Introduction

The audiogram is a graphical depiction of the frequency range of hearing and frequency-specific hearing sensitivity (hearing thresholds). It is a fundamental piece of information used to predict how underwater noise sources could affect marine mammal behavior and their hearing1. Possible effects of underwater anthropogenic noise on marine mammals include changes in hearing sensitivity such as temporary or permanent threshold shifts (e.g., refs. 2,3,4), stress responses (e.g., refs. 5,6,7), and disruptive behavioral changes (e.g., refs. 8,9,10). Data for audiograms can be acquired by both behavioral and electrophysiological methods. Behavioral methods require having an animal under stimulus control and training it to respond if it hears a sound (e.g., refs. 2,11,12,13,14,15). Data collection usually takes weeks to months, necessitating regular access to the animal. Electrophysiological methods, such as auditory evoked potentials, are an alternative and can be applied to temporarily captured or stranded restrained small whales16,17,18,19,20,21. However, hearing thresholds are commonly overestimated with this latter approach, as it loses effectiveness at the lowest frequencies of hearing. In addition, its application is limited by the size of the whale, i.e., as the brain to body mass ratio becomes smaller, electrophysiological signals become more difficult to record.

Audiograms are available for multiple species of odontocetes (toothed whales), but there are no empirically measured audiograms for any mysticete (baleen whale). Instead, mysticete hearing range and sensitivity are inferred using a combination of anatomical modeling (e.g., refs. 22,23,24,25), vocalization frequencies (for example, 20–24 kHz for the humpback whale, Megaptera novaeangliae26,27), and the assumption that if they behaviorally respond to a noise, they must have heard it19,28,29,30. Houser et al.23 used basilar membrane measurements from humpback whales to develop a band-pass filter model of auditory sensitivity. The model suggested peak hearing sensitivity between 2 and 6 kHz, with “good” hearing sensitivity from 700 Hz–10 kHz. Tubelli et al.24 later used finite-element modeling to create middle ear transfer functions for the humpback whale. Based on the location of acoustic stimulation along the whale’s head, two models were created—a tympanic membrane (TM) model and a tympanic bone (TB) model. The TM model suggested “best” hearing sensitivity at 1–3 kHz, with “good” sensitivity (to -40 dB) from 15 Hz–3 kHz. The TB model suggested greatest sensitivity at 2–3 kHz, with “good” sensitivity from 200 Hz–9 kHz. Collectively, the models suggest humpback whales are most sensitive between 1 and 6 kHz, with “good” hearing sensitivity up to ~10 kHz but not beyond. Though useful for predictions, the humpback whale models require validation to provide confidence in their predictive capability. Unfortunately, there is little potential for (trained) behavioral or electrophysiological approaches to be used to provide validation data because the size of humpback whales makes it difficult to maintain them under human care.

Behavioral observation audiometry (BOA) is a less precise alternative to classical behavioral audiometry. This method has historically been used in infants or small children incapable of fulfilling or comprehending instructions31. It relies on the observation of behavioral responses synchronized with the transmission of an acoustic stimulus. The responses are generally reflexive behaviors, such as flinching or orienting toward the sound source. Practitioners of BOA seek to find the “minimum response level” (MRL), which indicates the lowest level of sound to which a subject is responsive32,33,34. The MRL is not equivalent to the hearing threshold obtained through behavioral audiometry. It can be affected by the state of attention, mood, and prior experience with a signal, as well as the background noise level since these measures are rarely conducted in anechoic rooms. However, it offers a useful estimate of the upper bounds of hearing sensitivity when other options are not available. BOA was first used in marine mammals in Atlantic Walruses (Odobenus rosmarus rosmarus35. Applying this concept to baleen whales, it is possible to carry out a BOA experiment in the open ocean if animals reliably respond to a signal. Although ocean noise may be an issue, in that the detection of some frequencies will be masked by environmental noise, by making some assumptions about the ability of whales to detect signals in noise (e.g., by the application of known critical ratios for other marine mammal species), a function could be developed to estimate the humpbacks’ relative sensitivity to different frequencies of noise. This could yield an audiogram that regulators and stakeholders could use.

