Aquaculture 4.0: Smart Fish Farming

• Aquaculture 4.0 is a production paradigm that integrates IoT, AI, robotics, and digital twins to enable real-time, efficient, and welfare-aware fish farming.
• It leverages continuous sensor networks, advanced computer vision, and cloud analytics for precise feeding, water quality control, and behavioral monitoring.
• Demonstrated applications include underwater robotic inspections, blockchain-based data exchange, and experimental digital twins that optimize operational performance.

Aquaculture 4.0 denotes the aquaculture analogue of Industry 4.0: a transition from labor-intensive, manual, and reactive farming toward smart, data-driven, automated, and digitally integrated operations built around IoT technologies, data analytics, robotics, AI, and real-time decision-making (Akram et al., 16 Jul 2025). In the research literature, the same transition is also articulated through adjacent terms such as smart fish farming and precision aquaculture, emphasizing continuous sensing, cloud or edge connectivity, intelligent perception, predictive control, and data-driven intervention across feeding, water quality, welfare, infrastructure, and logistics (Yang et al., 2020). Operationally, Aquaculture 4.0 is often realized as a networked cyber-physical system in which sensors generate data, software models interpret it, communication layers distribute it, and control or advisory modules support monitoring, traceability, and smarter operational decisions (Papadopoulos et al., 2023).

1. Conceptual scope and operating logic

A central theme in the literature is that Aquaculture 4.0 is not a single device class or algorithmic family, but a production paradigm. Smart fish farming is described as the integration of IoT, big data, cloud computing, AI, and modern sensing technologies to achieve real-time data collection, quantitative decision-making, intelligent control, precise input use, and personalized management (Yang et al., 2020). Within that framing, deep-learning work has already organized major application areas into live fish identification, species classification, behavioral analysis, feeding decision-making, size or biomass estimation, and water quality prediction (Yang et al., 2020).

A later synthesis extends that scope from recognition and prediction to generation, planning, communication, and compliance. It maps Aquaculture 4.0 applications across sensing and perception, robotics and automation, planning and optimization, communication and reporting, traceability and regulatory compliance, and digital twins and simulation, while treating multimodal generative AI as an enabling layer for smarter and more adaptive decision-making (Akram et al., 16 Jul 2025). This broader formulation is consistent with works that focus on secure data exchange, autonomous inspection, precision feeding, and structural monitoring rather than only on computer vision or only on water-quality sensing.

The problem inventory addressed by Aquaculture 4.0 is similarly broad. Domain-specific AI work identifies disease outbreaks, feeding inefficiency, labor and regulatory burden, hatchery mortality, and unstable water quality as recurring bottlenecks; other studies emphasize net damage, biofouling, remote monitoring constraints, or high operational losses from delayed reaction to environmental anomalies (Narisetty et al., 28 Jul 2025). This suggests that Aquaculture 4.0 is best understood as a systems-level attempt to couple biological production, environmental sensing, digital infrastructure, and operational intelligence rather than as a narrow modernization of individual farm tasks.

2. Sensing, connectivity, and trusted data infrastructures

One of the clearest Aquaculture 4.0 patterns is the replacement of episodic manual observation with continuous sensor networks and end-to-end data pipelines. A cost-effective IoT architecture for a real aquaculture facility was structured into a local layer and an external layer: sensor-actuator nodes sent water data via LoRa/LoRaWAN to a gateway, which forwarded JSON-encoded records over HTTP to a cloud server and MongoDB-backed monitoring application (Teixeira et al., 2021). In validation, the node operated with deep sleep around $10\ \mu A$, woke every 10 minutes, and supported a full sensor-to-app pipeline that ran for 66 hours nonstop, stored 392 successful readings, and exhibited 2.3% packet loss; site-specific radio surveying showed reliable communication at about 43–104 m and failure at 117–143 m (Teixeira et al., 2021). The result is important because it demonstrates feasibility under realistic constraints while also showing that nominal long-range claims must be tempered by obstacles, vegetation, and farm geometry.

Low-cost embedded sensing has also been coupled to cloud logging and downstream machine learning for pond suitability assessment. An Arduino Uno and Ethernet Shield architecture combined pH, DS18B20 temperature, turbidity, and ultrasonic depth sensing with ThingSpeak storage and WEKA-based classification, using measurements from five ponds and labels for 11 fish categories (Islam, 11 Jan 2025). In that study, only three ponds satisfied the stated reference ranges, and Random Forest yielded 94.42% accuracy, 93.5% kappa statistics, and 94.4% Avg. TP Rate (Islam, 11 Jan 2025). The significance lies less in algorithmic novelty than in the full chain from sensor acquisition to operational pond-level decision support.

