Apparent source levels and active communication space of whistles of free-ranging Indo-Pacific humpback dolphins (Sousa chinensis) in the Pearl River Estuary and Beibu Gulf, China

Data collection

Acoustic recordings were made during June–July, 2014, in PRE (22°06′–22°11′S; 113°40′–113°45′E) and August 2014 in BG (21°30′–21°37′S; 108°40′–108°58′E), China (Fig. 1). Surveys were conducted from a 7.5 m recreational powerboat with a 140 hp outboard engine in PRE or a 6.8 m dolphin-watching vessel powered by 40 hp outboard engine in BG under Beaufort sea states ≤3 (on a scale of 12) with a randomly selected route rather than structured transects.

When a group of dolphins was sighted and the majority of whose members were engaged in slow or moderate movements (resting, milling, socializing or feeding) (Hawkins & Gartside, 2009), the vessel moved position to the side of the dolphin group. Groups were defined by the ‘10 m chain rule’ (Quick & Janik, 2012). If the dolphin group was traveling fast (Hawkins & Gartside, 2009), the boat would move swiftly ahead of their moving direction to await them passing by. During sound recording, the vessel’s engine was turned off. For each animal group, the GPS time, location (latitude and longitude), dolphin species, and behavior (traveling, socializing, milling, resting, and feeding) (Hawkins & Gartside, 2009) were recorded. The water depth and water quality, including temperature, salinity, and pH, were measured with a Horiba Multi-parameter Water Quality Monitoring System (model W-22XD; Horiba, Ltd., Kyoto, Japan) for sound propagation modeling. Recording was stopped when none of the dolphins of a group were within 50 m to the hydrophone arrays.

The two-dimensional cross-shaped array consisted of five Reson piezoelectric hydrophones, one in the middle and four on each end of the arms (model TC-4013, frequency range 1 Hz–170 kHz, sensitivity: −211 dB ± 3 dB re 1 V/µPa; Reson Inc., Slangerup, Denmark) (Fig. 2). Each hydrophone was equipped with a 1 MHz bandwidth Reson EC6081 voltage pre-amplifier with a band-pass filter (model VP2000, pass-band 0.1 to either 100 kHz or 250 kHz depending on sampling rate). The EC6081 employ the first order filters (one pole), which was a filter slope of 6 dB/octave in frequency. The hydrophones were connected via a 16-channel synchronized analogue-to-digital (A/D) converter to a laptop computer running LabVIEW 2011 SP1 software (National Instruments (NI), Austin, TX, USA). The A/D converter consisted of four high-speed, 16 bit resolution, data acquisition (DAQ) modules (NI 9223), incorporated in a compact DAQ four-slot USB chassis (NI cDAQ-9174). Each NI 9223 was a four-channel simultaneous A/D converter with a sample rate up to 1 MHz for each channel. Both VP2000 amplifier and NI cDAQ-9174 were powered by external battery packs.

A steel bracket was used to fix the distance between hydrophones. The bracket was made from a stainless cylinder-shaped bar with a cross structure as its backbone (bar diameter: 2.5 cm) and a reinforced stainless bar (bar diameter: 2 cm) at each quadrant (Fig. 2). A 5 cm extending bar (bar diameter: 0.3 cm) was affixed perpendicularly to the bracket plane at the center and end of each arms to mount the hydrophones, to minimize the interference of the bar to the sound (including reflection and/or shadowing) (Fig. 2). Inter-hydrophone distance along the backbone structure of the bracket was 1.47 m in PRE and 1.54 m in BG.

During the sound recording, the hydrophone arrays were deployed from the side of the boat so that the plate was in the horizontal plane at a depth of 1 m. Floats and attached weights limited array movement to reduce noise due to water flow (Fig. 2). The acoustic data were stored directly on the hard drive of a computer in technical data management streaming (TDMS) format and sampled at a rate of either 200 kHz or 512.828 kHz, giving a Nyquist frequency of 100 kHz and 256.414 kHz, respectively. The presence of signals was monitored in real-time by using both the PC screen for waveforms monitoring and a headphone connected to the center hydrophone. To minimize the chance of missing good signal, a three second pre-recording buffer was employed. Upon detecting a signal, a manual trigger was used to initiated a recording with the buffer included.

The Reson hydrophones were calibrated prior to shipment from the factory (Fig. S1). The remaining components of the recording system, including the amplifier, filter, A/D converter and laptops, were calibrated in the lab prior to the field survey by inputting a calibration signal generated by an OKI underwater sound level meter (model SW1020; OKI Electric Industry Co., LTD., Tokyo, Japan). Signal flow was also simultaneously monitored with an oscilloscope (model TDS1002C; Tektronix Inc., Beaverton, OR, USA). The noise floor of the recording system was about 65 and 55 dB re 1 µPa²/Hz at 100 Hz and 1 kHz, respectively, and flat at about 50 dB re 1 µPa²/Hz between 10 kHz and 100 kHz, which were lower than the ambient noise level at sea state 0 in our study (Fig. S1), and suitable for noise monitoring.

Sound propagation modeling

Multi-path propagation is inevitable in shallow waters, as bottom and surface reflections interfere with the signal propagation in a direct path. Following standard sound propagation theory (Au & Hastings, 2008; Aubauer, Lammers & Au, 2000; Urick, 1983), a custom-compiled sound propagation model (File S1) targeted on the impact of multi-path propagation on the original signals and took into account the hydrophone-animal geometry (such as animal depth, hydrophone depth, distance between hydrophone and animal) and site specific environment and bathymetry characteristics (such as water depth and bottom sediment contents) was adopted for this study (Fig. 3).