The critical ratio (CR) is used to obtain a first order estimate of the ability of an animal to detect a signal in noise36. It is defined as the difference between the sound level of a tone that is just detectable (dB re. 1 µPa), and the power spectral density level of background noise (dB re. 1 µPa2/Hz), across a “critical band” of frequencies masking the tone. Simply put, it is a measure of how much louder a signal must be than the ambient noise to be audible. Within captive marine mammals, CRs have been measured in several species, including the harbor porpoise (Phocoena phocoena37), bottlenose dolphin (Tursiops truncatus38,39), and killer whales (Orcins orcas40). The measured CRs are surprisingly consistent across species despite obvious differences in hearing range, sensitivity, and ear morphology. Generally, within odontocetes and pinnipeds, CRs tend to increase with increasing signal frequency, from ~17 dB at 100 Hz, to 23 dB at 10 kHz41,42,43. Since humpback whales reliably respond to frequency-modulated tones28, naval sonar signals44, and oil and gas exploration seismic air guns45, it seems reasonable to assume that a modified BOA approach can also be applied to humpback whales. Testing across a broad range of frequencies in noise should lead to a frequency- and noise-dependent MRL curve for this species.

The goal of the current study was to use frequency-modulated signals (up-sweeps), and a behavioral response paradigm, to determine the MRL of migrating humpback whales in relatively quiet (aside from singing whales) natural ocean noise, which occurs in the coastal waters east of Australia. Three of the frequencies used were predicted to be audible based on modeling and other less-controlled hearing studies (0.25, 1, and 4 kHz). A fourth, higher frequency (16 kHz), was hypothesized to be less audible based on prior modeling work23,24. The whales were exposed to low-level upsweeps and the received level and signal-to-noise ratio (SNR) of the upsweeps were estimated at the point where a behavioral change was determined statistically. The MRLs from the sample population were subsequently used to make inferences about frequency-dependent masked hearing in humpback whales.

Results

The behavioral change-point of groups

Upsweep stimulus exposure trials (n = 39) were designed to expose a group of approaching whales to tonal upsweeps with start frequencies of 0.25, 1, 4, or 16 kHz (one frequency per trial; see Supplementary Method 1 for further details). Each experiment tested one frequency, with at least eight trials per experiment (frequency), noting each trial exposed a unique whale group (Fig. 1). At some distance from the source, groups exposed to a sound source typically responded by slowing down or speeding up and turning away to avoid the source vessel (behavioral change-point), sometimes followed by heading directly towards the vessel (usually circling or staying close to the vessel). Control trials (n = 6) used a ‘silent’ source. Baseline behavioral data were collected on other groups when the source vessel was not present (n = 9).

Fig. 1: An example of a 0.25 kHz stimulus exposure trial.
figure 1

As the whale group (illustrated by the whale icon) traveled southwards, the sound source vessel (vessel icon) moved into position and anchored 5.7 km south of the group’s trajectory (11:36). When the group was 4.5 km from the source vessel, transmission of the upsweeps started (11:59). The group immediately changed its dive behavior by decreasing time spent on the surface, then altered its course by turning southeast (at ~3.8 km from the source). The group continued to travel on a southeastern course, displaying short surface intervals, until it passed by the source vessel. The color bar inlay corresponds to the times that the locations of the whales were recorded.

Full size image

The behavioral change-point for each group was statistically determined using measurements related to the group’s movement (e.g., course, speed, and dive distance from focal data, and course, speed, fluking and rolling behavior from tag data) and dive behavior (e.g., dive and surface times). These multivariate parameters were collapsed into two univariate time series using a Mahalanobis distance (MD) analysis to create MD scores (see Supplementary Method 2 for further details). A score of >1 indicated a significant change in behavior in the during phase compared to the before phase. Groups within the during phase of exposed trials had greater MD scores compared to groups within control and baseline trials (see Supplementary Materials for further details) indicating groups responded to the sound source and not the anchored vessel.

As intended by the experimental design, most groups did not respond at signal onset meaning the signal was assumed to be inaudible at the start of exposure. The behavioral change-point, i.e., the point at which the group first responded to the now audible signal, was defined using MD scores. Given MD scores within control and baseline groups could be >1, conservatively, the change-point for movement was taken as the first instance in the during phase where the MD score was >1.4 (i.e., the magnitude of behavioral change was above any changes within control and baseline groups; Fig. 2 and Supplementary Materials). A significant change in dive behavior was assumed to occur when the dive MD score was >2.5 for the same reason (Fig. 2 and Supplementary Method 2).