Aquaculture 4.0 research has also treated data infrastructure as a security and governance problem. A “future fish farm” proof of concept implemented a private-permissioned Hyperledger Fabric 2.3.0 network with two organizations, RAFT crash-fault-tolerant ordering with 3 orderers, CouchDB state storage, X.509 identity control via the Membership Service Provider, Go chaincode, and two private data collections separating shared operational data from fish-farm-only sensitive data (Papadopoulos et al., 2023). Shared fields included windspeed, rainfall, airpressure, temperature, waveheight, and watercurrent, while private fields included fdom, salinity, ph, turbidity, algae, orp, and nitrates; the PoC reported write transaction time of approximately 7.3 seconds, read transaction time of approximately 6.9 seconds, throughput of about 7–8 transactions per minute, and stable performance at 100,000 stored records (Papadopoulos et al., 2023). In Aquaculture 4.0 terms, this work extends digitalization beyond telemetry into privacy by design, immutable logging, traceability, auditability, and policy-enforced multi-party access control.

3. Computer vision, behavioral analytics, and precision feeding

Feeding is one of the most intensively digitized domains in Aquaculture 4.0 because it directly couples economics, welfare, and environmental loading. A real-time monitoring system for European sea bass in marine cages used underwater cameras, an automatic feeder controlled by a Raspberry Pi microcomputer, Python v3.9, YOLOv5 for detection, DEEPSORT for tracking, and group-level metrics including fish speed and a new Feeding Behavior Index (FBI) derived from density variation during feeding (Georgopoulou et al., 2024). The system monitored a population of over 10,000 fish in a 40 m diameter, 9 m deep cage under four feeding conditions—normal feeding, reduced feeding at 50% of normal, overfeeding at 150% of normal, and no feeding—and used 10th-degree polynomial smoothing, Gaussian Mixture Models, Kruskal-Wallis tests, and Dunn post hoc tests to extract behavioral phases and compare responses (Georgopoulou et al., 2024). Reduced feeding produced mean speed $0.61 \pm 0.10\ \text{bd/sec}$, significantly higher than normal feeding at $0.54 \pm 0.11\ \text{bd/sec}$ with Statistic $= 4.48$, $p\text{-value} = 0.03$; overfeeding and no feeding yielded $0.57 \pm 0.12$ and $0.56 \pm 0.15\ \text{bd/sec}$, and GMM clustering showed that under overfeeding the post-feeding cluster appeared 10 minutes before feeding stopped, whereas under reduced feeding the second feeding cluster extended beyond feeding (Georgopoulou et al., 2024). These results were explicitly interpreted as a basis for predicting satiation level and controlling feeding duration.

An integrated computer vision and IoT approach for tilapia feeding similarly turned size estimation and fish counting into actuator commands. The system combined pH, dissolved oxygen, and temperature sensors, two cameras, an STM32F103C8 microcontroller, a feeding pump, a pH pump, a load cell with HX711 amplifier, two servo motors, MQTT-based communication, backend inference, and mobile app monitoring (Hossam et al., 2024). On the vision side it used YOLOv8 for keypoint detection and fish counting, GLPN for depth estimation, and the classical morphometric relation

$W = aL^b$

with $a = 0.014$ and $b = 3.02$ for Tilapia to convert estimated length into weight and then into feed allowance (Hossam et al., 2024). The dataset contained 3,500 annotated images, and the reported keypoint metrics were Precision 94.50%, Recall 89.71%, AP@50 99.68%, and AP@75 94.16%; fish counting achieved Precision 96.21%, Recall 86.82%, AP@50 98.88%, AP@75 92.47%, and 94.5% counting accuracy on 100 frame pairs with average absolute error 0.7 fish per frame pair (Hossam et al., 2024). The paper also described a 58-fold production increase as a preliminary/theoretical maximum estimate rather than a demonstrated field outcome.

Another feeding-oriented line of work focuses on feed-particle dynamics as a proxy for fish response. A tuna nutriment tracking method used stabilization, YOLOv3 detection, minimum-cost data association, and quadratic trajectory mapping on a 419-frame real-farm video with nutriment objects as small as $0.61 \pm 0.10\ \text{bd/sec}$0 to $0.61 \pm 0.10\ \text{bd/sec}$1 pixels (Pradana et al., 2021). With human-annotated trajectories as reference, the method achieved an average error distance of 21.32 pixels with standard deviation 3.08 pixels at $0.61 \pm 0.10\ \text{bd/sec}$2, and ran at about 1.93 fps (Pradana et al., 2021). In the Aquaculture 4.0 context, the contribution is that feed tracking becomes an indirect sensing channel for behavioral state and feeding-machine decisions.