Since the energy flux density (EFD, dB re 1 µPa² s) is more meaningful in situations where considerable signal distortion occurs during propagation (Urick, 1983), the estimated transmission loss (TL) for each location was subsequently derived from the difference from the energy flux density of the received signal (EFD_r) and the energy flux density at the signal source (EFD_s) by the equation: (1)${\displaystyle \mathrm{TL}={\mathrm{EFD}}_r-{\mathrm{EFD}}_s}$ (2)${\displaystyle \mathrm{EFD}=10\times {log}_{10}\left\{{\int }_0^T\left(p^2\left(t\right)∕p_{\mathrm{ref1}}^2\right)dt\right\}}$ where p(t) is the sound pressure in µPa, then p_ref1 was 1 µPa² s. For each hydrophone and animal depth combination at a given water depth, the above obtained transmission losses, at varied separation distances between the hydrophone and animal were fitted to a geometric spreading loss model to estimate the environment–dependent transmission loss coefficient by the equation: (3)${\displaystyle \mathrm{TL}=k\times {log}_{10}\left(r∕r_{\text{0}}\right)}$ where k was the transmission loss coefficient, r was the distance between the animal and hydrophone, r_o was the reference range set as 1 m (Fig. 4). Frequency-dependent absorption was ignored here, since the sound absorption losses for a standard Sousa whistle, with a mean fundamental frequency of 6.35 kHz (Wang et al., 2013) as a function of site specific temperature and pressure at the PRE and BG were 0.31 and 0.30 dB/km, respectively, according to the Fisher and Simmons equation (Fisher & Simmons, 1977), and would be negligible over the ranges at which we actually recorded signals.

Acoustic data analysis

The peripheral four hydrophone channels were used for the acoustic localization of phonation animals, and the center hydrophone channel was used for detailed whistle characteristic measurement. Raven Pro Bioacoustics Software (version 1.4; Cornell Laboratory of Ornithology, NY, USA) was used to analyze the acoustic data in spectrogram (window type: Hann windows; FFT size: 8,192 and 16,384 samples for sampling frequencies of 200 and 512 kHz, respectively; frame overlapping: 80%). Only whistles with good signal-to-noise ratios (SNR ≥ 10 dB) on all five hydrophones and satisfying the criteria of no overlapping echolocation signal or whistles from different individuals were analyzed. In order to make the data more independent and to reduce the possibility of using multiple whistles from the same animal, for each dolphin encounter, we extracted only one signal for each whistle tonal type (Wang et al., 2013) for further analyzing.

Acoustic localization

Passive acoustic localization of vocalizing animals based on differences in the time of arrival of the same sound between all pairs of hydrophone receivers is a well-established technique (Au & Benoit-Bird, 2003; Janik, 2000b; Jensen et al., 2009a; Spiesberger, 1997; Spiesberger & Fristrup, 1990; Wahlberg, Møhl & Madsen, 2001; Watkins & Schevill , 1972). In this study, a custom-written package based on Matlab software (version R2010b; The Mathworks, Inc., Natick, MA, USA), named TOADY (King, Harley & Janik, 2014; Quick, Rendell & Janik, 2008; Quick & Janik, 2012; Quick & Janik, 2008; Schulz, Whitehead & Rendell, 2006; Schulz et al., 2008) was adopted for localizing phonating animals. The time delays were preserved on the simultaneous multi-track recording of signal input from all hydrophones. Signal waveforms from the different recording channels were cross-correlated to determine the difference in arrival time of a sound at each hydrophone pair. Before cross-correlation processing, a digital high-pass filter set to start rolling off just below the minimum frequency of the fundamental frequency contour of each whistle was used to eliminate any low-frequency background noise interference. The position of the largest peak in the resulting cross-correlation vector represents the amount by which the two signals are offset in time (Hayes et al., 2000). Signals with more than one equivalent peak and/or low cross-correlation maxima were discarded (Lammers & Au, 2003). The time delays were used to generate hyperboloid surfaces of possible source locations.

The standard hyperboloid can be estimated by rotating a standard hyperbola along its transverse axis. In detail, the standard hyperbola can be constructed by equations: (4)${\displaystyle x^2∕a_{ij}^2-y^2∕b_{ij}^2\text{=1}}$ (5)${\displaystyle a_{ij}=c\times t_{i-j}∕2}$ (6)${\displaystyle d_{ij}=\sqrt{{\left(x_i-x_j\right)}^2+{\left(y_i-y_j\right)}^2+{\left(z_i-z_j\right)}^2}}$ (7)${\displaystyle b_{ij}=\sqrt{{\left(d_{ij}∕2\right)}^2-a_{ij}^2}}$ where (x, y) represent the locus coordinates in two dimensions at the hyperbola located alone the hydrophone array plane, a_ij and b_ij represent the distance from the center to either vertex and the length of a segment perpendicular to the transverse axis drawn from each vertex to the asymptotes of the hyperbola between the hydrophone i and j, respectively. The symbols (x_i, y_i, z_i) and (x_j, y_j, z_j) represent the three-dimension coordinates of the hydrophone i and j, respectively, c represents the speed of sound in water in m/s, and t_i−j represents the time delay between the hydrophones i and j in seconds. The maximum allowable time delay between a pair of hydrophones in the array is limited to the direct-path propagation time between them (Helble et al., 2015) as: (8)${\displaystyle max\left(t_{i-j}\right)=d_{ij}∕c}$ where d_ij represent the separation of the hydrophone i and j in m. The standard hyperboloid was then rotated and further recast to the center of the spatial geometry of the corresponding array-pair positions.

Once all the hyperboloids were established, contours of the hyperboloids (hyperbolae) at varied assumed animal depths, ranging from the water surface to the bottom set at 0.5 m increments, were displayed in the graphical interface of the TOADY software for visual inspection the hyperbolic fixing (Fig. 5C). Four hydrophones resulted in six hyperbolae and yield four points of intersection (for each independent combination of a hydrophone triad, only two of the three time differences were linearly independent, and all three hyperbolae intersected at a single point) (Laurinolli et al., 2003). The localization accuracy was increased by inclusion of the depth function (Quick, Rendell & Janik, 2008), and animal depth was estimated as that where the surface area of the polygon formed by the hyperbola intersections was minimum (Quick, Rendell & Janik, 2008). The average of the hyperboloid intersections was taken as the best estimate of the sound source’s location (Clark & Ellison, 2000; Laurinolli et al., 2003; Schulz, Whitehead & Rendell, 2006; Schulz et al., 2008).

Ideally, all the four intersections occurred at one point (Fig. 5C). The location error was assessed by a linear error propagation model (Taylor, 1997), and the root-mean-square (rms) location error was estimated using the equation: (9)${\displaystyle {\epsilon }_{\mathrm{rms}}=\sqrt{{\epsilon }_x^2+{\epsilon }_y^2+{\epsilon }_z^2}}$ where ε_x, ε_y, and ε_z are the standard deviation (SD) of the hyperbolae intersections in the zonal, meridional, and vertical directions, respectively (Laurinolli et al., 2003; Schulz et al., 2008; Wahlberg, Møhl & Madsen, 2001).

