From disease detection to biomass forecasting: AI improves aquaculture risk strategy

Aquaculture has become the fastest growing food production sector in the world, expanding rapidly as global seafood demand rises. With that growth has come heightened exposure to biological threats, environmental instability and operational uncertainty, prompting a surge in artificial intelligence (AI) applications aimed at predicting and preventing losses.

In Artificial Intelligence in Aquaculture Risk Management: A Systematic Review by PRISMA, published in Applied Sciences, researchers examine how AI and machine learning are being used to manage these risks and whether predictive models are being effectively embedded into formal risk management systems.

AI expands across biological, environmental and operational risks

The review categorizes AI applications into four core risk domains: biological, operational, environmental and management frameworks. Of the 38 studies analyzed, 17 focused on biological risks, 15 on operational risks, three on environmental risks and three on formal management integration.

Biological risk dominated the research landscape. Machine learning models are increasingly deployed to predict disease outbreaks, detect parasites, analyze fish behavior and forecast mortality events. Deep learning architectures such as Convolutional Neural Networks and Long Short-Term Memory networks were widely used, particularly in image and video-based monitoring systems. These systems process underwater footage to detect visible symptoms, abnormal swimming patterns or early parasite infestations.

Other studies applied ensemble methods such as Random Forest and Gradient Boosting to large historical datasets of mortality, environmental variables and feeding rates. In several cases, models achieved high sensitivity in identifying high-risk periods weeks before outbreaks occurred. Early warning capabilities give farm managers time to adjust feeding, apply treatments or isolate affected cages.

Environmental risk applications, though fewer in number, showed strong technical maturity. LSTM-based models were used to predict dissolved oxygen levels, temperature shifts, ammonia concentrations and other water quality indicators. Accurate forecasting of dissolved oxygen depletion can prevent mass mortality events by triggering aeration systems before conditions reach critical thresholds. Some studies combined temporal convolutional networks with LSTM architectures to enhance prediction accuracy.

Operational risk applications addressed biomass estimation, feeding optimization, harvest timing, logistics and even financial volatility. Computer vision systems significantly reduced biomass estimation error compared with traditional sampling methods, which often exceed ten percent error margins. Improved biomass accuracy directly influences feed conversion efficiency, harvest planning and revenue projections.

In parallel, some researchers extended AI beyond production into financial risk modeling. Machine learning models were used to forecast salmon price volatility and calculate value-at-risk scenarios for biomass, linking biological uncertainty to economic exposure. These applications demonstrate that risk in aquaculture is not purely biological or environmental but deeply financial.

Across all domains, the distribution of algorithms revealed clear patterns. About 37 percent of studies relied on deep learning techniques. Ensemble methods accounted for nearly a quarter of applications, while traditional machine learning approaches such as Support Vector Machines and logistic regression remained relevant. Hybrid models combining statistical and AI methods also emerged, especially where mechanistic ecological knowledge complemented data-driven modeling.

Despite impressive predictive metrics reported in many studies, the authors caution that high accuracy on historical datasets does not automatically translate to reliability under novel conditions. Climate anomalies, emerging pathogens and unexpected operational disruptions can challenge models trained on past patterns.

The algorithm-to-action gap in aquaculture 4.0

Only three of the 38 analyzed studies explicitly linked machine learning outputs to structured risk management frameworks such as ISO 31000.

The majority of studies focused on risk identification and prediction. Few extended their models into standardized evaluation, treatment and monitoring protocols. This creates what the authors describe as an algorithm-to-action void. AI tools can detect anomalies, predict outbreaks or forecast water quality shifts, but their outputs often remain technical results rather than embedded decision rules within farm management systems.

The review highlights several reasons for this disconnect. First, disciplinary isolation persists. Data scientists develop high-performing models but may not engage with operational managers who make daily decisions. Conversely, farm managers may lack familiarity with AI tools or remain cautious about relying on black-box systems for high-stakes decisions.

Second, regulatory frameworks in aquaculture have not yet fully incorporated predictive analytics into formal risk management requirements. While financial industries operate under strict model governance standards, aquaculture remains in an early stage of integrating AI into compliance structures.

Third, trust and explainability remain barriers. Complex deep learning models can be difficult to interpret. Without transparency about which variables drive predictions, managers may hesitate to act on automated alerts. The review emphasizes the importance of explainable AI techniques, such as feature importance analysis and interpretable learning methods, to build operational confidence.

Quality assessment within the review reinforces this challenge. While 20 studies demonstrated high data quality and robust validation methods, only three achieved high operational integration. Thirty-two were classified as moderate in terms of linking technical outputs to unified management frameworks. Three showed low operational integration despite adequate technical performance.

This means that AI in aquaculture has matured at the predictive layer but remains underdeveloped at the governance and decision-support layer. The risk management cycle defined by ISO 31000 includes identification, analysis, evaluation, treatment and monitoring. Most AI studies concentrate on identification and analysis but stop short of embedding outputs into evaluation and treatment mechanisms.

Toward hybrid risk governance in smart aquaculture

The review does not merely diagnose the gap; it proposes a structural solution. The authors introduce the concept of a Contextualization Layer. This intermediate layer would translate raw model predictions into actionable business terms by filtering outputs through economic constraints, operational thresholds and ethical considerations.

In practice, this means that instead of presenting managers with abstract probabilities or regression outputs, AI systems would generate clear decision pathways. For example, a predicted drop in dissolved oxygen could automatically activate aeration pumps or generate tiered alerts based on risk severity and stock value. A disease risk forecast could trigger predefined treatment protocols or quarantine measures aligned with biosecurity standards.

Such integration would transform AI from a forecasting tool into a core component of risk governance. It would also align aquaculture with the broader digital transformation trend known as Aquaculture 4.0, where IoT sensors, digital twins and real-time analytics converge.

The review also sheds light on hardware and infrastructure challenges. Edge computing capabilities are needed to run machine learning models on low-cost devices, especially in smaller facilities. High-quality sensor networks remain expensive, and data scarcity limits model generalizability across species and geographic regions.

Emerging solutions include satellite-based dissolved oxygen estimation, soft sensors for chemical oxygen demand prediction and federated learning approaches that allow farms to train collective models without sharing sensitive raw data. These developments aim to reduce barriers to adoption while maintaining predictive strength.

From a policy perspective, the authors suggest that regulators and industry bodies could accelerate adoption by encouraging or requiring structured risk management plans that incorporate predictive analytics. Certification schemes or incentives tied to AI-based monitoring could shift the industry from reactive crisis management toward proactive risk mitigation.