Can AI See Fish on Sonar?

First, a disclosure

Our app does not connect to a fish finder, a transducer, or any sonar hardware. Nothing in this article describes a feature we sell. We are writing it because “AI fish finder” is a question people genuinely search, and the science behind it is better documented — and more interesting — than the marketing around it.

What sonar actually measures

An echosounder is a sound-timing instrument, and everything on the screen follows from that. NOAA’s survey documentation describes it directly: the transmitter emits short pulses of sound, and when a pulse hits an object such as a school of fish, the sound bounces off and scatters in many directions. The receiver measures what comes back and how long it took.

Why do fish show up at all? Air. The swim bladder’s air reflects sound far more strongly than bone or flesh, because its density differs so sharply from water. NOAA’s ocean service adds the sizing intuition: the resonant swim bladder provides the greatest contrast, and the larger the fish, the stronger the returning echo.

So the screen shows reflected energy, positioned by time. Strong return, something reflective; deep return, something far. Nothing in the physics attaches a name to the mark.

Why the same fish is not the same echo

If echo strength mapped neatly to fish size, sonar reading would be arithmetic. It does not, and the reasons are measured.

The ICES reference work on acoustic target classification lists them: the dominant behavioural factor is the fish’s orientation — its tilt angle relative to the incoming sound wave — and backscatter further varies with swim bladder morphology, life stage and depth. The echo measurements themselves are described as stochastic, shifting with unobserved changes in the environment and the target.

Depth deserves its own line, because the mechanism is elegant and unforgiving. Pressure compresses gas, so the swim bladder that makes a fish visible shrinks as it descends. A study of Mueller’s pearlside modelled the effect: at about 85 metres, the acoustic size of the bladder was roughly half its surface value, and maximum target strength fell with depth accordingly. The identical fish, deeper, reads smaller.

Tilt a fish, age it, or sink it, and its echo changes. That is what any classifier — human or machine — is working with.

What AI has genuinely achieved

Within those limits, the published results are impressive and honest about their scope.

The landmark study trained a convolutional network on Norwegian North Sea survey data — multifrequency scientific echosounders at 18 to 200 kHz, labelled by experts across a decade. It distinguished sandeel schools from other species and from background with an F1 score of 0.87, and separated sandeel schools from other-species schools at 0.94. Genuinely strong — at what is essentially a one-species problem: this target, versus everything else, in one sea.

The more revealing paper is the follow-up, because it reports what happened in production. The same institute integrated three models into its real survey pipeline and found biomass estimates generally similar to manual annotation — with variation across years. One model misclassified a surface layer as sandeel and was prone to seabed contamination. The stated conclusion: full automation was not yet feasible, but predictions could serve as starting points for manual scrutiny. Performance changed by year for reasons the training data did not fully explain.

That is the current frontier, run by the people most motivated to make it work: AI as a strong first pass, a person still checking.

The net at the end of the argument

Here is the detail that settles the species question. NOAA’s acoustic surveys — calibrated instruments, multiple frequencies, trained analysts — verify their echograms by fishing. The survey description says it without embarrassment: areas are ground-truthed using a net to collect a sample, and the catch identifies the species, size, sex and age of what the sonar showed.

The ICES report frames it as a rule: for mixed-species aggregations, classification to species requires ancillary information such as trawl catches or camera images. Frequency response helps, shape and depth help, prior knowledge of the water helps — and the confirmation is still a net.

A consumer unit has one transducer, no calibration exercise, and no trawl. When it prints a species name beside a mark, it is making an educated guess from context — depth, size, water, season. Sometimes a good guess. Never a measurement.

How to read a screen honestly

None of this makes sonar less useful; it clarifies what kind of instrument it is. Sonar is superb at the questions sound can answer: how deep, how much structure, whether something reflective is present and at what depth it is holding. Those are exactly the questions an angler most needs answered.

Treat any species label — AI-assisted or not — as a hypothesis to test with local knowledge: what lives here, what holds at this depth in this season, what the fish you actually catch turn out to be. That last check is the angler’s version of the trawl, and no algorithm has retired it yet. The same logic runs through photo identification: detection is the strong suit, naming is the hard part, and the honest systems say which is which.