Signal extraction

Whistles with successful source location estimates were extracted for sound parameter analysis using the center hydrophone channel. The extracted whistle was assigned to one of the following six tonal types based on its fundamental time-frequency contour as: flat, down, rise, U-shape, concave and sine. All tonal types were mutually exclusive (Fig. 6, for detailed definition, see Wang et al., 2013). A three-step procedure was applied to extract the candidate whistles (Fig. 7). A 2-s signal was extracted for each candidate whistle (the whole signal in Figs. 7A and 7B). The actual whistle was subsequently measured from the start and end points of the fundamental contour (Fig. 7C) and further extracted it as the portion containing 98% of the total cumulative energy, which started at the time when 1% of the cumulative signal energy was reached (t_1%ce, in Fig. 7E) and ended when 99% of the cumulative signal energy was reached (t_99%ce, in Fig. 7E). Whistle duration was derived from the time difference between the 1st and 99th cumulative energy percentiles (in Fig. 7E). A 500 ms ambient noise selection was extracted either before of after (in Fig. 7B) each whistle from the 2-s signal, with a gap of over 0.2 ms from either sides of the whistle fundamental contour (in Fig. 7B). All spectrograms were computed with 25 ms Hann windows (5,000 and 12,820 samples, zero-padded to 8,192 and 16,384 samples for sampling frequencies of 200 and 512 kHz, respectively) for FFT computation with 80% overlap for a temporal resolution of 5 ms and an interpolated spectral frequency resolution of 24.4 and 31.1 Hz, respectively.

Apparent source levels and source energy flux density

For each whistle, the root-mean-square sound pressure levels (SPL_rms, dB re 1 µPa) and energy flux density (EFD) were calculated using the following equations (Au & Hastings, 2008): (10)${\displaystyle SP{L}_{\mathrm{rms}}=10\times {log}_{10}\left\{1∕T\times {\int }_0^T\left(p^2\left(t\right)∕p_{\mathrm{ref2}}^2\right)dt\right\}}$ where p(t) was the sound pressure in µPa, and p_ref2 was 1 µPa. SPL_rms critically relies upon the signal window size (T) in Eq. (10) (Madsen, 2005). Bottlenose dolphins integrate pure-tone acoustic energy in the same way as humans (Johnson, 1968b), with the integrating time constant for the pure-tone range from 1 kHz to 8 kHz approximately 200 ms (Johnson, 1968b; Plomp & Bouman, 1959). The representative range of the fundamental frequencies of the Sousa whistle averaged at 6.4 kHz with the minimum and maximum fundamental frequency average at 5.1 kHz and 7.7 kHz, respectively (Wang et al., 2013). Here, we assumed that the integration time constant from bottlenose dolphins also applied to Sousa. Both whistles and matched noise samples were consecutively cut into segments of 200 ms with two adjacent slices overlapping by 95%. A measure termed SPL_rms200 was taken as maximum SPL_rms value from the 200 ms slices of each whistle, and SPL_noi was derived from the average SPL_rms value of the 200 ms, slices of each matched noise sample. Absolute pressure levels were derived by incorporating the sensitivity of the hydrophone and the amplifier gain (Au & Hastings, 2008). Apparent source levels (ASLs) and source energy flux density (SEFD) were estimated from the received apparent sound pressure levels and energy flux density by compensating for the transmission loss using the site-specific transmission loss model.

Power spectral density and one-third octave band levels

Power spectral density (dB re 1 µPa²Hz⁻¹), the averaged sound power in each 1 Hz band (Sims et al., 2012b) were calculated using Welch approach for each whistle over its 98% energy windows and their corresponding noise to assess the detailed acoustic energy distribution. Their one-third octave band levels (dB re 1 µPa²) were further calculated to assess how cetaceans auditory systems perceive sound and were impacted by ambient noise (Madsen et al., 2006). All power spectral density and one-third octave band levels were computed with 0.2 s slice window, with 95% overlap between two slices for FFT computation, resulting in an interpolated spectral frequency resolution of 3.05 and 3.91 Hz for sampling frequencies of 200 and 512 kHz, respectively.

Active communication space

Detection of a tonal signal against a continuous broad-band noise background will be effectively masked by only a relatively narrow band of frequencies centered on the tonal stimulus, namely the critical bandwidth (Fletcher, 1940). The critical ratio is another measure of auditory filter width and an indirect method for estimating critical bandwidth (Au & Moore, 1990). At the detection threshold, the signal power equals the noise power, so that the auditory filter width is the ratio of the threshold intensity of a tone over the ambient noise power spectral density at the frequency in question (Fletcher, 1940).

The active communication space is a combined function of the signal source level, the dolphin auditory threshold, the habitat-specific transmission loss, and the masking level at the one third octave band center frequency in question (Janik, 2000a; Jensen et al., 2012; Quintana-Rizzo, Mann & Wells, 2006). The masking level is determined by the noise one-third octave band level or the masked tone threshold, whichever dominated. The masked tone threshold is the sum of the noise power spectral density and the critical ratio at the frequency in question (Janik, 2000a; Jensen et al., 2012; Quintana-Rizzo, Mann & Wells, 2006). The active communication space of each whistle is estimated as the maximum range at which the signal can still be detected in at least one of the one-third octave bands analyzed after accounting for the transmission loss (Miller, 2006). For whistle signals, the one-third octave band that determines the maximum range is always at the peak frequency of the signal one-third octave band levels (Fig. 8).