Fig. 2: Examples of the Mahalanobis distance (MD) scores for dive and movement behavior comprised from focal follow and tag data from four trials.
figure 2

Each graph shows the Mahalanobis distance (MD) score from focal follow data (upper four) and tagged individual movement and dive behavior (lower four) from four trials (a 0.25 kHz stimulus exposure trial, which is outlined in Figs. 1, and 1, 4, and 16 kHz stimulus exposure trials). Blue and red horizontal dashed lines show the level at which a significant change in behavior was assumed, i.e., an MD score that exceeded the dashed line. Graphs also indicate the onset of the upsweeps (vertical green dashed lines).

Full size image

For a subset of whales, suction-cupped tag data provided high-resolution movement data (3D pitch, roll, heading, body acceleration, and dive depth), potential startle responses, fluking patterns and swim speed using a suite of tag sensors (see Supplementary Method 2 and Supplementary Fig. 1). Within the tag data (n = 4 exposed groups consisting of one tagged group per experiment), as with the focal follow data, MD scores relative to the before phase were created. In all four instances, the time of the movement change-point detected by the land data closely matched the time change that was detected in the tag data (Table 1). To summarize, two independent data collection platforms produced similar times for behavioral change-points.

Table 1 The time of the dive and directional movement change-points measured using the land observation and tag datasets from the four tagged animal trials showing that the independent data platforms detected the behavioral change-points at similar times
Full size table

Frequency-dependent response thresholds

No significant change in behavior was detected within the during phase of any control or baseline trial as estimated using standardized MD scores (Supplementary Method 2). By contrast, a movement and dive change-point occurred within the during phase for 37 of the 39 groups exposed to 0.25 kHz (n = 9), 1 kHz (n = 8), 4 kHz (n = 10), and 16 kHz (n = 10) upsweeps. One group did not respond to the 250 Hz signal, and one group did not respond to the 4 kHz signal, though both groups came within 400 m of the source vessel and would have been exposed to clearly audible signals.

Signal received levels, noise levels (one-third octave in dB re 1 µPa and corrected to power spectral density in PSD; dB re 1 µPa2/Hz), and SNRs were measured at the behavioral change-point (first response) of groups for each exposure trial. At 0.25 kHz and 1 kHz, noise was dominated by singing whales, and at 4 kHz and 16 kHz, noise levels varied due to patchy occurrences of snapping shrimp. Results are summarized in Table 2.

Table 2 Summary of received levels when responses were observed, ambient noise measurements (measured as one-third octave and power spectral density), and mean resultant signal-to-noise ratios (SNR using one-third octave, SNRTO, and power spectral density, SNRPSD, noise) when responses occurred for each test frequency
Full size table

The average of the lowest two received levels and two lowest SNRPSDs observed across the sample population was used to estimate the MRL for each frequency. The received level at each behavioral change-point, and the associated MRLs, follow the noise profile of the study area, with a decline in response received levels with increasing frequency (Fig. 3). This is because the upsweeps were likely masked. Response SNRPSDs and associated MRLs (Fig. 4) were close to what would have been the just masked hearing threshold in odontocetes in this background noise when incorporating the known critical ratios of odontocetes42, suggesting MRLs measured here were close to likely masked detection thresholds for humpback whales.

Fig. 3: The received level at which each group behaviorally responded (red circles) across the tested frequencies, the minimum response level for each frequency (blue crosses and line), and median and 10th percentile background noise level for the study site (black lines).
figure 3

Response received levels (red circles) for each trial are shown along with the minimum response level (MRL; dB re 1 µPa), being the mean of the lowest two received levels for each frequency (blue crosses and blue line) are plotted across frequencies. The grand median noise PSD (dB re 1 µPa2/Hz) calculated across all trials (thick solid black line) is shown as a representation of the noise spectrum in the study area, as well as the noise 10th percentile (thin dotted line) for reference.

