The Scottish salmon industry has a problem.
Scottish salmon survival for the 2022 year-class was 61.8%, which is the lowest in over three decades.
99% of the fish survive each month, but that small fraction that die adds up, and across an eighteen-month seawater phase nearly four fish in ten are lost.
So how do we detect this before it's too late? Our preliminary research shows that even before individual fish show signs of illness, the behavior of the entire school changes.
Good salmon
- Coordinated schooling, steady milling
- Moderate, stable swimming speed
- The school holds a depth band
- Normal feeding and rest rhythm
Bad salmon
- Cohesion falls, fish isolate
- Speed variance up, or mean speed down
- The school rises to the surface, when there is low oxygen
- Erratic darting, or lethargy
Why not track each fish
Camera systems count lice on individual fish, and need precision optics and computer vision to do it. This is necessary when you need a lice count per pen for a regulator. However, it is an expensive tool to answer simply, “is anything wrong here?”
Watching the school costs less, needs far less precision, and survives noise, murk and occlusion.
Current state of the art
Camera-based monitoring is well established and well funded: TidalX AI, Aquabyte, OptoScale and Dundee-based Ace Aquatec all sell computer-vision systems for lice detection, biomass estimation and welfare.
School-level hydroacoustics is established too, though to a lesser extent. Bluegrove, formerly CageEye, already runs echosounders on Scottish sites to monitor distribution and appetite. So our innovation is not sonar, and it is not school-level sensing.
Our difference is what the signal is used for. Bluegrove's product is designed to optimize feeding schedules. Ours flags that something is wrong, since deviations from each pen's own behavioural baseline can act as an early warning.
How much can be saved
We price against the value of the loss avoided.:
Every figure there needs checking with a producer, and farmgate price is volatile. The point is that the ratio survives being wrong by an order of magnitude.
Pipeline in action
Everything below is computed from a single simulation run. We have developed this pipelines as a proof of concept, but further development of the simulated fish behaviour, and metric tuning is needed. This shows how a cheap echosounder could be utilised to monitor the behaviour of the salmon as a group rather than tracking individual fish.
How each the metric are calculated
The animation below shows how raw sensor data is processed to generate each of the four metrics.
Each value is derived from the sensor output, not from the ground truth fish positions. The shaded band surrounding each trace is that pen's own rolling baseline, ±2.5σ. Sv represents the acoustic backscatter strength: the higher the Sv, the higher the likelihood of fish being present (approximately).
What building it told us
We ran a series of simulations to test our assumptions about how the sensor data translates into meaningful insights about the fish. These findings must be taken with caution, but can provide a starting point for further investigation.
Placement is very important
Preliminary research shows that salmon mill in a torus around the pen, usually in small schools. A downward transducer positioned at the centre of the pen may not detect any fish. In our simulation, eight metres off centre it samples about 40 of 180. This should be validated with real-world data in Phase 2.
Baselines do not necessarily transfer
In a experiment where echosounder data from multiple pens was collected, healthy pens judged against a different pen's baseline false-flagged three times in five. Against their own baselines, there was no false flagging. This implies that baselines are specific to each pen and cannot be naively applied to other pens.
Every pen learns its own band, and keeps learning — except while something looks wrong.
Depth beats agitation
Surface-dwelling behaviour flags in every pen at 10% prevalence. Agitation needs about half the school affected. So this is hypoxia and gill-health early warning that sometimes catches lice — a narrower claim, and the one the evidence supports.