The active communication space for each whistle can be modeled by the equations: (11)${\displaystyle k\times {log}_{10}\left[\mathrm{mean}\left(\mathrm{ACS}\right)\right]=\text{TL}=\left[\mathrm{mean}\left(\mathrm{Sig\_TOBL}\right)\right]\left(\mathrm{fp}\right)-max\left\{\mathrm{ML(fp)},\mathrm{AT(fp)}\right\}}$ (12)${\displaystyle k\times {log}_{10}\left[max\left(\mathrm{ACS}\right)\right]=TL=\left[max\left(\mathrm{Sig\_TOBL}\right)\right]\left(\mathrm{fp}\right)-max\left\{\mathrm{ML}\left(\mathrm{fp}\right),\mathrm{AT}\left(\mathrm{fp}\right)\right\}}$ (13)${\displaystyle \mathrm{ML}\left(\mathrm{fp}\right)=max\left\{\left[\mathrm{mean}\left(\mathrm{Noi\_TOBL}\right)\right]\left(\mathrm{fp}\right),\mathrm{MTT}\left(\mathrm{fp}\right)\right\}}$ (14)${\displaystyle \mathrm{MTT(fp)}=\left[\mathrm{mean}\left(\mathrm{Noi\_PSD}\right)\right]\left(\mathrm{fp}\right)+\text{CR(fp)}}$ where ACS was the active communication space of the whistle in the near simultaneous ambient noise conditions obtained from the matched noise sample, mean(Sig_TOBL) and max (Sig_TOBL) were the averaged and maximum one-third octave band level for all the 200 ms slices for each whistle, f_p was determined by the peak frequency of the averaged one-third octave band levels for all the 200 ms slices for each whistle, mean(Noi_PSD) and mean(Noi_TOBL) were the averaged power spectral density and one-third octave band levels of all the 200 ms slices from the matched noise sample for each whistle, ML was the masking level, MTT was the masked tone threshold, and AT was Sousa auditory threshold. The Sousa audiogram with a frequency span of 500 Hz–38 kHz (which cover the fundamental contour range of Sousa whistles of 520 Hz–33 kHz (Wang et al., 2013)) was estimated by fitting a third-order polynomial curve to the auditory thresholds between 5 kHz and 38 kHz (Li et al., 2012). CR was the dolphin critical ratio (Johnson, McManus & Skaar, 1989), and was obtained by following the equation: (15)${\displaystyle \mathrm{CR}=19.8+0.075\times f^{1∕2}}$ where CR was in dB and f was the frequency in Hz. The equation was obtained by applying a least-square fit to the bottlenose dolphin critical ratio data (Johnson, 1968a; Moore & Au, 1982).

In cases where the masking level was always higher than the relevant Sousa auditory threshold, i.e., the active communication space was noise-limited, the theoretical active communication space determined by the Sousa auditory threshold alone was also calculated. The active communication space determined by auditory threshold alone (ACS_at) can be modeled as: (16)${\displaystyle k\times {log}_{10}\left[\mathrm{mean}\left({\mathrm{ACS}}_{\mathrm{at}}\right)\right]=\mathrm{TL}=\left[\mathrm{mean}\left(\mathrm{Sig\_TOBL}\right)\right]\left(\mathrm{fp}\right)-\mathrm{AT}\left(\mathrm{fp}\right)}$ (17)${\displaystyle k\times {log}_{10}\left[max\left({\mathrm{ACS}}_{\mathrm{at}}\right)\right]=TL=\left[max\left(\mathrm{Sig\_TOBL}\right)\right]\left(\mathrm{fp}\right)-\mathrm{AT}\left(\mathrm{fp}\right).}$

Statistical analysis

Descriptive statistics of all measured acoustic parameters were obtained and presented in the form of mean, SD, and ranges if they were normal distributed. For those parameters with a grossly skewed distribution, descriptive parameters of median, quartile deviation (QD), 5 percentile (P5), and 95 percentile (P95) were adopted. The Levene’s test for equality of error variances and Kolmogorov–Smirnov goodness-of-fit test were used to analyze homogeneity of variance and the distributions of the data, respectively. Nonparametric statistical analyses (Zar, 1999) were adopted if data were not normally distributed. The Kruskal–Wallis test (Zar, 1999) was used to examine the difference in the mean of the transmission loss coefficient of different test signals running in the sound propagation model. The Mann–Whitney U-test (Zar, 1999) was used to analyze differences between transmission loss coefficients, as well as acoustic parameters between sites. Differences in apparent source levels and energy flux density across different whistle tonal types was analyzed by the Kruskal–Wallis test (Zar, 1999), and Duncan’s multiple comparison test (Zar, 1999) was used for post hoc comparisons of acoustic parameter among tonal types. Statistical analyses were performed using SPSS 16.0 for Windows (SPSS Inc., Chicago, USA). Differences were considered significant at p < 0.05.

Ethical statement

Permission to conduct the study was granted by the Ministry of Science and Technology of the People’s Republic of China. The research permit was issued to the Institute of Hydrobiology of the Chinese Academy of Sciences (Permit number: 2011BAG07B05).

Sound propagation modeling

Sixteen whistles with estimated source distance from one of the 5 hydrophone channels close to 1 m were selected as a proxy for whistle sources and imported to the sound propagation model. The sound propagation speed in water (c_w) was calculated as 1,538 m/s and 1,535 m/s, for the Pearl River Estuary and Beibu Gulf, respectively, according to the Medwin equation (Medwin, 1975) based on the average measured site-specific temperature, salinity, and water depth (Table 2). The sound speed in air (c_s) was 343 m/s, and in bottom sediment (c_b) was 1,535 and 1,742 m/s for Pearl River Estuary and Beibu Gulf, respectively. The impedances z_s and z_w were 400 and 1.54 × 10⁶ Pam⁻¹ s, z_b was 2.19 × 10⁶ and 3.45 × 10⁶ Pam⁻¹ s for Pearl River Estuary and Beibu Gulf, respectively (Urick, 1983), since the sediment types were different, with clay silt in Pearl River Estuary and fine sand in Beibu Gulf (TQ Zeng & HW Su, pers. comm., 2014). The hydrophone depth was set at 1 m to mirror the real hydrophone setup during sound recording. The water depth in Pearl River Estuary and Beibu Gulf was set as a range from 5 m to 9 m and from 2 m to 8 m, respectively, to mirror the measured site-specific depth (Table 2). The maximum distance between animal and hydrophone was set at 50 m (Fig. 4), corresponding to the maximum localized whistle distance of 49.6 m (see next paragraph). The reflective coefficients at the air–water interface were averaged at −0.09 for both of the Pearl River Estuary and Beibu Gulf, whereas the reflective coefficients at the water-bottom interface were averaged at 0.24 and 0.46 for Pearl River Estuary and Beibu Gulf, respectively. No significant difference in the mean transmission loss coefficient k was observed among different testing signals within each site (Pearl River Estuary: Kruskal–Wallis χ² = 16.82, df = 15, p = 0.33; Beibu Gulf: Kruskal–Wallis χ² = 5.02, df = 15, p = 0.91). Thus, we pooled all test signals within the sites to calculate average transmission loss. From the pooled data we estimated an average k in Pearl River Estuary of −17.3 ± 1.0 (95% CI [−17.4:−17.2]), which was significantly different from that in Beibu Gulf of −14.6 ± 0.8 [95% CI (−14.7:−14.5)] (Mann–Whitney U-test: z = 1,532, df = 1,119, p < 0.01) (Fig. 9).