Full size image
Fig. 4: The received signal-to-noise (SNR) at which each group behaviorally responded (red circles) across the tested frequencies, the SNR minimum response level (blue crosses and line), and odontocete critical ratios (black diamonds and black line).
figure 4

Frequency-dependent response signal-to-noise ratios (SNRPSD, red circles) at which groups behaviorally responded and the associated MRL curve (blue line) calculated across all trials, taken as a mean of the lowest two SNRPSDs for each frequency. Also shown is the power equation fit to the aggregate data for odontocete critical ratios (taken from 42; black line).

Full size image

Discussion

Behavioral observation audiometry (BOA) is a method used to test the hearing of human infants and persons incapable of responding to verbal instruction46. It correlates behavioral reactions of an individual to acoustic broadcasts of varying stimulus levels to obtain a MRL to the stimulus. Historically it was recommended that three or more behavioral response threshold sessions be run, and that the two lowest stimulus levels that correlate with an observed response are averaged to determine the MRL. Here, we used a modified BOA approach to determine humpback whale MRLs to one-third octave frequency sweeps presented at four frequencies ranging from 0.25 to 16 kHz. These frequencies ranged from the lower end of the humpback whale’s presumed hearing range (0.25 kHz), through the frequency region of estimated peak hearing sensitivity (1 and 4 kHz), to a frequency likely approaching or surpassing the upper end of their hearing range (16 kHz). While responses at 0.25–4 kHz were not surprising, responses at 16 kHz provide the first direct evidence of hearing in humpback whales at a relatively high frequency. Since minke whales (Balaenoptera acutorostrata) may have ultrasonic hearing47, it appears that the hearing range of baleen whales extends to much higher frequencies than previously thought. This has large implications for the management of anthropogenic ocean noise since baleen whales were thought of as low-frequency hearing specialists. In other words, mitigation measures based on the assumed high-frequency roll-off of hearing may have to be revised since it seems baleen whales (at least the minke and humpback whale) are not as insensitive to high frequencies as believed.

The use of BOA in infants has largely been replaced by visual reinforcement audiometry and electrophysiological audiometry (e.g., auditory brainstem responses). The BOA method has been criticized because of the rapid habituation of subjects to repeated exposures, the potential for observer bias, and a large variability in responses48,49. The cross-sectional approach used in this study resolves issues of habituation and observer bias, as only one trial per subject group was carried out. Since animals were migrating southwards, there was a negligible chance of repeatedly exposing the same animals. In addition, an objective response metric for behavioral change was used, where the behavior of subject groups before the stimulus presentation was statistically compared to behavior during the stimulus presentation using two independent data collection platforms. Though response variability remains an issue, this was addressed by testing multiple subject groups to capture the variability across the test population. Since only one stimulus presentation could be made for each migratory group of humpback whales, the approach to determining the MRL was modified—the average of the lowest two received levels associated with a response and observed across the sample population was used to calculate the MRL for each frequency. It should be noted, however, that the BOA worked well in this study because of the high sample size and predictable behavior of the migrating humpback whale groups. In addition, these results are only valid for humpback whales but, if possible, a similar approach could be used in other species.

The SNRPSDs at MRLs estimated here were close to predicted critical ratios (CRs) measured in odontocetes, phocid seals, and otariid seals, with all species following similar patterns41. Assuming similar critical ratios exists within the mysticetes, the MRLs presented here are potentially close to masked detection thresholds for the humpback whale. This suggests that the frequencies used for these tests were within their peak sensitivity range of hearing, and not at the upper end of their hearing curve, where hearing is likely to be unmasked by the ambient noise. However, although the SNRs corresponding to the MRLs are comparable to CRs measured in other marine mammals, the present study did not meet the conditions for measuring CRs. First, the subjects of CR studies are trained to detect and report the tones and have an expectation of what to listen for37,38,40. The present study relied on a natural, untrained (avoidance) response, to the stimulus. Second, CRs are determined by measuring the SNR for a tone at its detection threshold in broadband Gaussian background noise of constant spectral density. By contrast, in the present study, even though the noise PSD varied by only a few dB across the relevant noise bands, the variation was due to uncontrolled biotic and abiotic sources.

