FROM FRAGMENTED FARM DATA TO INTELLIGENT FEEDING: A MACHINE LEARNING FRAMEWORK FOR INTEGRATED DECISION SUPPORT IN AQUACULTURE

In this study, we present a machine learning–driven framework that transforms heterogeneous farm data into a unified basis for feeding decisions. Multiple data sources—including environmental measurements, feeding logs, biomass and welfare indicators, and production records—are integrated into a single analysis-ready dataset. The integration aligns data across site, cage, fish group, and time, with a focus on feeding events. This enables consistent analysis of feeding performance in its biological and environmental context (Fig. 1).

/>

Fig. 1: Integrated data framework showing ingestion, harmonization, and temporal alignment of multi-source farm data into a unified analysis-ready dataset.

A key contribution is the creation of a structured feature space that captures interactions between fish condition, environment, and farm operations. The dataset includes variables such as fish weight, growth trends, temperature during feeding, and management indicators like days since last treatment. Feature importance analysis shows that feeding behaviour is driven by a combination of factors rather than a single variable (Fig. 2). In particular, fish size, recent growth, and management history play a central role, confirming and extending previous findings from explainable machine learning approaches (Saad et al., 2024). This highlights the need to combine biological, environmental, and operational data when modelling fish appetite.

Fig. 2: Feature importance analysis showing the relative contribution of biological, environmental, and management variables in feeding predictions, highlighting the multi-factor nature of fish appetite dynamics.

Based on this dataset, a machine learning framework is developed to support feeding decisions through three stages (Fig. 3). First, a feed prediction model learns current feeding practice from historical data. Second, an outcome model estimates expected growth and feed efficiency (FCR). Third, both models are combined into a simulation layer that explores alternative feeding strategies and evaluates their expected impact. This creates a closed-loop approach where data is used not only for analysis but also to guide future decisions.

Fig. 3: Machine learning framework integrating feed prediction and outcome models into a simulation-based decision layer for evaluating and optimizing feeding strategies under varying production conditions.

Retrospective simulations using real farm data indicate that model-informed feeding adjustments can improve growth performance and reduce feed conversion ratio (FCR) in a large share of cases (Fig. 3). While these results are exploratory, they demonstrate the potential of integrating multi-source data and machine learning to move from reactive feeding toward predictive and adaptive strategies. This represents a shift from experience-based decision making to data-supported optimization under real production conditions.

More broadly, this work highlights a key shift in aquaculture. The challenge is no longer collecting data, but structuring and connecting it across systems. Traditional models rely on simplified relationships between fish size and environmental conditions (Aunsmo et al., 2014; Remen et al., 2016) and do not capture real production dynamics. In contrast, the proposed framework enables learning directly from operational data.

This approach provides a scalable foundation for intelligent feeding systems. It bridges data integration and modelling, enabling future developments such as real-time optimization and autonomous feeding. Ultimately, it supports data-driven feeding decisions for improved performance, reduced waste, and more sustainable production.

Acknowledgements

The work described has received funding from the internal RACE Autoföring project and has been carried out at the ACE facility of SINTEF Ocean. We are also grateful for the contribution of the industry partners, SALMAR, Piscada, and SpillFree in making this project possible.

References

[1] Aunsmo, A., Krontveit, R., Valle, P. S., & Bohlin, J. (2014). Field validation of growth models used in Atlantic salmon farming. Aquaculture, 428, 249-257.

[2] King, D. M., Dunn, M., Kerhervé, S., Nissen, O., & Saad, A. (2025). Robust Inference of Fish Behaviour from Underwater Acoustic and Environmental Data. Procedia Computer Science, 270, 5490-5499.

[3] Remen, M., Sievers, M., Torgersen, T., & Oppedal, F. (2016). The oxygen threshold for maximal feed intake of Atlantic salmon post-smolts is highly temperature-dependent. Aquaculture, 464, 582-592.

[4s] Saad, A. et al. (2024), Optimizing feeding strategies in aquaculture using machine learning: Ensuring sustainable and economically viable fish farming practices. Procedia Computer Science, 246, 4712-4721.