Acoustic localization

Of all the analyzed whistles, the average estimated distance between the center hydrophone and phonating animals was 6.8 ± 4.2 (SD) m (range: 2.1–23.8 m) in Pearl River Estuary and 8.4 ± 7.0 (SD) m (range: 2.2–49.6 m) in Beibu Gulf (note that the distribution of the distance between hydrophone and phonating animal may vary if a different hydrophone was chosen). The average localization error (ε_rms) was 0.3 ± 0.2 m and 0.5 ± 0.4 m in Pearl River Estuary and Beibu Gulf, respectively.

Apparent source levels and source energy flux density

Significant differences in whistle duration was observed between Pearl River Estuary (mean ± SD: 0.50 ± 0.19 s) and Beibu Gulf (mean ± SD: 0.44 ± 0.20 s) (Mann–Whitney U test, z = − 2.0, df = 241, p = 0.04) (Table 3). On the other hand, no significant differences were found in all the measured apparent source levels and source energy flux density between Pearl River Estuary and Beibu Gulf (Mann–Whitney U test, p < 0.05) (Table 3), thus they were pooled according to tonal classes for further analysis.

The 242 successfully located whistles consisted of 66 flat, 28 down, 28 rise, 64 U-shape, 15 convex, and 41 sine whistles (Table 1). Significant differences were observed in the whistle duration, apparent source levels, and source energy flux density among the six tonal types (whistle duration: Kruskal–Wallis χ² = 29.27, df = 5, p < 0.01; ASL_rms: Kruskal–Wallis χ² = 17.02, df = 5, p < 0.01; ASL_rms200: Kruskal–Wallis χ² = 12.08, df = 5, p = 0.03 and source energy flux density: Kruskal–Wallis χ² = 11.16, df = 5, p = 0.04). In detail, the duration of sine whistles (mean ± SD: 0.60 ± 0.3 s) was significantly longer than flat (mean ± SD: 0.43 ± 0.2 s), rise (mean ± SD: 0.43 ± 0.2 s) (Duncan’s multiple-comparison test; p < 0.05) and U-shape (mean ± SD: 0.39 ± 0.2 s) whistle types (Duncan’s multiple-comparison test; p < 0.01). Sine whistles had significantly lower ASL_rms200 andsource energy flux density (Duncan’s multiple-comparison test; p < 0.05) and ASL_rms (Duncan’s multiple-comparison test; p < 0.01) than down whistles (Fig. 10).

One-third octave band levels

The average peak frequency across the one-third octave band levels for all the 200 ms whistles slices across all whistles was 4.9 ± 1.0 kHz and 6.1 ± 2.6 kHz for Pearl River Estuary and Beibu Gulf, respectively, which were significantly different (Mann–Whitney U test, z = − 3.0, df = 241, p < 0.01) (Table 3). No significant difference was observed in the ambient noise sound pressure levels between Pearl River Estuary (SPL_noi; mean ± SD: 122.3 ± 5.0 dB) and Beibu Gulf (mean ± SD: 122.2 ± 6.3 dB) (Mann–Whitney U test, z = − 1.0, df = 241, p = 0.32) (Table 3). However, their one-third octave band noise level property was significantly varied: in Pearl River Estuary, the one-third octave band level at frequency band of 4.47–5.63 kHz (centered at 6.3 kHz and corresponding to the peak frequency of whistles in Beibu Gulf) was significantly higher than the band of 5.63–7.08 kHz (centered at 5 kHz and corresponding to the peak frequency of whistles in Pearl River Estuary) (mean ± SD: 95.8 ± 3.2 dB and 97.1 ± 4.2 dB, respectively; Wilcoxon signed ranks test, Z = − 4.99, p < 0.001, N = 33), whereas, an opposite trend was observed in Beibu Gulf (mean ± SD: 99.3 ± 4.3 dB and 96.9 ± 4.2 dB, respectively, for the two frequency bands; Wilcoxon signed ranks test, Z = − 8.59, p < 0.001, N = 209). In addition, the one third octave level at frequency band of 4.47–5.63 kHz in Pearl River Estuary was significantly higher than that in Beibu Gulf (Mann–Whitney U-test: z = − 4.02, df = 241, p < 0.001), although no significant difference in the frequency band of 5.63–7.08 kHz was observed between them (Mann–Whitney U-test: z = − 1.68, df = 241, p = 0.09).

Active communication space

We calculated the mean and maximum active communication spaces across the 200 ms segments for each analyzed whistle. The median of these results were 14.7 ± 2.6 m and 17.1 ± 3.5 m, respectively, in Pearl River Estuary, and 34.2 ± 4.8 m and 43.5 ± 4.6 m, respectively, in Beibu Gulf. Both measures were significantly smaller in Pearl River Estuary than in Beibu Gulf (Mann–Whitney U test: z = − 5.5, df = 241, p < 0.01 and z = − 5.8, df = 241, p < 0.01, respectively). The largest mean active communication spaces for any whistle in Pearl River Estuary and Beibu Gulf were estimated to be at 40.7 m and 209.7 m, respectively, while the largest maximum active communication spaces were estimated to be 51.1 m and 266.8 m, respectively. Significant differences were observed in the mean and maximum active communication space among six tonal types, which follow from the observed differences in apparent source level (mean active communication space: Kruskal–Wallis χ² = 25.56, df = 5, p < 0.01; maximum active communication space: Kruskal–Wallis χ² = 23.80, df = 5, p < 0.01). In detail, both of the mean and maximum active communication space of flat was significantly shorter than that in down and U-shape (Duncan’s multiple-comparison test; p < 0.01). The mean active communication space of flat was significantly short than that in rise (Duncan’s multiple-comparison test; p < 0.01). Of all the localized whistles, the averaged one-third octave band levels of all matched 200 ms noise samples for each whistle at the frequency determined by the peak frequency of the averaged one-third octave band levels of all the whistle 200 ms slices ((mean(Noi_TOBL))(fp)) were significantly higher than Sousa auditory threshold at the same frequency (AT(fp)) (Mann–Whitney U test, z = − 9.25, df = 241, p < 0.01), indicating that active communication space was mainly noise-limited. However, 62 out of 242 whistles (26.6 %) with the (mean(Noi_TOBL))(fp) lower than its corresponding AT(fp), indicating that their active communication space were auditory-threshold limited, as opposed to noise-limited, we also estimated the theoretical auditory-threshold-limited active communication space for all whistles. The theoretical mean auditory-threshold limited active communication space in Pearl River Estuary (median ± QD: 24.3 ± 4.8 m) was significantly shorter than that in Beibu Gulf (median ± QD: 60.7 ± 12.2 m) (Mann–Whitney U test, z = − 4.2, df = 241, p < 0.01), and the maximum was also significantly shorter in Pearl River Estuary (median ± QD: 35.7 ± 4.6 m) that in Beibu Gulf (median ± QD: 74.3 ± 25.3 m) (Mann–Whitney U test, z = − 4.4, df = 241, p < 0.01) (Table 3). The biggest mean auditory-threshold limited active communication space in Pearl River Estuary and Beibu Gulf were estimated to be 106.5 m and 457.2 m, respectively, whereas the biggest maximum ACS_at were estimated to be 109.3 m and 463.8 m, respectively. Significant difference was observed in the mean and maximum auditory-threshold limited active communication space among six tonal types (Kruskal–Wallis χ² = 17.04, df = 5, p < 0.01 and Kruskal–Wallis χ² = 15.28, df = 5, p < 0.01, respectively). In detail, both of the mean and maximum auditory-threshold limited active communication space of flat were significantly shorter than that in U-shape (Duncan’s multiple-comparison test; p < 0.05).