In part, this explains the variation in response received levels observed. Some noise sources were continuous (e.g., wind dependent noise, traffic noise), while others were impulsive or fluctuated temporally and spatially during the experiments (e.g., humpback whale song, snapping shrimp). Since the characteristics of natural/anthropogenic noise affect whether masking is increased or released (e.g., via dip listening, comodulated masking release50), it is likely that the SNRs measured here are also affected by the type of noise the whales experienced in the environment. For example, if we assumed that the whales first detected and responded to signals when background noise was temporarily low, as might occur during dip listening in between song chorusing, the 10th percentile noise PSD curve could be used as the reference to calculate the SNR at signal detection. This would make no difference at higher frequencies, given the 10th percentile and median noise levels were similar. However, at song-dominated frequencies, where song levels can temporarily fluctuate, using the 10th percentile noise PSD would increase the response SNR by 7 dB at 0.25 kHz and by 3 dB at 1 kHz. The degree to which the actual CRs and the SNRs measured here differ is unknown, but Branstetter et al.51 found a 22 dB range in masked detection thresholds (CRs) of dolphins exposed to different types of natural, synthetic, and anthropogenic noise. It should be noted that the CRs of odontocetes found in the literature also shows variation between subjects and studies. Though this range might provide a reasonable upper bound by which humpback CRs and the SNRs measured here differ, the differences will likely also vary as the frequency of the signal and noise increase as the dominant noise source will change with increasing frequency (e.g., humpback song vs. snapping shrimp). Regardless, these data likely provide the upper bounds of hearing sensitivity in noise, meaning, if applied to policy development on ocean noise mitigation, a more conservative measure of hearing sensitivity should be considered.

To summarize, the MRL curve derived with BOA in the present study is a proxy for the shape of the humpback whale’s masked audiogram, which could not have been achieved in baleen whales using classical behavioral audiometry. The curve makes no assumptions about absolute hearing sensitivity but suggests that hearing sensitivity extends at least to 16 kHz. Since hearing at 16 kHz was masked, it is likely the upper frequency limit of hearing for humpback whales could well be above this frequency, and thus higher than estimated from anatomical models. Extending the MRL curve to encompass lower and higher frequencies than those tested here is an obvious next step toward understanding the full frequency range of hearing in the humpback whale.

Methods

Behavioral data collection

Trials were conducted over three field seasons (2021, 2022, and 2023) at Peregian Beach, 100 km north of Brisbane, Queensland, Australia, each year during September and October. Humpback whales were migrating from their breeding grounds inside the Great Barrier Reef, to feeding grounds in the Southern Ocean. Given this field site was positioned on a near-shore migratory corridor for a large population of humpback whales (>30,000), the predictable movement of the whales allowed behavioral responses to be detected reliably52 with negligible risk of repeating trials on the same group. Ethical approval was granted by the University of Queensland Animal Ethics Committee. All relevant ethical regulations for animal use were complied with.

The sample unit was a group of whales, defined as those whales surfacing synchronously and within 100 m of each other, or a tagged individual within the group. Teams of land-based and boat-based observers recorded every surface behavior and location as the group moved southwards through the study area (see Supplementary Method 1) to produce a continuous time-series. Observations were entered in real-time into networked tracking software (SeaScape, Eric Kniest, affiliated with University of Queensland). Mapped whale tracks were available, in real-time, to all networked computers (i.e., on all land- and boat-based computers). Each trial was divided into three phases; a tagging phase, before phase, and during phase.

“Tagging” phase

For some groups, an attempt was made to tag one of the whales in the group using a suction-cup acoustic tag (Acousonde 3B, Acoustimetrics, CA, USA), if sea conditions were suitable (sea state ≤3, wind speed <12 knots). During a tagging attempt, the tagging vessel (Pelagia, 5.4 m rigid-hulled inflatable vessel with single 100 hp 4-stroke outboard motor) accelerated to within ~5 m of the group, and, if successful, the tag was deployed from the end of a 6 m carbon-fiber pole. Once tagging attempts had ceased (and regardless of whether a tag was deployed), the group was given ~10 min to “recover” to what looked like pre-tagging behavior before starting the experiment53.

