Social calls in humpback whale mother-calf groups off Sainte Marie breeding ground (Madagascar, Indian Ocean)

Study area

We collected acoustic data during winters 2013, 2014, 2016, and 2017 in the coastal waters off Sainte Marie island, Madagascar, South Western Indian Ocean (between latitudes 17° 19′ and 16° 42′ South, and longitudes 49° 48′ and 50° 01′ East), where mother-calf pairs come in these relatively calm and shallow waters (Trudelle et al., 2016).

Data collection and tagging procedure

We used Acousonde 3B (www.acousonde.com) attached to females or calves via four suction cups, a non-invasive attachment system, to record sounds. We deployed tags from a 6.40 m rigid motorboat using a 5 m handheld carbon-fiber pole. The boarding team consisted of staff experienced in successfully approaching mother-calf pairs: one operator (boat pilot), photographer, note-taker, and tagger. While the boarding team may change between outings, the team composition was always respected and each team member held one role within a whole outing. The tags were placed on the back, near the dorsal fin of the animal. Calves were tagged using one of the two approaches described in Stimpert et al. (2012) and in Huetz et al. (2022) in order to minimize disturbance to the mother-calf pair. Tagging efforts were terminated if the pair displayed avoidance behavior or if the calf was not successfully tagged within 30 min. Immediate behavioral response of the animals to tagging was recorded as in Stimpert et al. (2012). All mother-calf pairs were photo-identified to avoid double-sampling within the calving season. Tagged animals were not followed after tag deployment to avoid any further disturbance of their behavior. After tag deployment or an aborted attempt, the boat slowly moved away in the opposite direction of the mother-calf pair. The tag was retrieved after few hours or the following day, when it detached itself from the animal (usually as a consequence of rubbing against the mother, surface active behavior, etc.). The Fisheries Resources Ministry, Madagascar, fully approved all experimental protocols (Research and Collect permits #44/13-MPRH/SG/DGPRH, #43/14-MRHP/SG/DGRHP, #28/16-MRHP/SG/DGRHP, and #26/17-MRHP/SG/DGRHP). This present study complies with the European Union Directive on the protection of animals used for scientific purposes (EU Directive 2010/63/EU).

For each spotted group with mother-calf pair, approaches similar to touristic boats, complying with Madagascar’s Code of Conduct for whale watching activities (Inter-ministerial decree March 8th, 2000), were adopted: the boat’s speed was reduced gradually at 800 m distance from the spotted group. The observation area for mother-calf groups was set at a 200 m radius around them. We used this distance to observe the groups before tag deployment. We noted the group composition: lone mother-calf pair (MC) or mother-calf pair accompanied by one or several escorts (MCE). Behavioral observations and photo-identification were obtained concurrently for each group, and the calf’s relative age was estimated using the angle of furling of its dorsal fin (Cartwright & Sullivan, 2009; Huetz et al., 2022): C1 (neonate)–calf presenting some folds, scars, and skin color that tends to be light grey dorsally and white ventrally and with less than 44° dorsal fin furl (Faria et al., 2013), C2: very young but non-neonate calves having more than 45° but less than 72° dorsal fin furl, and C3: older calves that have unfurled dorsal fins (>72°). Depending on the opportunity and on how the group behaved, we tagged the calf or the mother using either a passive or active approach, as described in Stimpert et al. (2012) and Huetz et al. (2022).

We did not follow the animals after tag deployment to avoid further disturbance of their behavior. This means that change in group composition during the data recording is unknown. MC could be joined by other escort(s) or escort(s) in MCE might have left. Our tagging processes lasted 21 min on average and never exceeded 30 min (i.e., our maximum duration to tag a whale). Beside the mother-calf recordings, we also punctually recorded several male songs using a towed Aquarian H2a hydrophone deployed from the boat (engine off).

Vocal repertoire and acoustic analysis

We downloaded the audio files from the tags as *.MT files, converted them to *.WAV format using GoldWave software (GoldWave Inc., St. John’s, Canada), and analyzed them using Avisoft SASLab Pro version 5.207 (Avisoft Bioacoustics, Nordbahn, Germany). We produced spectrograms of the acoustic recordings using 1,024-point Fast Fourier Transform, 75% overlap, and Hamming window.

