Sound-recording acoustic tags attached to marine animals are commonly used in behavioural studies. Measuring ambient noise is of interest to efforts to understand responses of marine mammals to anthropogenic underwater sound, or to assess their communication space. Noise of water flowing around the tag reflects the speed of the animal, but hinders ambient noise measurement. Here, we describe a correlation-based method for stereo acoustic tags to separate the relative contributions of flow and ambient noise. The uncorrelated part of the noise measured in digital acoustic recording tag (DTAG) recordings related well to swim speed of a humpback whale (Megaptera novaeangliae), thus providing a robust measure of flow noise over a wide frequency bandwidth. By removing measurements affected by flow noise, consistent ambient noise estimates were made for two killer whales (Orcinus orca) with DTAGs attached simultaneously. The method is applicable to any multi-channel acoustic tag, enabling application to a wide range of marine species.
Sound-recording acoustic tags are commonly employed to study movement and acoustic behaviour of marine mammals (Marshall, 1998; Johnson and Tyack, 2003; Akamatsu et al., 2005; Goldbogen et al., 2006; Johnson et al., 2009), and are a key instrument in behavioural studies aimed at understanding the impact of anthropogenic sound (e.g. Tyack et al., 2011; Miller et al., 2012; Dunlop et al., 2013; Goldbogen et al., 2013). Acoustic tags are equipped with hydrophones to record vocalizations of the tagged and surrounding animals and other sound types such as exposure signals, e.g. sonar sounds or other control sounds (e.g. Tyack et al., 2011; Curé et al., 2012), or ambient noise. Ambient noise levels are of interest to efforts to understand the responsiveness of marine mammals to anthropogenic sound such as sonar signals or shipping noise (Ellison et al., 2012; Dunlop et al., 2013), or to estimate the communication space available to them (Miller, 2006; Clark et al., 2009). However, movement of the animal generates flow noise around the hydrophone (Haddle and Skudrzyk, 1969; Goldbogen et al., 2006), potentially masking other sounds at lower frequencies and limiting the capability to reliably measure ambient noise. Flow noise can also be exploited to estimate the animal's speed through the water (Goldbogen et al., 2006). Speed estimates can be used to obtain more accurate underwater movement paths via track-reconstruction methods (Wensveen et al., 2015), and are also used to indicate feeding attempts (lunges) of rorqual whales (Goldbogen et al., 2006; Doksæter Sivle et al., 2015).
Because of the turbulent nature of water flow around the receiving hydrophones, pressure fluctuations generated by the flow are uncorrelated between two hydrophones for frequencies higher than f>(U/D), with D being the spacing of the two hydrophones and U the hydrophone speed through the water (Corcos, 1967). In contrast, ambient noise is correlated for , with cs being sound speed in water (Cox, 1973). The relative contributions of flow and ambient noise can thus be estimated by calculating the coherence of the sound field measured on closely spaced hydrophones (e.g. Beerens et al., 1999; Barclay and Buckingham, 2013). This approach enables reliable measurements of ambient noise levels by removing data affected by flow noise, and provides a direct measurement of flow noise over a larger bandwidth than using a single hydrophone.
Here, we apply a correlation-based method to estimate the flow and ambient noise contributions in a recording made with acoustic version-2 digital sound recording tags (DTAGs) deployed on a humpback whale and two killer whales. DTAGs are commonly equipped with two hydrophones separated by 2.5 cm (Johnson et al., 2009), which theoretically can provide estimates of the contributions from flow and ambient noise for a frequency range between ∼0.1 and ∼6 kHz for an animal swim speed of 2.5 m s−1.
MATERIALS AND METHODS
The acoustic tag used in this study was the version-2 DTAG; a suction-cup tag equipped with one or two hydrophones, depth and acceleration sensors (Johnson and Tyack, 2003; Johnson et al., 2009). The DTAGs sampled audio at 96 kHz (mn12_180a) or 192 kHz (oo09_144a/b) with 16-bit resolution, and had a flat frequency response with a 400 Hz one-pole high-pass filter. The acoustic sensitivity of the hydrophones, determined from calibration measurements, was −189±3 dB re. 1 µPa−1 (mean±s.d., N=3 tags) (P. J. Wensveen, Detecting, assessing, and mitigating the effects of naval sonar on cetaceans, Appendix III, PhD thesis, University of St Andrews, St Andrews, UK, 2016). These DTAGs were attached to a humpback whale and two killer whales during studies to investigate behavioural responses of marine mammals to sonar sounds (Miller et al., 2012; Doksæter Sivle et al., 2015) (Table S1). Animal experiments were carried out under permits issued by the Norwegian Animal Research Authority (permits S-2007/61201 and S-2011/38782), in compliance with ethical use of animals in experimentation. The research protocol was approved by the University of St Andrews Animal Welfare and Ethics Committee and the WHOI Institutional Animal Care and Use Committee.