Feeding intelligence in Aquaculture 4.0 also includes explicit control theory and reinforcement learning. A population fish-growth model with biomass $0.61 \pm 0.10\ \text{bd/sec}$3, population size $0.61 \pm 0.10\ \text{bd/sec}$4, and manipulated variables $0.61 \pm 0.10\ \text{bd/sec}$5 was used to compare bang-bang control, PID, model predictive control, and Q-learning under different unionized ammonia exposure scenarios (Aljehani et al., 2023). The study reported that Q-learning prevented fish mortality and achieved good tracking errors under different levels of unionized ammonia, but potentially underfed the fish; in the most advanced formulation, MPC$0.61 \pm 0.10\ \text{bd/sec}$6 jointly optimized feeding, temperature, dissolved oxygen, and ammonia, yielding mortality $0.61 \pm 0.10\ \text{bd/sec}$7, RMSE 2.723%, and food consumption 2487.9 g for population 10 (Aljehani et al., 2023). This establishes an explicitly closed-loop interpretation of Aquaculture 4.0 in which perception and forecasting are tied to constrained control.

4. Robotic inspection, autonomy, and welfare-aware operation

Aquaculture 4.0 increasingly assigns inspection and monitoring to autonomous or semi-autonomous robotic platforms. An underwater robotic system for aquaculture net pens combined a Blueye Pro ROV X, a topside server unit connected through WiFi, rope-based distance estimation using Canny edge detection, Hough transform, and triangulation, a vision-based feedback control law for yaw and depth regulation, and YOLOv5 for detecting biofouling, vegetation, holes, and plastic debris (Akram et al., 2023). The vehicle was designed to traverse the net from top to bottom, maintain a reference distance of 200 cm, and inspect the net in real time; evaluation on unseen real-time images in a pool setup showed successful detection of plant, holes, and plastic, while the rope pair provided a usable visual reference for distance tracking (Akram et al., 2023). The paper directly targeted the cost, risk, and repeatability limitations of diver-based or teleoperated inspection.

Surface autonomy has been studied in parallel. The “Oboat” unmanned surface vehicle adopted a special round-shaped structure and redundancy motion design to balance mobility, stability, and portability while limiting disturbance to the aquaculture environment (Zhang et al., 2022). Its monitoring workflow combined line-of-sight navigation with PID control, environmental sensing of pH, temperature, dissolved oxygen, and turbidity/TDS-related measurements, ultrasonic fish-detection radar used to generate $0.61 \pm 0.10\ \text{bd/sec}$8 and $0.61 \pm 0.10\ \text{bd/sec}$9 occurrence labels, hybrid TSP + CPP path planning, and logistic regression for fish existence prediction (Zhang et al., 2022). In field trials the robot tracked a route of seven GPS turning points, grouped fish occurrence sites into five ROIs, and generally followed the intended intra-regional zigzag route despite natural GPS error on the order of 10 meters (Zhang et al., 2022). This workflow makes exploration, labeling, prediction, and revisit part of a single operational loop.

A distinct contribution of the robotics literature is its insistence that automation is not behaviorally neutral. An industrial-scale study at SINTEF ACE Korsneset used an Argus Mini $0.54 \pm 0.11\ \text{bd/sec}$0 ROV with dimensions

$0.54 \pm 0.11\ \text{bd/sec}$1

equipped with two Ping360 sonars and a stereo camera to examine how Atlantic salmon in cages of about 170,000 fish reacted to robot maneuvers (Evjemo et al., 2024). Sonar-based semantic segmentation, stereo vision, and a blind behavioral review all showed that high thruster activity increased fish distance, forward motion produced the strongest reaction, turning was weaker and more variable, and active upward movement was more disturbing than passive buoyant ascent in some cases (Evjemo et al., 2024). The significance is that Aquaculture 4.0 robotics must be designed around welfare-aware constraints rather than only around task completion.