Social calls that were clearly audible and distinguishable were aurally and visually identified, classified, and then compared with social call catalogues available in the literature (Dunlop et al., 2007; Dunlop, Cato & Noad, 2008; Zoidis et al., 2008; Stimpert et al., 2011; Rekdahl et al., 2013, 2017; Fournet, Szabo & Mellinger, 2015; Epp, Fournet & Davoren, 2021; Indeck et al., 2021). Call types qualitatively similar to call types described in these catalogues (reference catalogue hereafter) were given the same name. New names were assigned for the remaining call types onomatopoeically, as the behavioral context or biological function of these new call types is still unknown and the main acoustic structure can be common to several call types (e.g., several call types can be low-frequency upsweep calls). Distant song units could not be confused with social calls in our dataset as songs can be easily recognized by their well-organized temporal pattern. Additionally, we used 16 of the recorded male songs to determine if some social calls were similar to males’ song units.

For each identified social call, we measured nine temporal and spectral characteristics (Table 1). Similar to Dunlop et al. (2007), we categorized social calls as either low-frequency sounds (LF), mid-frequency harmonic sounds (MF), high-frequency harmonic sounds (HF), amplitude modulated sounds (AM), noisy and complex sound (NC), or pulsed sounds (PS). LF corresponds to calls with peak frequency below 160 Hz. MF corresponds to calls with a peak frequency ranging from 170 to 550 Hz. HF corresponds to calls with a peak frequency above 700 Hz. AM corresponds to sounds consisting of a combination of long harmonic and amplitude modulated components with peak frequency ranging from 20 to 300 Hz. NC corresponds to broadband calls or harmonic calls with additional noise-like features. PS corresponded to low-frequency sounds repeated rhythmically. Representative spectrograms (Hamming window, FFT window size: 1,024 pts, 90% overlap) for the identified social calls were generated using the Seewave package (Sueur, Aubin & Simonis, 2008) in R software (R Core Team, 2021).

To obtain an overview of the intensity of sounds recorded in mother-calf groups, we measured the received level (in dB re 1μPa RMS) of the most common and aurally and visually easily identifiable call types using the Root Mean Square (RMS) function in Avisoft SASLab Pro. We considered only 10 good quality calls with a signal-to-noise ratio above 10 dB and without overlap for each selected call type. Using such criteria, calls selected for such measurements are likely produced by individuals constituting the focal mother-calf group (i.e., the mother, the calf, and possibly the escort if present). Mother-calf pair don’t gather with other pairs, so the only calls detected from our recordings can only come from individuals of the focal groups. Songs are highly different, so no confusion could be made. A similar procedure was used previously on humpback whale mother-calf pairs in Western Australia (Videsen et al., 2017).

Statistical analysis

To validate our aural-visual classification of calls by categories and by types, we performed random forest (RF) analyses (Thiebault et al., 2019; Indeck et al., 2021) using the randomForest (Liaw & Wiener, 2002) package in R. We calculated the global accuracy of the RF classification of calls by categories and by types (defined as accuracy = 1 − OOB, where OOB stands for out-of-bag error, the misclassification error rate) and computed the Gini index, which gives the importance of the variables used for the classification. We only included acoustic variables that were primarily measurable for most of the calls: Dur, Fmax, Q25, Q50, Q75, and PR. We included only call types for which we had at least six exemplars. The number of variables to be randomly selected at each split was set at 2 (as we had only six variables), and the number of trees grown was set at 500. We used a balanced RF design to maintain equal sample sizes of each category or type in the classification and avoid the over-representation of the most represented classes (Chen, Liaw & Breiman, 2004). In this design, each tree of the RF is built with the same number of calls per category or per type (i.e., the smallest number of calls for a given category or type).

Tag deployments

We performed 62 successful tag deployments (35 on calves and 27 on mothers) during the four years of data collection. Acoustic data were usable for 21 deployments (21 different mother-calf groups): seven deployments on mothers and 14 deployments on calves (Table 2). The other deployments were not analyzed due to the high background noise level or their short durations (less than 30 min). Background noise was mainly present when the tag was placed in a higher position close to the dorsal fin and thus often out of the water, especially on calves for which surfacing activities occurred very often. Of the 21 studied groups, we identified 12 as MC and nine as MCE. Three groups had a C1 calf, three had a C2 calf, and 15 had a C3 calf. We detected social calls in all of the studied groups.

Social call classification

A total of 2,033 social calls were clearly distinguishable. Aural-visual characteristics allowed the classification of these calls into 30 call types representing five of the six main categories suggested by Dunlop et al. (2007) (Table S1): low-frequency sounds (LF), mid-frequency harmonic sounds (MF), high-frequency harmonic sounds (HF), amplitude modulated sounds (AM), and pulsed sounds (PS). We did not find any call corresponding to the noisy and complex sound (NC) category. Eleven call types were named after qualitatively similar calls in the reference catalogue we used and 17 were given new name as they appeared qualitatively different (Table S1). Audio examples of call types are provided (Audio S1).