Speed-through-water estimates of the humpback whale were derived from depth rate for periods when the animal was swimming at high absolute pitch (≥70 deg). Pitch angles were smoothed with a 5-s moving average filter to suppress effects of fluking motion.
Flow noise and ambient noise estimation
with bandwidth w=200 Hz, l=1, 2 and k=1, 2 the hydrophone indices and fc the center frequency. Here, X(f) is proportional to the Fourier transform of the sound pressure, and X*(f) its complex conjugate, with |X2(f)| equal to the mean-square sound pressure spectral density (International Organization for Standardization, 2015), measured in a time window of 0.17 s.
where pref=1 μm. The SPLs measured on a DTAG attached to a humpback whale were analyzed in the 1–2 kHz frequency band (the same band as the sonar transmissions in the controlled exposure experiments) for an 8 h period prior to the first sonar transmission. Subsets of noise measurements were created for times during which: (1) SPLuncorr<SPLcorr, and (2) SPLcorr≥SPLuncorr+6 dB (Fig. 1). The first subset was representative of flow noise even though correlated noise levels were higher, whereas the latter was unlikely to be affected by flow noise and was therefore representative of ambient noise. A 6 dB margin was adopted to account for uncertainties in the coherence measurement that are due to the short integration time [determined by measuring C(f) for artificial time series consisting of coherent and incoherent Gaussian noise; Fig. S1]. The SPL measured from the humpback whale tag was filtered with a 5-s running average before comparing it with animal swim speed.
The two tagged killer whales were in close proximity of one another, allowing for cross-comparison of the estimated ambient noise. A 30-min period prior to the first sonar transmission was considered. Noise levels were expressed as SPLs, measured in 1/3-octave (decidecade; International Organization for Standardization, 2015) bands centred at 1 and 2 kHz. The ambient noise distribution for the killer whale tags was computed using data points at times where SPLcorr≥SPLuncorr+6 dB. We defined ambient noise as the contribution of all sound, except acoustic self-noise and transient sounds emitted by marine mammals and sonar. Thus, environmentally driven sound sources, such as breaking waves and rain, as well as continuous anthropogenic sound sources, such as shipping noise, were considered to contribute to the ambient noise. The correlated sound could contain contributions of transient sounds that were not considered as part of the ambient noise (Fig. 2). For instance, the 1–2 kHz band overlapped with the frequency range in which killer whales vocalize (Ford, 1989; Miller, 2006), and the sound of bubbles released during surfacing events. The start and end times for these sounds were manually selected for both tags. Noise segments that overlapped such transient signals were removed from the ambient noise analysis.
In all cases, data points for which the animal depth was less than 2 m were removed to avoid surface splashes affecting the measurements.
RESULTS AND DISCUSSION
Analysis of the humpback whale tag data showed that the uncorrelated part of the sound field, SPLuncorr, at frequencies of 1–2 kHz was strongly related to swim speed (R2=0.87; Fig. 1). This was similar to the correlation of the total SPL (R2=0.86) in the 0.01–0.5 kHz band commonly used to derive swim speed from flow noise (e.g. Goldbogen et al., 2013) (Fig. 1). No clear relationship with swim speed (R2=0.17) was found for data points where the correlated part of the sound field, SPLcorr, exceeded the uncorrelated part by more than 6 dB for the 1–2 kHz band. Those measurements were not affected by flow noise, and thus provided a measure of the ambient noise at the animal's location.
For the killer whale tags, high levels of flow noise were found to be more commonly present near 1 kHz than 2 kHz (Fig. 3, Table S2). A striking difference in flow noise levels was observed between the two tags. At 2 kHz, tag oo09_144b showed higher contributions of flow noise than tag oo09_144a, which was caused by a different location and orientation of the DTAG on the animal. Tag oo09_144a was positioned near the dorsal fin, but tag oo09_144b slid to the lower left side of the body soon after it was deployed (Fig. S2). The mean flow noise level was 10–19 dB higher than the mean ambient noise level (depending on the frequency and tag; Table S2). The total noise level measured on the DTAG therefore was not always a reliable measure of the ambient noise levels for these frequencies.
The ambient noise 1/3-octave band SPLs at 1 kHz ranged between 90 and 118 dB re. 1 µPa2 (mean=96.3 dB re. 1 µPa2, s.d.=3.4 dB), and at 2 kHz between 83 and 118 dB re. 1 µPa2 (mean=96.7 dB re. 1 µPa2, s.d.=5.7 dB). The ambient noise level was lower at the surface than at greater depths for both tagged animals. A clear increase in SPLcorr was observed as one animal (oo09_144a) dove between 20 and 60 m. Those maximum levels coincided with the depth at which the minimum sound speed was observed using CTD (SAIV SD200) measurements obtained in the same area (Fig. 3).