Each experiment tested one frequency, with at least eight trials per experiment (frequency), and with each trial exposing a unique whale group. During each trial, the migrating whales moved towards an anchored ‘source vessel’ with a deployed transducer placed directly in the path of the group. The source level and the distance to each group at the start of exposure were set so that the whales were unlikely to audibly detect the signal initially but would swim towards the source until the signal was detectable (Fig. 1). Control trials (n = 6) were also carried out where the transducer was deployed from the anchored source vessel but not operated to determine if groups responded to the experimental set-up rather than the sound source. Additional baseline behavioral data were collected on other groups when the source vessel was not present (n = 9) to quantify natural migratory behavior.

“Before” phase

The trial started when the focal follow began. During this period, the sound source vessel (Carmena, a 5.6 m center console aluminum boat) was positioned for the experiment ~5–6 km ahead of the path of the group. The position of the source vessel was recorded every 0.5 to 3 minutes using network-linked GPS, so its location was available to the land-based team and focal-follow vessel in real-time. As the whales were in the southward migration phase, their movement behavior was predictable—each group tended to travel at a consistent speed, course, and distance from shore. The path of the focal group could therefore be predicted with reasonable accuracy. This allowed the focal-follow team (on Pelagia) to direct the sound source vessel to anchor ~5–6 km ahead of, and in line with, the predicted path of the focal group. Once anchored, the sound source vessel deployed an underwater transducer and was ready for signal transmission (Fig. 1). In control trials the underwater transducer was deployed but no signal was transmitted. The minimum time for the before phase was 20 min (~4 group dive cycles) and the maximum was 90 min, depending on the speed with which the focal group approached the source vessel. For baseline trials, the before phase was the first 60 min of the focal follow.

“During” phase

For baseline groups, the during phase started approximately one hour after the focal follow began (average length of the before phase for stimulus exposure and control groups). For control and stimulus exposure groups, the start of the during phase was determined by the team on Pelagia, who radioed the sound source vessel to start the stimulus transmission once the group came within ~4.5 km. After the start of the transmission (exposure trials) or start of the during phase (control trials), the focal-follow vessel and land-based observers continued to follow and observe the group until the group passed the sound source vessel at closest approach (Fig. 1), which marked the end of the trial. The during phase for baseline groups lasted approximately 40 minutes, to match the length of a typical stimulus exposure and control trial.

Acoustic stimuli consisted of a one-third octave linear upsweep that was 2 s in duration, repeated every 4 seconds (2 s on, 2 s off duty cycle; see Supplementary Method 3). Land-based observers were blind to the treatment (silent control or stimulus transmission). Using the results of in situ test transmissions, source levels for each stimulus were tailored so that the SNR at 4 km was below the anticipated masked hearing threshold for the stimulus frequency band. This was because the goal of the experiment was to measure when whales first responded to the signal (i.e., a response threshold) rather than respond at signal onset. The anticipated response threshold was based on measured critical ratios of odontocetes, pinnipeds and sirenians43, as well as the responses of humpback whales to a 2 kHz upsweep in a previous experiment at this site28.

Behavioral and movement data consolidation and analysis

Focal data provided group-based data on movement (course and speed) and dive behavior. For a subset of whales, tag data provided high-resolution movement (3D pitch, roll, heading, body acceleration, and dive depth), potential startle responses, fluking patterns, and swim speed (see Supplementary Method 2). Each trial dataset (group-based or tagged-based) was considered as a multivariate time-series of dive and movement measures and each trial potentially included four different analyses: the first related to group-level movement behavior (using course, speed, and dive distance), the second related to group-level dive behavior (using dive time and surface interval), the third related to individual (tagged) dive responses (dive time, surface interval, ascent and descent speed), and the fourth related to individual (tagged) movement responses (fluking, rolling, heading, OBDA, jerk, and speed). All measures were estimated per dive cycle.

Five measurements related to the group’s movement and dive behavior for each dive cycle were extracted from the land-based data. Group course (1) and speed (2) were measured from the end of one surface interval to the start of the next surface interval using theodolite positions. Dive distance (3) was estimated using the last position of the group before a ‘long’ dive and the first position of the group when re-sighted at the start of the next surface interval assuming a linear swimming track. The durations of both long dives (4) and surface intervals (5) (i.e., the time the first group member reappeared after a dive and the time the last member of the group disappeared before a deep dive) were also measured. For individual tagged whales, movement measurements included speed (averaged per minute), the number of flukes (summed per minute), and the mean roll and heading, and the maximum OBDA and jerk values for each minute (see Supplementary Fig. 1). Dive measures included dive and surface interval time and descent and ascent speed.