A closely related stereo-vision study quantified fish responses to intrusive objects through caudal-fin detection and 3D reconstruction in industrial sea cages. The pipeline used YOLOv8, ByteTrack, SuperGlue, triangulation, and RAFT-Stereo testing, and compared several pre-processing strategies including DWT, morphological area opening, CLAHE, and Sea-thru (Alvheim et al., 28 May 2026). For the foreground caudal fins dataset, MO-WT augmented training gave Precision 0.898, Recall 0.895, and mAP50 0.961; for the full caudal fins dataset it gave Precision 0.841, Recall 0.781, and mAP50 0.860, while triangulation was more reliable than RAFT-Stereo in this underwater setting (Alvheim et al., 28 May 2026). Biologically, fish kept closer to smaller objects than to larger ones, closer to white objects than to yellow ones, and showed less conclusive shape effects (Alvheim et al., 28 May 2026). This is an important corrective to simplistic automation narratives: in underwater aquaculture, classical geometry can outperform a learned depth model, and welfare-relevant behavior depends on object properties as well as robot kinematics.

5. Digital twins, experimental platforms, datasets, and domain models

Digital twins provide one of the most explicit Aquaculture 4.0 formulations of cyber-physical integration. A multifidelity digital twin for net cages fused low-fidelity FhSim simulations with sparse high-fidelity field measurements from the SINTEF ACE fish farm through a nonlinear autoregressive Gaussian process framework (Katsidoniotaki et al., 2024). The modeled cage had 50 m diameter, 18 m depth, solidity ratio 0.21, 12 bridle lines, and 321 nodes; the low-fidelity database covered 1000 sea states spanning current velocities $0.54 \pm 0.11\ \text{bd/sec}$2 to $0.54 \pm 0.11\ \text{bd/sec}$3, significant wave heights $0.54 \pm 0.11\ \text{bd/sec}$4 to $0.54 \pm 0.11\ \text{bd/sec}$5, peak wave periods $0.54 \pm 0.11\ \text{bd/sec}$6 to $0.54 \pm 0.11\ \text{bd/sec}$7, and wave/current directions $0.54 \pm 0.11\ \text{bd/sec}$8 to $0.54 \pm 0.11\ \text{bd/sec}$9 (Katsidoniotaki et al., 2024). The twin ingested online metocean data and predicted mooring line loads and cage deformation in near real time; PCA retained $= 4.48$0 components at a 93% explained-variance threshold, and a comparison showed that GCNs outperformed the GP+PCA surrogate for full deformation prediction (Katsidoniotaki et al., 2024). Because an MAE $= 4.48$1 reference line was treated as operationally relevant for UUV safety, this work linked structural health monitoring directly to autonomous inspection and collision avoidance.

Experimental platforms are equally important because many Aquaculture 4.0 methods need realistic but controllable development environments. An enhanced octagonal tank of internal dimensions $= 4.48$2 and approximate volume $= 4.48$3 was built as a sea-cage replica for fish behavioral analysis, with a submersible LED line positioned near the surface facing downward to simulate artificial daylight and on-site lighting conditions (Voskakis et al., 2024). The sensor suite comprised a stereo video system, a monocular camera, an event camera, and a Ping360 360° sonar, while specially designed enclosures and a bottom deployment hole hid the sensing infrastructure to produce a “clean” open arena less likely to alter fish behavior (Voskakis et al., 2024). The work is infrastructural rather than algorithmic, but its significance is methodological: Aquaculture 4.0 requires sensor validation under conditions that preserve ecological realism.

Public datasets have begun to expand the empirical base of smart aquaculture beyond water-only monitoring. AQUAIR logged six indoor environmental quality variables—temperature, relative humidity, CO$= 4.48$4, VOC, PM$= 4.48$5, and PM$= 4.48$6—plus a proprietary score every 5 minutes for 84 days in a closed hatchery room of about 75 m$= 4.48$7 at Amghass Station 3, producing 23,856 cleaned rows (Sabiri et al., 28 Sep 2025). The sensor was mounted according to ISO 16000-1 at 1.5 m above the floor and more than 1 m from the nearest water surface, and the preprocessing pipeline normalized timestamps, anchored records to a 5-minute grid, linearly interpolated single-channel gaps of 10 minutes or less, and used a rolling Hampel filter with window $= 4.48$8 and threshold $= 4.48$9 for outlier correction (Sabiri et al., 28 Sep 2025). The descriptive statistics included median CO$p\text{-value} = 0.03$0 758 ppm and median PM$p\text{-value} = 0.03$1 12.2 $p\text{-value} = 0.03$2, while diurnal analysis showed pronounced feeding-time peaks in CO$p\text{-value} = 0.03$3, VOC, PM$p\text{-value} = 0.03$4, and PM$p\text{-value} = 0.03$5 after morning feeding and evening cleaning (Sabiri et al., 28 Sep 2025). The dataset is significant because it extends Aquaculture 4.0 sensing to the coupled air–water environment of indoor systems.

