Salmon ventilation: cameras flag hidden stress

The underwater cameras already installed in some marine pens can do more than show where salmon are swimming. An Australian and Norwegian team has developed a computer-vision method that detects fish heads, classifies each visible mouth as open or closed, and tracks those changes to estimate ventilation rate. On an independent test set of 100 Atlantic salmon, automated estimates were strongly correlated with manual observations (Pearson’s correlation coefficient: 0.82).

This creates a practical opportunity: farms may be able to detect a population-level change remotely without imposing another capture on fish already exposed to shifts in temperature and dissolved oxygen, parasites, handling or treatment. The output is not a diagnosis of gill disease and it does not directly measure oxygen consumption. It is an early warning that must be interpreted with water measurements, behaviour, feeding performance and health observations.

Reading a rapid movement in a commercial pen

Atlantic salmon, Salmo salar, draw water through the mouth and across the gills for gas exchange. The frequency of mouth-opening cycles is therefore a visible proxy for ventilatory activity. It can change with temperature, swimming activity, water quality, stress, gill condition and other physiological demands.

The method reported in Aquacultural Engineering was developed with video from commercial sea farms. A convolutional neural network finds visible heads and distinguishes open from closed mouths. A multiple-object tracker then connects detections across frames so that complete cycles can be counted. This design addresses a field problem rather than a controlled studio scene: fish cross the image at different depths and angles, obscure one another and leave the frame, while turbidity, motion blur and camera biofouling can reduce visibility.

The researchers do not attempt to maintain a lifelong record for each fish. They combine many sufficiently clear tracks to build a pen-level indicator. That distinction matters. A group median can reveal a collective shift, but it cannot reliably identify every distressed individual or represent fish that never enter a usable part of the image.

Associated with gill condition, but not a disease signature

A follow-up study applied the approach to 240 videos from commercial farms in Tasmania. Median ventilation rates ranged from 85.7 to 156.5 mouth-opening cycles per minute across pens, with an overall mean of 106.1 ± 12.1. Higher rates were associated with worse gross gill scores after the analysis accounted for dissolved oxygen, temperature and fish weight.

The effect was small, however, and the score distribution was narrow: 223 of the 240 pens, about 93%, had a score of 1 or less. The fish also had no laboratory confirmation of amoebic gill disease. Gross scoring described visible gill condition and did not establish a single cause.

A rise in ventilation should therefore not be translated automatically into “gill disease”. It may accompany low oxygen, warmer water, vigorous activity, irritation, infection or several pressures at once. Conversely, a stable pen median cannot rule out a focal lesion or a problem affecting a minority of fish. The measurement is useful within a decision pathway, not as a standalone verdict.

Following change around a freshwater bath

A third field study compared 47 paired videos recorded before and after freshwater baths at commercial Tasmanian farms. Median ventilation fell by an average of 2.56 cycles per minute after treatment. The automated method detected this change alongside records of water temperature and dissolved oxygen.

The result shows that camera data can document a physiological change around an intervention. It does not prove that the decrease came from removal of one specific pathogen, nor that every decrease is beneficial. Handling conditions, the post-treatment interval, water differences and activity can also affect ventilation. Operational monitoring therefore needs a comparable reference window and records of the pen, time of day, temperature, oxygen, feeding and recent events.

Turning a curve into an actionable alert

Farms should first establish normal variation for each site rather than import a universal threshold. An alert might identify a persistent departure from the pen’s own pattern under comparable time, temperature and activity conditions. Image quality belongs in the same dashboard: poor contrast, particles, lens biofilm, crowding in front of the camera or too few trackable heads can make an estimate unstable.

When a change appears, staff can confirm oxygen and temperature sensor readings, observe swimming and appetite, look for mortality or surface behaviour, and decide whether a targeted examination is justified. If the signal persists or is accompanied by other abnormalities, proportionate sampling and veterinary investigation remain necessary to identify the cause and select an intervention.

Automated ventilation could also compare periods before and after an operation, prioritise pens for inspection, or document recovery. It should not trigger repeated handling on the strength of one graph. An alert protocol needs to state who reviews the signal, which checks follow, when observation becomes action, and how false positives are recorded.

Technical limitations are part of the evidence

The algorithm does not see every fish in a commercial pen. The studies discard sequences that are too poor and require enough usable tracks; in the gill-score analysis, pen estimates were based on observations from at least 200 fish. This filter improves the stability of the estimate, but it also means that missing data cannot be interpreted as an absence of risk.

Performance may shift with camera type, lighting, season, fish size, swimming speed and site characteristics. Before operational use, local validation should compare a sample of automated and human counts, quantify rejected footage and challenge the system under difficult conditions. Video governance, equipment maintenance and traceability of model versions also become components of the health-monitoring programme.

There is a further sampling question. Fish close to a camera may not represent the full population if depth, appetite, current or health status affects where they swim. Positioning, observation windows and the number of cameras should therefore be documented, and trends should be combined with environmental and production data rather than treated as a census.

More observation with less handling

Computer vision can extract repeated physiological information from infrastructure already operating inside a sea cage. It may reduce exploratory captures and shorten the time needed to recognise a population-level change, offering potential welfare and husbandry benefits. Its strength is continuous observation; its weakness is the non-specific nature of the signal.

Vetofish can help farms define a local baseline, integrate ventilation with health and water-quality indicators, design escalation rules and organise veterinary assessment of alerts. The right goal is not to ask a camera to name a disease. It is to use the camera to know earlier when a closer look is warranted.