Underwater Acoustic Monitoring With AI

Passive acoustic monitoring in the ocean is dominated by a constraint that has nothing to do with machine learning: the instrument is on the seabed for months, you cannot reach it, and every choice you made before it went in is fixed until you get it back.

Sound behaves differently down there

Sound travels at roughly 1,500 m/s in seawater — about 4.4 times its speed in air — and the exact value depends on temperature, salinity and pressure, which is why sound rays bend rather than travel straight. In the deep ocean that bending creates the SOFAR channel, a depth at which the sound speed is at a minimum and energy is trapped, allowing very low-frequency calls to propagate for hundreds of kilometres.

Absorption is strongly frequency dependent, and this is the fact that organises the whole field. Low frequencies lose very little energy per kilometre; high frequencies are attenuated quickly. So a 20 Hz baleen whale call may be detectable across an ocean basin, while a 130 kHz porpoise click is gone within a few hundred metres. A hydrophone listening for the first is sampling the region; one listening for the second is sampling a small volume around itself, and the two designs have almost nothing in common despite using the same hardware.

Two calibration terms come with the territory. Hydrophone sensitivity is quoted in dB re 1 V/µPa, and underwater sound pressure levels are referenced to 1 micropascal rather than the 20 micropascals used in air — so an underwater dB figure is about 26 dB larger than an airborne one describing the same pressure, plus a further difference from the media’s impedances. Airborne and underwater levels are not comparable numbers, and treating them as such is a standard error.

The sample rate is a modelling decision

Nyquist sets the floor: to represent a signal component at frequency f you need to sample above 2f. That turns your target species into a number before anything else is decided.

• Baleen calls, roughly 10–100 Hz. A 1 kHz sample rate is generous. Instruments listening only for these can run for a year on one battery.
• Fish choruses and vessel noise, roughly 50 Hz–2 kHz. An 8 kHz rate covers it comfortably.
• Dolphin whistles, roughly 2–20 kHz. 48 kHz.
• Odontocete echolocation clicks, to 150 kHz and beyond. 384 kHz or higher, and now every downstream number is 384 times the baleen case.

A rate chosen too low does not degrade gracefully. Energy above Nyquist aliases down into the band you are keeping, so a high-frequency click train appears as a spurious low-frequency structure that a detector will happily learn. The anti-alias filter in front of the converter is the only defence, and it is in the instrument, not in your notebook.

A season of data, derived

Assume a single-channel instrument recording continuously at 256 kHz with 16-bit samples, deployed for a 120-day field season. Every number below follows from those three inputs.

  bytes per second = 256,000 samples/s * 2 bytes = 512,000 B/s

  bytes per day    = 512,000 * 86,400 s
                   = 44,236,800,000 B
                   = 44.2 GB per day  (41.2 GiB)

  per 120-day season = 44.2 GB * 120
                   = 5,308 GB
                   = 5.3 TB per instrument

  a 10-instrument array = 53 TB per season

Change one input and watch it move. At 24-bit resolution instead of 16 the season is 8.0 TB. At a 48 kHz rate — enough for whistles, not for clicks — the same season is 995 GB, a factor of 5.3 smaller. At 1 kHz for baleen work it is 20.7 GB, small enough to email.

The compression question comes up immediately and has an uncomfortable answer. Lossless codecs such as FLAC typically achieve something in the region of a 30–50% reduction on broadband ocean noise, which is real but does not change the order of magnitude. Lossy codecs achieve far more and are usually inadmissible: they are psychoacoustic, tuned to what a human ear does not notice, and everything they discard is in the quiet high-frequency region where clicks live. If archived audio is the evidentiary record for a survey, lossy compression destroys the record to save disk that costs less than the ship time did.

Duty cycling, and what it costs you

The standard mitigation is not to record continuously. A duty cycle of 10 minutes in every 40 records 25% of the time, and every number above divides by four: 1.3 TB per instrument per season, with battery life and disk both extended by the same factor.

What it costs is stated exactly by the arithmetic. You are blind for 75% of the deployment, so any density or rate estimate derived from the recording carries that as a sampling fraction, and events shorter than the off-period can be missed entirely. For an animal that vocalises many times an hour this is a fair trade — the estimator is unbiased if the off-periods are uncorrelated with behaviour. For a rare event, or for anything with a diel rhythm that happens to beat against your cycle period, it is not. Choosing a cycle period that is not a divisor of 24 hours avoids the worst version of that resonance.

The other mitigation is on-instrument detection: run a small detector in the field and store only triggered segments plus summary statistics. That converts a storage problem into a decision you can never revisit, because whatever the detector rejected is gone. Deployments that do this almost always keep a continuous low-rate channel or a periodic full recording alongside, precisely so the detector’s misses remain discoverable later.

Detection on a recording you cannot re-take

The modelling itself is conventional: compute a spectrogram, run a convolutional or transformer classifier over time-frequency patches, post-process frame scores into events. What is unconventional is the feedback loop, or rather its absence. You get the disk back in six months. If the gain was set too high and the loudest calls clipped, or too low and the quiet ones sat in the converter’s noise floor, that is the season.

So the pipeline is built backwards from that. The first pass over recovered data is not detection, it is data quality: check the proportion of samples at or near full scale to find clipping introduced by a gain setting, look for the flat-line signature of a flooded or failed element, check the recording timestamps for clock drift against the deployment log, and look for the broadband noise floor rising with sea state so detection thresholds can be conditioned on it rather than fixed.

Two more constraints shape the modelling. Labels are extremely expensive — an analyst annotating spectrograms produces a few hours of labelled audio a day against terabytes of unlabelled — so self-supervised pretraining on the unlabelled archive followed by a small supervised head is the norm rather than an optimisation. And class imbalance is extreme: a target species may vocalise in well under one percent of recorded time, which means a detector reported at 99% accuracy has said nothing, and the metrics that matter are precision at a fixed recall and the false-positive count per hour of recording. The same prevalence trap is worked through in full for a home sound detector, where the arithmetic is identical and the numbers are easier to hold in your head.