Sound propagation modeling

Depending on local conditions, the spreading loss constant k normally ranges from spherical spreading loss (k = 20) to cylindrical spreading loss (k = 10) (Urick, 1983). The transmission loss coefficient in Pearl River Estuary (mean ± SD: −17.3 ± 1.0; range: from −20.5 to −15.5) followed an almost spherical spreading model and was comparable to that calculated for shallow waters of Koombana Bay in Western Australis of −18, which was derived from playback experiments with variable receiver and sender locations at an approximate homogenous water depth of 5–7 m, using pure tone sounds with a frequency span of 1–7 kHz (Jensen et al., 2012) and the empirical and theoretical derived transmission loss coefficient at the Antarctic Peninsula of −17.8 (Širović, Hildebrand & Wiggins, 2007). The transmission loss coefficient in Beibu Gulf (mean ± SD: −14.6 ± 0.8; range: from −16.5 to −12.1) was equidistant between spherical and cylindrical spreading. The range of the transmission loss coefficient estimated in Pearl River Estuary and Beibu Gulf (Fig. 9) was also comparable to that estimated with a water depth less than 3 m and in mud or sand sediment (k range from −26.6 to −14.5) and in channels with water depth between 3 m and 5.3 m (k range from −24.5 to −12.8) in Sarasota Bay, Florida (Quintana-Rizzo, Mann & Wells, 2006). Therefore, our transmission loss estimates were well within the range of currently published findings in similar habitats.

Acoustic localization

The potential factors that influence the source location include ambient noise in the transmission paths, variability in sound speed in the media (Clark & Ellison, 2000; Tiemann, Porter & Frazer, 2004), multipath arrivals of a signal at the hydrophone (Clark & Ellison, 2000; Hayes et al., 2000), and array configurations (Janik, 2000a). However, we estimated low localization error in both of the Pearl River Estuary and Beibu Gulf, with error percentages (ε_rms divided by mean localized distance between hydrophone and animal (r)) of 0.3/6.8 = 4.4% and 0.5/8.4 = 5.9%, respectively. Therefore, error due to source localization was not expected to be large in our results.

The low localization error in this study may be ascribed to the following counter error analysis strategies, firstly, the use of waveform, rather than spectrogram cross-correlation technique. In order to determine the time delay between two hydrophones, a cross-correlation function can be applied to either signal waveforms or spectrograms (Janik, 2000a). Compared with waveform cross-correlation, spectrogram cross-correlation technique may introduce a slightly larger error (Clark & Ellison, 2000). The error was due to the time grid spacing resolution during spectrogram calculation by averaging over small time slices of the original signals (Janik, 2000a). Secondly, the inclusion of the scanning depth function (an extra dimension to the array plane) in the TOADY localization program in conjunction with the two dimensional array recording systems makes possible the three dimensional localization of the animal and helps to increase the accuracy of the localization (Quick, Rendell & Janik, 2008), since using a two dimensional array in a three dimensional environment may generate some errors (Janik, 2000a). On the other hand, the optimal source location in passive acoustic localization can be achieved when phonation animals inside the array and the source-to-hydrophone distance of the same order of magnitude as inter-hydrophone spacing (Madsen & Wahlberg, 2007). Additionally, source-to-array distance to a range of three to four times the hydrophone array dimension can still be localized with high confidence (Watkins & Schevill , 1972). Of all the localized whistles in this study, those with a source-to-hydrophone distance smaller than 3 m, 12 m, and 15 m, which mimics the distance of one time, three times, and four times of the maximum inter-hydrophone spacing, account for 27.3%, 83.9% and 93.9%, respectively. In addition, the 95% confidence interval of the mean distance between the animal and hydrophone (Pearl River Estuary: 5.4–8.3 m, Beibu Gulf: 7.4–9.3 m) was well within three times of the maximum inter-hydrophone spacing. This also account for the high acoustic localization in this study.

