AI in Aquaculture

In this context, artificial intelligence means statistical and machine learning methods applied to the data a farm already generates: sensor time series, feeding and production records, images and video. The applications are prediction, classification, anomaly detection and decision support.

• A recirculating farm produces a continuous multivariate record that no operator can review in full.
• Many valuable signals are slow drifts across several parameters at once, which is exactly what human review of individual gauges misses.
• Repetitive estimation tasks — counting, sizing, classifying behaviour — are laborious and are performed inconsistently by people.
• The economics of the sector reward small persistent improvements in feed conversion and survival, which are the kind of gains statistical methods can find.

• Correlation found in one farm's record does not establish a mechanism, and does not transfer to another farm.
• Model performance degrades as the system changes — new species, new diet, new equipment — so monitoring the model is an ongoing task.
• Never place a model in a control path where a wrong output has an irreversible consequence for a live biomass.
• The uncertainty of a prediction is part of the prediction. An output without it invites misuse.

• Design the data infrastructure first: consistent sensor placement, calibration records, timestamped logging, local and remote retention.
• Record the context alongside the measurements — feeding, handling, maintenance, stocking, mortality — because those events explain most of the variation a model will encounter.
• Keep raw data as well as processed data, since future analysis will want questions that were not anticipated.
• Define the decision each analytical output supports before commissioning any modelling work.
• Keep models advisory. Retain the deterministic control and alarm layer as the safety system.

• Continue calibrating instruments; the value of the entire analytical layer rests on them.
• Record events accurately and promptly, since retrospective reconstruction is unreliable.
• Treat any model output as a prompt to investigate, and record what the investigation found — that record is what improves the model.
• Monitor model performance over time and expect it to degrade after any material change to the system.
• Be sceptical of vendor performance claims and require validation in the system concerned.

• Audit data quality on a schedule: gaps, calibration records, sensor drift, event record completeness.
• Re-validate models against recent data periodically, not only at deployment.
• Keep a record of model versions, training periods and the changes between them.
• Clean and reposition cameras and optical sensors, whose performance degrades with fouling faster than most operators expect.

• Attempting modelling before establishing reliable, calibrated measurement.
• Transferring a model between farms, species or diets and expecting the reported performance.
• Placing a model in a control path where a wrong output harms a live biomass.
• Evaluating on data that overlaps the training period, and mistaking the result for predictive performance.
• Presenting outputs without uncertainty, so they are read as measurements.
• Failing to record operational events, leaving the record without the explanations for its own variation.
• Treating a discovered correlation as a mechanism.

Hybrid approaches combine mechanistic models of the system — mass balance, gas transfer, nitrification kinetics — with statistical models fitted to the farm's data. The mechanistic component constrains the result to physically possible behaviour, which is a meaningful advantage in a domain where a purely statistical model can produce outputs that violate conservation of mass.

Data volume is a constraint the sector faces collectively. Individual farms rarely have enough history to train models for rare events, and pooling data across farms raises questions of comparability, commercial confidentiality and whether systems are similar enough for the pooling to be valid.

Machine learning earns its place in aquaculture where a long, calibrated, consistently recorded history supports a decision that repeats. Anomaly detection and vision-based measurement are the most established; most else is developing. Instrument first, record events alongside measurements, keep models advisory rather than in the control path, and validate in the system concerned rather than accepting figures from another one.