. The diverse nature of aquaculture operations, ranging from small family farms to extensive industrial facilities, complicates the widespread deployment of AIoT systems. Small-scale farms especially might struggle with the high expenses of installing widespread sensor networks and handling the data they generate. Additionally, setting up and maintaining AIoT systems demands specialised knowledge in data science, machine learning, and IoT device management. Many aquaculture operators, particularly those in small to medium-sized businesses, may not have the expertise to implement and maintain these technologies effectively. Beyond the technology itself, a critical barrier is the need for a skilled workforce. Setting up, maintaining, and effectively utilizing AIoT systems demands specialised knowledge in areas such as data science, machine learning, IoT device management, and cybersecurity. Many aquaculture operators, particularly those in small to medium-sized businesses or remote locations, may lack the expertise to implement and maintain these advanced technologies effectively. This highlights a significant need for targeted technical training programs and comprehensive workforce development initiatives within the aquaculture sector to bridge this skills gap and ensure successful AIoT adoption and utilization. The complexity of interpreting AI-driven insights and translating them into practical farming decisions also requires adequate training.

A further limitation lies in the generalisation and transferability of AI models across different species and applications. Many AI models are tailored explicitly to particular species, farming methods, or environmental conditions, restricting their broader applicability. For example, a model developed using data from Atlantic salmon farming in net pens might not perform effectively when used for shrimp aquaculture in ponds or tilapia in recirculating systems, as their behaviours, environmental needs, and disease profiles differ significantly. This limited flexibility often requires repeated retraining or tailored adjustments to the model for each new application, which can be time-consuming and expensive, thus slowing down the broader implementation of AIoT technologies across the diverse aquaculture industry.

Finally, the environmental impact of AIoT systems must be considered. The energy-intensive nature of data processing and monitoring equipment can contribute to environmental degradation, potentially undermining aquaculture’s sustainability goals. Furthermore, the successful deployment of AIoT systems relies heavily on adequate physical and digital infrastructure. In many remote aquaculture locations, existing infrastructure may be insufficient to support the required stable power supply and reliable internet connectivity for extensive AIoT deployments. Establishing and maintaining the necessary communication networks through technologies like LoRa, 5G, or satellite internet can be challenging and costly in these areas, limiting the feasibility of real-time data transmission and analysis. The physical robustness of equipment in harsh marine or aquatic environments, subject to factors like biofouling and corrosion, also presents an infrastructure challenge requiring durable and easily maintainable designs.

7. Future Directions in AIoT for Aquaculture

While current AIoT applications significantly impact aquaculture through smart feeding, water quality management, behaviour monitoring, disease detection, biomass estimation, and automation, substantial work remains to realise their full potential. Future endeavours should prioritise enhancing real-time adaptability, building resilience against environmental variations, ensuring extensibility, and achieving cost-effectiveness across diverse aquaculture settings. An essential focus should be improving real-time decision-making, enabling aquaculture facilities to respond swiftly to dynamic environmental changes. This would facilitate continuous operations, even in remote environments, allowing for the optimisation of critical parameters like feeding intensity and water quality to support fish health and environmental sustainability.

Furthermore, adapting AIoT systems to various species and environmental conditions represents an ongoing challenge. Many models that perform well under specific conditions struggle with generalisation when faced with variations in lighting, water turbidity, or species behaviour. Future studies should focus on creating adaptive algorithms and multimodal systems to address these challenges. This can be accomplished by utilising diverse training datasets encompassing various species and environments or leveraging transfer learning methods to enhance models’ ability to generalise across different aquaculture settings.

Scalability and cost-effectiveness are crucial for more widespread use of AIoT, particularly among small aquaculture operators. Innovations in sensor integration with mobile-compatible systems and using cloud-based platforms for data processing can significantly reduce hardware requirements and make AIoT solutions more accessible. Integrating diverse data sources, such as visual, acoustic, and environmental data, into unified monitoring platforms is also essential. As demonstrated in biomass estimation and behaviour analysis, this multimodal approach provides a comprehensive view of fish health and population dynamics, fostering resilience by mitigating reliance on single data sources.

Enhancing early disease detection and prevention by real-time biosensing is another critical area for development. Future research should focus on integrating real-time biosensors that track pathogen levels, enabling early detection and intervention. AIoT solutions that are mobile and compatible with offline use can enhance disease detection in remote and resource-constrained aquaculture facilities. Recent advancements in biosensing technologies, such as microfluidic devices and lab-on-a-chip platforms, show promise for rapid and sensitive detection of disease biomarkers in aquaculture environments. Combined with AI-driven analysis, these technologies can enable early-warning systems and automated responses to mitigate disease outbreaks, representing a significant step forward in proactive aquaculture management. Additionally, the integration of advanced automation and robotics has the potential to transform aquaculture operations significantly. Future robotic systems should prioritise adaptability, efficiency, and reliability in diverse environments facilitated by enhanced sensor robustness, algorithm efficiency, and improved energy solutions. Multi-sensor fusion combined with AI can lead to highly autonomous robotic systems capable of performing routine tasks, reducing labour, and improving operational efficiency.

Energy independence is paramount for aqua farms located in remote or offshore areas. Integrating renewable energy sources like solar and wind power with AIoT-enabled energy optimisation can reduce reliance on conventional energy sources and minimise operational costs. Finally, digital twin technology, involving the development of virtual aquaculture system models, offers promising potential for simulation in real time, optimisation, and predictive analytics. The development of digital twins in aquaculture is gaining momentum, with initial implementations focusing on modelling water quality dynamics and fish growth. These virtual models are used to optimise feeding strategies, predict environmental impacts, and improve overall farm management. As the technology matures, digital twins are expected to play an increasingly important role in decision support and automation in aquaculture. Mobile and offline-compatible AIoT solutions can extend disease detection capabilities to remote and resource-limited aquaculture facilities. Advanced automation and robotics are also poised to revolutionise aquaculture operations.

8. Conclusions

While AIoT holds immense promise for revolutionising aquaculture through enhanced efficiency, sustainability, and productivity, its full potential is yet to be realised. The paper highlights that current AIoT applications offer significant benefits in water quality management, disease detection, smart feeding, and behaviour analysis. However, significant research gaps remain, particularly in broodstock management, multimodal system integration, crustacean aquaculture applications, predictive analytics for sustainability, and model generalisation.

Challenges such as high initial costs, variability in the environment, data privacy difficulties, scalability problems, and the requirement for specialised technical knowledge also hinder the successful implementation of AIoT. To overcome these hurdles, future research and development should focus on enhancing real-time adaptability, improving species and environmental adaptability, ensuring cost-effectiveness, developing integrated multimodal systems, enhancing disease detection through biosensing, advancing automation and robotics, implementing sustainable energy solutions, and leveraging digital twin technology.

Ultimately, we paper conclude that AIoT can drive a more efficient, sustainable, and resilient aquaculture sector. However, achieving this requires addressing existing gaps and challenges through strategic investments, innovative solutions, and a commitment to continuous improvement and adaptation.