Can Background Noise Help Drifting Larvae Hear Their Reef?

Oyster larvae drift with the current and cannot swim away from the sound they are exposed to. They settle when they hear a healthy reef, and vessel noise is known to reduce that settlement. The standard explanation is masking: noise drowns out the cue, so more noise always means worse detection. Physics and neurobiology suggest a different possibility. In a sensor that fires only when its input crosses a threshold, a moderate amount of random noise can push a faint, otherwise undetectable signal over that threshold. Too much noise swamps it again. This effect is called stochastic resonance, and it produces an inverted-U curve of detection against noise level. We propose to test whether whole, non-motile larvae show this inverted-U in settlement behavior when a faint reef cue is presented under graded, calibrated noise. The test is designed so that the masking account and the stochastic-resonance account make opposite, pre-registered predictions.

This matters for two reasons. First, whether peripheral noise benefits carry through to whole-organism behavior is an open question in stochastic-resonance research. Single sensory cells show the effect, but behavioral and fitness consequences are rarely tested. Second, environmental noise policy for passive drifters is written entirely in terms of masking and energy thresholds. If larvae near the edge of a reef's acoustic range detect it better at some noise levels than in silence, then "less noise is always better" is the wrong model for distant cues. The response to noise would depend on how strong the cue is.

Core Claim

Model larval cue detection as a population of noisy threshold units. That model makes three predictions a linear masking model cannot. First, for a reef cue below the larvae's detection threshold, the settlement gained by adding the cue (cue present minus cue absent, at the same noise level) rises and then falls as noise increases. Second, the noise level giving the largest gain shifts as the cue gets stronger, and the inverted-U disappears once the cue is clearly above threshold, where noise only hurts. Third, removing the noise energy from the frequency band the cue occupies abolishes the benefit, even when total noise power is held fixed.

A simple version of the model shows where these predictions come from. Let $A$ be the received cue level, $\theta$ the larval detection threshold and $\sigma$ the noise standard deviation in the cue band. The probability that a threshold unit fires in a listening window is roughly $\Phi((A-\theta)/\sigma)$ with the cue and $\Phi(-\theta/\sigma)$ without it, where $\Phi$ is the standard normal cumulative distribution. The cue's benefit is

$$D(\sigma) = \Phi\!\left(\frac{A-\theta}{\sigma}\right) - \Phi\!\left(\frac{-\theta}{\sigma}\right).$$

When $A<\theta$, $D$ is zero in silence, peaks at an intermediate $\sigma$, and decays at high noise. When $A>\theta$, $D$ falls monotonically. The no-cue term also predicts that noise alone raises settlement somewhat, through "false alarm" settlement. That rise follows a specific curve tied to the same $\theta$, so it is part of the test rather than a nuisance.

What Prior Work Establishes and Leaves Open

The behavioral starting point is Comparison of the effects of reef and anthropogenic soundscapes on oyster larvae settlement. That study played reef, vessel, combined, off-reef and silent treatments to Pacific oyster pediveliger larvae in 100-liter tanks. Reef sound raised settlement 1.64-fold over a no-speaker control and 1.44-fold over off-reef sound, and adding vessel noise to reef sound cut settlement about 1.29-fold. The design used one cue level and one noise level, and it interpreted the result as masking. A single noise level cannot reveal a non-monotonic curve, so that study motivates this proposal and supplies its protocol, but it cannot decide between the two accounts. One detail in its numbers matters here. Vessel noise alone raised settlement 1.21-fold, not significantly. Our design has to account for that hint that noise by itself acts as a weak trigger.

The mechanism comes from Statistical and Dynamical Interpretation of ISIH Data from Periodically Stimulated Sensory Neurons. That work showed that crayfish water-motion hair receptors and an electronic excitable-neuron model both reach peak output signal-to-noise ratio at an intermediate noise level (about 10 mV rms in the circuit). It also argued that linear noise subtraction cannot describe such systems. The work stopped at single cells in an adult, mobile animal driven by a pure sinusoid. What Is Stochastic Resonance? Definitions, Misconceptions, Debates, and Its Relevance to Biology extends the concept to aperiodic signals and to populations of threshold units, which is the right formulation for a complex reef soundscape. It also frames the debate over whether biological stochastic resonance is optimal or merely compensatory. Our cue-strength manipulation speaks to that debate directly. Wei Xu showed that the effect depends on the shape of the noise distribution when noise is heavy-tailed. That matters because real marine noise, including distant pile driving, is impulsive, and it motivates our secondary arm.

The injury literature constrains the choice of organism. Sea bass work shows a sharp jump in vulnerability across life stages. Juveniles held 45 m from monopile driving suffered 0% immediate mortality and 9% versus 10.1% control delayed mortality (In Situ Mortality Experiments with Juvenile Sea Bass). Early larvae died within 10–30 minutes above 200 dB (Pilot study on the impact of impulse noise on sea bass larvae). This fragility comes from the swim bladder, the gas-filled buoyancy organ (The impact of underwater noise on the early life stages of fish). In a fish larva, any noise effect on behavior would be confounded with gas-cavity stress. Oyster larvae have no swim bladder, and our noise levels stay far below any injury range, so behavior can be read as information processing.

The 10.1% control mortality in the sea bass field study is a second constraint. Handling and batch effects can be as large as the effect being measured. The design therefore randomizes treatments within larval batches and treats batch as a block.