Apparent source level and noise levels

The apparent source levels of Sousa whistle obtained in this study, with the ASL_rms over its 98% energy window of 137.4 ± 6.9 dB (range: 114.1–160.4 dB) and ASL_rms over 200 ms running windows of 139.5 ± 6.9 dB (range: 115.6–161.4 dB), were similar to the estimated mean and range of the ASL_rms of Atlantic spotted dolphin (Stenella frontalis) and bottlenose dolphin at Gulf of Mexico (Frankel et al., 2014), but lower than those found for other dolphin species, including Hawaiian spinner dolphins (Stenella longirostris) (Watkins & Schevill, 1974), common dolphin (Deiphinus delpliis) (Fish & Turl, 1976), white-beaked dolphins (Rasmussen et al., 2006), baiji (Lipotes vexillifer) (Wang et al., 2006), short-finned pilot whale (Globicephala macrorhynchus) (Fish & Turl, 1976) and killer whale (Orcinus orca) (Miller, 2006), and bottlenose dolphin (Fish & Turl, 1976; Janik, 2000a; Jensen et al., 2012; Tyack, 1985) in other regions (Table 4). The observed no significant difference in broad band noise level between the two regions in our study may be due to the fact that, in order to model the active communication range of whistles in real time noise conditions, the analyzed noises were derived from sound files with whistle recorded in good SNR and successfully localized. However, the majority of good whistles were obtained either far away from the construction or navigation region in Pearl River Estuary or without dolphin tourism boat nearby in Beibu Gulf and represent an environment with less anthropogenic noise pollution. Besides, the fact that the successfully localized whistles in this study were from a less anthropogenic impacted environment may, in part, account for the low source level of the humpback dolphin whistles, since dolphins may use higher amplitude sound in a noisy environment. Additionally, the frequency of dolphin whistles tend to modify according to environmental ambient noise, and the bottlenose dolphin was observed to produced whistles with lower (Morisaka et al., 2005) or higher (May-Collado & Wartzok, 2008) frequencies at higher ambient noise conditions. In the present study, both the averaged peak frequency of whistles and one third octave band level of noise were significantly different between Pearl River Estuary and Beibu Gulf. However, the reasons why humpback dolphins emitted whistles with peak frequency coincidence with the noise one third octave frequency band with higher noise level in both of these two regions deserve further investigation.

Active communication space

Active communication space has been estimated for bottlenose dolphins (Janik, 2000a; Jensen et al., 2012; Quintana-Rizzo, Mann & Wells, 2006), killer whale (Miller, 2006), white–beaked dolphins (Rasmussen et al., 2006) and baiji (Wang et al., 2006). In these studies, the whistle source levels used to model the active communication space were either the highest and lowest source levels (Rasmussen et al., 2006), the mean source level (Wang et al., 2006), the highest and mean source levels (Janik, 2000a), the three ‘representative’ source levels (Quintana-Rizzo, Mann & Wells, 2006), the source level of each whistle over its 95% energy window (Jensen et al., 2012) or on the one-third octave band level scale (Miller, 2006). The noise condition was either extrapolated from literature at other habitats at sea-state 0 and sea-state 4 or 6 (Janik, 2000a; Miller, 2006), or using the noise level at the habitat in question under an optimal conditions of sea-state 0–1 and without an adjacent boat (Jensen et al., 2012; Rasmussen et al., 2006), or under normal recording conditions (Quintana-Rizzo, Mann & Wells, 2006; Wang et al., 2006). The propagation model was either adopted from the spherical transmission attenuation model (Rasmussen et al., 2006; Wang et al., 2006), the Marsh and Schulkin shallow-water transmission model (Marsh & Schulkin, 1962) which assumes the transducer and receiver were in the middle of the water column (Janik, 2000a; Miller, 2006) or derived by the sound transmission experiment using playback signal of computer-generated whistle mimic tone at varied distance with the transducer and receiver at a fixed depth of 1 m (Quintana-Rizzo, Mann & Wells, 2006), or with varied transducer and receiver combination (Jensen et al., 2012). However, when using the signal ASL_rms for modeling active communication space, the critical band theory of the mammalian auditory system was not considered, and may therefore have integrated the energy from harmonics, which was the integer times of the fundamental frequency and tends to be directional (Lammers & Au, 2003; Miller, 2002; Rasmussen et al., 2006), and more heavily affected by absorption. Additionally, the optimal or averaged noise condition, rather than the real-time ambient noise level used for modeling active communication space, may not shed much light for the biologically relevant active communication space of the signal. This was further corroborated by the findings that the modeled active communication space of baiji whistles with a ASL_rms of 143.2 dB will reduced from 6.6 km at normal noise level to a range of 22–220 m in a noisy boat traffic conditions. Furthermore, the Marsh and Schulkin model or playback experiment with the transducer and receiver at a fixed depth may not represent the transmission loss pattern of the site, because the sound transmission loss may vary within the water column (Quintana-Rizzo, Mann & Wells, 2006). Transmission loss near the surface or the bottom of the water column was observed to be much higher than that at the center of the water column (Janik, 2000a). One-third of an octave approximates the effective filter bandwidth of mammalian hearing systems (Fletcher, 1940; Richardson et al., 1995). The signal one-third octave band level information provided us with a starting point and an appropriate way for investigating how a mammal’s auditory system perceives sound and the extent of the masking effect of the ambient noise (Blackwell, Lawson & Williams, 2004; Madsen et al., 2006). In this study, models integrating the signal one-third octave band levels for each individual whistle, its corresponding real-time noise conditions, and site-specific transmission attenuation were adopted to estimate the active communication space in a more meaningful way. It should be noted that sound detection and discrimination thresholds may differ; for example, in all the birds of budgerigars (Melopsittacus undulates), zebra finches (Taeniopygia guttata) and canaries (Serinus canaria), the thresholds for discrimination between calls of the same species was observed higher than the thresholds for detection of those calls (Lohr, Wright & Dooling, 2003). The weakest portions of an emitted signal will always be lost during transmission before it is detected by a receiver, but whether or not the transmitting animal can be discriminated by a receiving animal will depend on how much of the information in the signal is needed to solve this task. Individual discrimination information of bottlenose dolphins was encoded in the frequency modulation contour of their signature whistles (Janik, Sayigh & Wells, 2006). Thus, the active communication space of signature whistles was determined by the weakest portion of the signal modulation contour. However, whether signature whistles also exist in humpback dolphins is still unknown and deserve further research. Since whistles were narrowband signals, we believe that the active communication space calculated by using the mean(Sig_ TOBL), as obtained by a running average of all the one-third octave band levels for each whistle over its 98% energy windows, will be similar to that based on the bandwidth of the entire whistle.