Wind-generated ambient noise is expected to be fairly constant with depth for these frequencies (Ainslie, 2010). The expected wind-generated 1/3-octave band SPLs at 1 and 2 kHz, for the conditions during the experiment (sea state 1; Miller et al., 2012), are 74 and 82 dB re. 1 µPa2, respectively (Wenz, 1962). The mean ambient noise levels measured on the tags were approximately 15 to 22 dB above those predictions. Two vessels were known to be nearby during this period: MS Strønstad, at a distance between 40 and 300 m from the animals, and RV H. U. Sverdrup II, between 8 and 11 km distance. Locations of other vessels commonly present in the area were not measured. The lower levels near the surface and increase in noise as the animal passed through the acoustic channel indicated that ships were likely the dominant sound sources.
The ambient noise levels measured during deep dives (>20 m depth) were consistent between the two tagged killer whales, with mean values agreeing to within 2 dB. For shallow dives, the mean SPLcorr on tag oo09_144b was systematically lower by 3.8–7.1 dB than that on tag oo09_144a. Both animals showed little rolling behaviour during shallow dives. Because of the low placement of tag oo09_144b on the animal, body shielding of the ship noise could have led to consistently lower measured ambient noise levels (P. J. Wensveen, The effects of sound propagation and avoidance behaviour on naval sonar levels received by cetaceans, MPhil thesis, University of St Andrews, St Andrews, UK, 2012). Body shielding likely occurred occasionally for both animals during deeper dives where more rolling behaviour was observed, leading to more similar measured ambient noise levels during deep dives. These measured ambient noise levels are useful in the context of behavioural responses of killer whales to sonar, as they may be used to assess whether faint sonar signals were audible to the whales.
Our analysis demonstrated that estimated flow noise was strongly related to animal swimming speed even when the ambient noise levels exceeded the flow noise levels (Fig. 1). This confirmed that the uncorrelated part of the noise was caused by flow noise. By removing time segments affected by flow noise, a realistic measure of ambient noise on DTAGs was obtained.
An important caveat is that this method worked in a limited frequency range. For low frequencies (roughly below 200 Hz), the sound pressure generated by flow noise on two DTAG hydrophones also became correlated (Fig. 2). The adopted integration time and bandwidth also limited the accuracy at which the coherence function could be measured. For the short time windows adopted here, the method required periods in the recordings during which the flow noise was substantially lower (≥6 dB) than the ambient noise, to estimate the ambient noise levels. This hindered measuring the contribution of ambient noise for low frequencies, which are of interest when investigating the effects of low frequency sound sources, such as shipping sound and airguns (e.g. Clark et al., 2009).
Ambient noise levels at even lower frequencies may be obtained by increasing the integration time, or placing two separate tags in close proximity on the same animal. An increase in integration time would allow for more reliable correlation measurements when flow noise dominates, at the expense of a decreased time resolution. A larger hydrophone separation would lower the frequencies at which flow noise becomes correlated (Corcos, 1967) and at which ambient noise becomes uncorrelated. Finally, future users should keep in mind that the dynamic range of recording systems can be limited by self-noise at the low end and that measuring very low ambient noise levels requires relatively sensitive hydrophones.
High flow noise levels, followed by a sudden strong decrease in flow noise levels, are used to acoustically detect lunge feeding attempts by baleen whales (Goldbogen et al., 2006, 2013; Simon et al., 2012; Doksæter Sivle et al., 2015). The low frequency band (typically <500 Hz) is used to detect lunges, as these provide a good proxy for flow noise (Simon et al., 2012) (Fig. 1). However, surfacing events as well as whale vocalizations also lead to strong peaks in low frequency noise. The flow noise separation method proposed in this study could improve the automatic detection of lunges, by providing a direct measure of the flow noise.
We gratefully acknowledge the 3S science team and the captains and crews of the MS Strønstad, and the RV H. U. Sverdrup II, for collecting data during the 3S studies. We thank Michael Ainslie for discussions and review of the manuscript, and the three anonymous reviewers, whose questions and comments significantly improved the manuscript. We further thank the sponsors, in particular René Dekeling (NL MoD) and Mike Weise (ONR), for supporting this work.
A.M.v.B.-B. conceived the study, processed and analyzed DTAG data, and wrote the paper. P.J.O.M., P.J.W. and F.I.P.S. contributed to data collection and interpretation. S.P.B. provided the code for measuring the coherence function, and input on flow noise theory. P.J.W. analyzed animal swim speed, and carried out the DTAG calibration. F.I.P.S. and A.M.v.B.B. performed manual audits of the DTAGs for detecting transient sounds. All authors contributed to the manuscript editing.
A.M.v.B.B. and P.B. were funded by The Netherlands Ministry of Defence. Fieldwork efforts and support for P.M. and F.S. was provided by the US Office of Naval Research [award numbers N00014-08-1-0984 and N00014-10-1-0355]. P.W. received a PhD studentship with matched funding from The Netherlands Ministry of Defence (administered by The Netherlands Organisation for Applied Scientific Research, TNO) and UK Natural Environment Research Council [NE/J500276/1].
The authors declare no competing or financial interests.