Multivariate analysis of whale behavior followed DeRuiter et al.54 and Stimpert et al.55 for all whales (tagged and non-tagged) and is outlined in further detail in the Supplementary Method 2. A Mahalanobis distance analysis was used to estimate a behavioral change-point for each group. Briefly, the mean, variance, and covariance measures in the before phase were used as the measure of distribution. The preliminary behavioral change-point for each group was allocated when the Mahalanobis distance (MD) of measures within the during phase increased above the before phase 95th percentile. Since the MD values across all groups were highly variable and dependent on the group’s underlying behavioral variability, standardized MD scores were created by dividing each MD value by the 95th percentile. A value of >1 indicated a change in behavior relative to the before period for that group.

Statistics and reproducibility

To determine if groups behaviorally responded to the upsweep signals, and not during control or baseline trials, MD scores within the during phase of baseline, control and upsweep exposure trials were statistically compared (see Supplementary Method 2). To conservatively assign a change-point, the highest group MD score for movement within the during phase of baseline groups and control (1.4) was used as the threshold for a significant behavioral response. This was, for example, a significant deviation in course and/or a significant change in travel speed. Dive behavior MD scores were more variable, therefore, a significant change in dive behavior was assumed to occur when the dive MD score was greater than 2.5. This was, for example, a dive which was substantially shorter and/or an extended surface interval within the during phase. Changes in movement and dive behavior of these magnitudes did not occur in control and baseline groups.

The time of the change-point for each trial was noted with the same procedure carried out for the tag data. For tag data, a change-point was usually related to a substantially shorter dive and/or increased surface interval, an increase in rolling behavior, a temporary cessation of tail fluking combined with a reduction in speed and/or a temporary change in heading. If a change-point was not detected, it was assumed the individual or group did not respond.

Signal received levels and noise levels were then measured at the behavioral change-point (first response) of groups for each exposure trial. From these values, the SNR at behavioral change-points were estimated using median sound pressure levels (SPL) of noise at the time and site of each trial. Noise was measured over the one-third octave, where the signal level was highest (dB re 1 µPa) and corrected to power spectral density (PSD; dB re 1 µPa2/Hz) to provide two SNRs at the point of response for each trial. The average of the lowest two received levels and two lowest SNRPSDs observed across the sample population was then used to estimate the MRL for each frequency. This follows the recommended method for estimating the MRL from repetitive trials performed in the same subject56 but modified here for the cross-sectional design of the experiment (i.e., single exposures across multiple subject groups). The received levels at each behavioral change-point (Fig. 3), the response SNRPSDs (Fig. 4), and the associated MRLs (Fig. 3 and Fig. 4), were then plotted against the grand median of median noise values measured across all trials to produce a frequency-dependent MRL curve (Fig. 3 and Fig. 4).

Acoustic recordings and analysis

During each trial, a series of sound recordings was made by the focal vessel (Pelagia). Recordings started where a putative change in behavior was believed to have occurred during the focal follow (to be confirmed by the later change-point analysis as described above), as well as between the putative change-point position and the source vessel (see Supplementary Method 4 and 5 for details). A continuous sound recording also occurred at the sound source vessel to provide a record of transmission and a source level measurement.

Analysis of the acoustic recordings were to (i) estimate the received level of the upsweep where the change in behavior occurred, and (ii) measure the noise levels for the same location. In many trials, upsweep signals within the recordings were masked by noise. Therefore, received levels for each trial were estimated by fitting upsweep received levels from the various recordings closer to the source to a simple log-linear regression model and assuming spherical spreading to 20 m from the source. Median noise was measured in the relevant one-third octave band from the recordings closest to the behavioral change-point since this noise was the best representation of noise at the group at the time it responded. These levels were then used to estimate the median noise power spectral density (PSD) level by subtracting the appropriate band-width correction for the relevant one-third octave. Noise levels were subtracted from the estimated received level to provide a SNRTO and SNRPSD for each trial.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.