The small active communication space calculated in this study is a result of ambient noise limiting detection ranges. The averaged one-third octave band noise levels were 98.7 ± 4.5 and 96.8 ± 4.2 dB re 1 µPa at the band which accommodates the peak frequency of whistles at 4.9 kHz and 6.1 kHz for Pearl River Estuary and Beibu Gulf, respectively. These results were almost 20 dB higher than the one-third octave band noise levels at the same frequency band (estimated to be less than 75 dB) for the estimation of the whistle communication range of the white-beaked dolphins in Faxafloi Bay, Iceland (Rasmussen et al., 2006), almost 30 dB higher than the ambient noise level (estimated to be 66.8 dB) for the estimation of the whistle communication range of baiji in the Shishou reserve (Wang et al., 2006), about 50 and 30 dB higher than the ambient noise level at sea state 0 and 4 respectively used for the estimation of the whistle communication range of bottlenose dolphins at Moray Firth (Janik, 2000a), and almost equal to the background noise spectrum level plus critical ratios for the estimation of the whistle communication range of bottlenose dolphins in Sarasota Bay (Quintana-Rizzo, Mann & Wells, 2006). The high ambient noise level observed in both of the Pearl River Estuary and Beibu Gulf at less anthropogenic impacted areas may due to the waves and biological noise, such as snap shrimps, and deserves further research.

The flourishing year round dolphin-watching industry makes the Beibu Gulf dominated by dolphin-watching boats of about 7 m length and equipped with 40 horse power engine during the day time, whereas the fast developing local economy makes Pearl-River Estuary dominated by different kinds of hydrofoil ferries between Hong Kong, Zhuhai, and Macao throughout the day, with the ferry length of 27.4–47.5 m and speed of 28–45 knots (Z-T Wang, 2014, unpublished data). Since different vessels tend to had very different acoustic signatures (Hermannsen et al., 2014). The varied vessel traffic condition between these two regions might also be a clue to the sources contributing to the ambient noise. Longtime noise recordings will be needed for a better view of the soundscape difference between these two regions.

The apparent source levels between different tonal classes was only significant varied between sine and down. The significant difference in active communication space between flat and down, and between U-shape and rise may be due to their frequency band variations (Wang et al., 2013), since the active communication space was determined by the one-third octave band level at the whistle peak frequency and the corresponding noise one-third octave levels at the same frequency band. In addition, the significant differences in apparent source levels and its active communication space among different whistle tonal types may be associated with their different functions in different behavioral context. This was further corroborated by the findings that different whistle tonal classes were in relation with different behavior states in Indo-Pacific bottlenose dolphins (Hawkins & Gartside, 2010) and some whistle types were used to convey specific information on their specific behavioral context (Hawkins & Gartside, 2010).

Acoustic cues play an important role in mediating social structure and were critical for aquatic animals, especially cetaceans that rely heavily on acoustic for communication. The active communication space calculated in this study can be used for defining dolphin groups in a more biologically meaningful way during field surveys. As determined by the mean active communication range, humpback dolphins within the distance of 14.7 m in Pearl River Estuary and 34.2 m in Beibu Gulf might be able to keep acoustic contact and can be defined as the same dolphin group. In a more quieter environment, where dolphin auditory threshold determine the active communication range, animals which apart from each other at a distance of less than 24.3 m in Pearl River Estuary and 60.7 m in Beibu Gulf can be grouped into the same dolphin group.

Auditory masking of communication signals may interfere the acoustically mediated social interactions. During our field survey in Beibu Gulf, when fast moving dolphin-watching boat was presented even at a distance of more than 1,000 m away, the whistles recorded will be severely masked with the generated vessel noise dominating the ambient noise. The negative impacts of vessel noise on cetacean have been widely documented. The vessel noise recorded at the heavily ship-trafficked marine habitats in Denmark from a range of different ship types was observed substantially elevated ambient noise levels across a wide frequency band of 0.025–160 kHz and can cause hearing range reduction of over 20 dB on harbor porpoises (Phocoena phocoena) even at a distances of over 1 km away (Hermannsen et al., 2014). Small vessels noise can significantly mask acoustically mediated communication in cetaceans and those travelling at 5 knots in shallow waters of the Koombana Bay, Western Australia can reduce the communication range of 26% on bottlenose dolphins within a distance of 50 m (Jensen et al., 2009a). The impact of vessel traffic on local humpback dolphins, especially the hydrofoil ferries at heavy shipping lanes of the PRE, deserves further research.

Limitations

Cetacean audiograms can vary greatly among individuals of different age groups (Houser, Gomez-Rubio & Finneran, 2008; Popov et al., 2007), and the threshold may shift with the presence of masking noise (Johnson, 1968a). Therefore, the audiogram adopted in this study may not be considered representative of the auditory sensitivity of Sousa as a whole, and more auditory studies covering different dolphin age groups are needed for deriving an accurate audiogram of Sousa. The source level of dolphin whistles can vary depending on the number, age, and sex of the individuals in a group, as well as their behavioral context (Fish & Turl, 1976; Frankel et al., 2014). The maximum apparent sound pressure level of 160.4 dB (rms) observed here does not necessarily represent the capability of the species. Due to the trade-off between high SNR and increased active space for high amplitude sound and decreased detection probability by predator with low amplitude sound (Morisaka & Connor, 2007), dolphins might not produce whistles at their maximum levels, even if rewarded in a conditioning procedure (Janik, 2000a).

The directional pattern of outgoing signals and any directional hearing capability will impact the active communication range of animal vocalization (Au, 1993). Received sound pressure levels will be maximum when the phonating dolphin is pointed directly at the receiver, and vice verse. In this study, both the directivity index (DI) and the beam pattern (BP) of the outgoing signal transmission and sound receiving at the frequency of the Sousa whistle were assumed to be 0 dB. The assumption of the 0 dB directivity index of the sound receiving system was further corroborated by the findings that bottlenose dolphins’ sound receiving directivity index increased with frequency and followed the equation (Au, 1993): (18)${\displaystyle D{I}_{\mathrm{receiving}}=16.94\times {log}_{10}\left(f\right)-14.69}$ where DI was directivity index, and f was frequency in kHz. If this was also applicable and can be extrapolated for Sousa, then the directivity index for signals with frequencies lower than 7.37 kHz will be close to 0.

Previously, 4 cm and 6 cm radius circular piston projectors were applied to model the outgoing signal transmission beam pattern for spinner dolphin (Lammers & Au, 2003) and white-beaked dolphin whistles (Rasmussen et al., 2006), respectively.

The same method (Au, 1993; Wang et al., 2015) was adopted in this study, and the modeled beam pattern has a directivity index of 3 dB for a 4 cm circular piston model and 6 db for a 6 cm circular piston model (Fig. S2), suggesting directionality would also have a minor effect on the results reported here. These are only theoretical results, however, and need to be further corroborated by empirical measurements.