At the knowledge-model level, Aquaculture 4.0 has also moved toward domain-specific LLMs. AQUA was introduced as the first LLM tailored for aquaculture, supported by AQUADAPT, an agentic framework with Data, Expert, QA, Cleanup, and Scoring agents for synthetic QA generation and quality control (Narisetty et al., 28 Jul 2025). The reported corpus comprised 55,105 documents, the expert-defined taxonomy covered 11 major categories and 60+ subcategories, the final filtered dataset contained approximately 3 million QA pairs, and the model was fine-tuned from Mistral-7B-Instruct-v0.3 using LoRA for 2 epochs on 8 NVIDIA H200 GPUs in 32 hours (Narisetty et al., 28 Jul 2025). The fine-tuned GPT-4.1 judge achieved Spearman’s $p\text{-value} = 0.03$6, Kendall’s $p\text{-value} = 0.03$7, Pearson $p\text{-value} = 0.03$8, exact match 63.1%, off-by-1 match 91.7%, MAE 0.42, pairwise consistency 88.5%, and weighted Cohen’s $p\text{-value} = 0.03$9, while the final AQUA model achieved BLEU-4 49.19, ROUGE-1 51.45, ROUGE-2 30.98, and ROUGE-L 45.09 on a held-out validation set (Narisetty et al., 28 Jul 2025). In Aquaculture 4.0 terms, this is a move from isolated predictive models toward expert-aligned advisory systems and IoT-linked edge reasoning.

6. Limitations, misconceptions, and prospective directions

The literature is consistent that Aquaculture 4.0 remains constrained by data, deployment, and validation bottlenecks. Deep-learning reviews identify the need for large labeled datasets as a central bottleneck restricting further applications, alongside high computational demand, overfitting risk, low interpretability, incomplete and unbalanced data, and public datasets that do not fully represent real farming conditions (Yang et al., 2020). Generative-AI reviews broaden that list to data scarcity and fragmentation, real-time performance constraints, trust and explainability, environmental costs, and regulatory uncertainty (Akram et al., 16 Jul 2025). These are not marginal implementation details: they define the boundary between a promising prototype and an operational aquaculture system.

A common misconception is that digital sophistication alone guarantees field readiness. Several influential studies remain proof-of-concept or experimental platforms rather than full production deployments: the Hyperledger fish-farm system used synthetic data and a Minifabric-based testbed, the autonomous net-inspection system was validated in a lab pool, the enhanced tank was designed as a transfer platform for later open-cage use, and the tilapia feeding system relied on a single controlled environment and a limited fish-size range (Papadopoulos et al., 2023). This does not diminish their importance, but it does mean that Aquaculture 4.0 evidence is often strongest at the level of architecture, feasibility, and subsystem validation rather than generalized farm-scale performance.

Another misconception is that automation is automatically welfare-compatible. Industrial ROV studies show that forward motion and high thruster activity can provoke strong avoidance responses, while stereo-vision studies of intrusive objects show systematic differences in fish distance as a function of object size and color (Evjemo et al., 2024). Related work also shows that a more advanced learned depth estimator, RAFT-Stereo, was less reliable than classical triangulation in underwater sea-cage imagery (Alvheim et al., 28 May 2026). The broader implication is that Aquaculture 4.0 must remain behavior-aware and methodologically plural: welfare constraints, sensor physics, and environmental optics can outweigh nominal model sophistication.

Future directions in the literature point toward denser integration rather than replacement of current components. The blockchain fish-farm work proposes more functions, more participants, more realistic deployments, and AI techniques for extracting further insights from stored data (Papadopoulos et al., 2023). Precision feeding research proposes changepoint analysis, neural networks, prediction of future feeding demand, and extension to gilthead seabream and salmon (Georgopoulou et al., 2024). Domain-model work outlines multi-agent automation, computer vision for fish health and behavior monitoring, regional adaptation, continual learning from new publications and farm telemetry, IoT-linked edge advisory systems, and an eventual “Aquaculture Intelligence” platform (Narisetty et al., 28 Jul 2025). This suggests that the mature form of Aquaculture 4.0 is likely to be a layered environment in which sensing, robotics, digital twins, secure data exchange, closed-loop control, and expert-aligned AI operate as interoperable components of a single bio-cyber-physical production system.