From the validation literature we take a discipline rather than a technique. Calibrate a model on one set of conditions, then test its predictions on conditions it never saw, and trace the residual mismatch to physical causes such as mounting compliance or boundary effects. This is the pattern in Design and Experimental Verification of a PZT Pump with Streamlined Flow Tubes and Effect of Reinforcing Bar on Rayleigh Wave Propagation on Concrete Structures. Small tanks have their own boundary artifacts, so we measure the delivered sound field rather than trusting speaker settings.

Research Program

The work has four steps.

Delivered-field calibration. We reproduce the five-tank playback arrangement from the oyster study. We map the field each tank actually delivers, at jar positions, with a calibrated hydrophone. We also measure a particle-motion proxy, because bivalve larvae likely sense water motion rather than pressure. We report how much the sensor perturbs the field. All later analyses use the measured cue and noise levels in the cue's band, not nominal playback gains.

Calibration experiment. We measure the threshold model's parameters from conditions that do not involve both cue and noise together. One set is settlement versus cue amplitude in silence, across attenuated reef recordings. The other is settlement versus noise level with no cue. Fitting $\theta$, a gain from detection to settlement, and the number of effective units to these two curves fixes the model.

Pre-registered prediction test. We then predict, before running it, the full surface of settlement over cue level and noise level. We test it on a factorial grid. Cue level takes four values: silence, two attenuations expected to fall below threshold, and one clearly above. Noise takes six band-matched Gaussian levels. Noise type takes two values: broadband, and broadband with the cue band notched out at equal total power. Settlement is scored after 24 hours on shell substrate, as in the original protocol, with jars as replicates and larval batch as a random effect.

Secondary arm. We repeat the sub-threshold cue condition with heavy-tailed impulsive noise matched in variance. The question is whether the optimal noise level moves as Wei Xu would predict. This links the result to impulsive anthropogenic sources.

Baselines

The naive reference is the no-speaker, no-noise condition, together with the plain reef cue in silence. For context, the cited oyster study reports that reef sound produced a 1.64-fold settlement increase over no-speaker controls, and that vessel noise added to reef sound produced a 1.29-fold drop. No reported result was found on Lacuna for settlement under graded noise levels or graded cue amplitudes. There is therefore no published curve to beat, only these single-point effects. Our silence conditions should reproduce the reported direction and rough magnitude, which serves as a check that our system is working.

The strongest competitor is a linear energetic masking model. In it, settlement is a monotonic function of in-band signal-to-noise ratio, plus an additive noise-alone term fitted from the no-cue curve. It gets the same calibration data, the same number of free parameters (up to two more), and the same fitting procedure. It is evaluated on the same held-out cue-and-noise conditions.

The second competitor is an acoustic-index model. It regresses settlement on the Acoustic Complexity Index of each delivered stimulus mixture, as the oyster study's interpretation suggests, and is fitted and evaluated identically.

The threshold model counts as winning on two conditions. It must have higher held-out log-likelihood on the cue-and-noise conditions, decided by a pre-specified comparison with batch-blocked cross-validation. And the cue benefit must show a significant interior maximum in noise for the sub-threshold cues and no interior maximum for the supra-threshold cue.

The key ablation is the band-notched noise. The threshold account predicts that removing in-band noise abolishes the benefit. Any account driven by total noise power predicts no change.

Why not just…?

Why not just accept linear masking, since the oyster data already show noise reduces settlement? The existing evidence is a single noise level paired with a strong, near-field cue. In our model that is the supra-threshold regime, where both accounts predict a decline, so the published 1.29-fold drop cannot separate them. The decisive evidence is the sub-threshold cue rows of the factorial grid. Masking predicts a cue benefit that falls monotonically with noise. If that is what we see, the masking account is supported and the stochastic-resonance hypothesis fails for this system. That would be a useful result either way.

Why not just say noise is itself a mild settlement trigger or arousal signal? This is the most serious threat. The 1.21-fold settlement rise under vessel noise alone, though not significant, suggests it. An additive arousal effect would raise settlement in the cue and no-cue conditions equally, leaving the cue benefit flat across noise levels. It would also be insensitive to whether the cue band is notched out. Two outcomes would refute arousal as the sole explanation: an interaction in which the cue benefit peaks and the peak moves with cue strength, and the loss of that peak under band-notched noise. If noise raises settlement but these signatures are absent, we will report that arousal explains the data.

Why not just use acoustic indices such as the Acoustic Complexity Index, which already describe what larvae respond to? Adding noise generally changes these indices monotonically. They also contain no threshold, so they cannot produce a benefit whose optimum depends on cue strength. We fit the index model as a full competitor rather than dismissing it. If it matches the threshold model on held-out conditions, the simpler description wins.

Why not just record from larval sensory cells, as in the crayfish work? Single-cell recordings would show peripheral enhancement, which has already been demonstrated. They would not show whether the enhancement survives into a settlement decision, and that is the open question. Behavior is the endpoint that matters ecologically, and it is the one missing from the literature.

Decisive Checks

The hypothesis lives or dies on three observations from the pre-registered grid. The first is an interior maximum in cue benefit for sub-threshold cues. The second is a shift, then disappearance, of that maximum as the cue strengthens. The third is the loss of the benefit under band-notched noise at matched power. We also require the threshold model, calibrated without any combined cue-and-noise data, to predict the combined surface better out of sample than the masking and index models.

Two outcomes would sharpen the conclusions. A positive result in the impulsive-noise arm would mean distant pile driving could act as information as well as hazard for drifters near a reef's acoustic edge. A null result across the grid would show directly that the masking assumption holds for passive larvae.