Low-frequency sounds (LF)

LF was the most represented category (46%, N = 925/2,033). Eleven call types were within the LF category: “100 Hz sound”, “bass”, “boom”, “gru”, “snort”, “burp”, “guttural”, “thowp”, “wop”, “bark”, and “drum” (Fig. 1). “Bass” and “wop” (Figs. 1B and 1I respectively) were the most common LF calls (heard in eight groups each), followed by “100 Hz” and “thowp” (Figs. 1A and 1H) (seven groups), “gru” and “snort” (Figs. 1D and 1E) (six groups), and by “boom” (Fig. 1C) (four groups). The remaining LF calls were rare (heard in two groups for “drum”, Fig. 1K, and only one group for “guttural”, “burp”, and “bark”, Figs. 1G, 1F and 1J respectively). “Burp”, “bark”, and “drum” were only heard in MC groups. The remaining LF calls were heard in both MC and MCE groups. “Bass” (Fig. 1B) was a harmonic sound with a fundamental frequency below 40 Hz on average, and can sometimes be masked by background noise. “Wop” (Fig. 1I) and “thowp” (Fig. 1H) were brief harmonic upsweep sounds similar in frequency but different in duration. The “100 Hz” call (Fig. 1A) was a long, relatively flat call. “Gru” (Fig. 1D) was a short harmonic sound like “snort” but with more spaced harmonics. “Snort” (Fig. 1E) can be produced in sequences. “Boom” (Fig. 1C) was a harmonic sound produced either in sequence or alone. “Boom” was frequently produced in series with “100 Hz” following a well-defined order (100 Hz–boom–boom–100 Hz). “Guttural” (Fig. 1G) appeared to be a “composite call” consisting of non-overlapping “gru” and “heek” (MF) not separated with a silence. “Drum” (Fig. 1K) was a very short call always produced in series, and “burp” (Fig. 1F) was a short harmonic sound with several close harmonics. “Bark” (Fig. 1J) was a short harmonic sound with ascending frequency modulation.

Mid-frequency harmonic sounds (MF)

MF was the second most represented category (35%, N = 718/2,033). Eight call types were within the MF category: “groan”, “downsweep”, “woohoo”, “trumpet”, “heek”, “whoop”, “wiper”, and “creak” (Fig. 2). “Heek” was the most common MF call (heard in nine groups), followed by “whoop” (seven groups), “downsweep” and “woohoo” (five groups), and by “trumpet” (three groups). “Groan”, “wiper”, and “creak” were uncommon (heard in one group only) and were only heard in MC groups. The remaining MF calls were heard in both MC and MCE groups. “Heek” (Fig. 2E) was a short MF call with a variable frequency modulation pattern (ascending, descending, or modulated). In some instances, heek was produced with “gru” (LF), with a short silence separating the two vocalizations to constitute a “combined call”. “Whoop” (Fig. 2F) was a long upsweep call starting with a flat part and fast ascending frequency. “Downsweep” (Fig. 2B) was a long call showing a descending frequency slope with well-spaced harmonics. “Woohoo” (Fig. 2C) was a long-duration call (i.e., several seconds) showing a variable frequency-modulated pattern. “Trumpet” (Fig. 2D) was a call produced alone or associated with “gru” or “slight snort”. “Downsweep”, “woohoo”, and “trumpet” were found in humpback whale songs recorded around the study site. “Groan” (Fig. 2A) was a long harmonic call. “Wiper” (Fig. 2G) was a short harmonic sound with a U-shape frequency modulation pattern, always produced in series (four to five repetitions) but with a random temporal pattern. “Creak” (Fig. 2H) was a composite call constituted by two different non-overlapping successive sounds without silence, and it showed the widest frequency bandwidth.

High-frequency harmonic sounds (HF)

HF was the third most represented category (10%, N = 202/2,033). Two social calls were within this category: “squeak” and “ascending shriek” (Fig. 3). Both were quite common: squeak (Fig. 3A) was heard in seven groups, and “ascending shriek” (Fig. 3B) was heard in four groups. “Squeak” and “ascending shriek” were heard in both MC and MCE groups. “Squeak” was a very short call with frequencies above 1 kHz, and “ascending shriek” was one of the longest social calls with the highest frequencies amongst all. Both call types were also found in humpback whale songs recorded around the study site.

Amplitude modulated sounds (AM)

AM was the fourth most represented category (8%, N = 167/2,033). Five social calls were within the AM category: “door”, “whine”, “trill”, “bug sound”, and “AM grunt” (Fig. 4). “AM grunt” (Fig. 4E) and trill (Fig. 4C) were the most common MF calls (heard in six and five groups, respectively). The remaining calls were relatively uncommon since they were heard only in one group each. “Door” (Fig. 4A) was heard in an MC group. “Whine” (Fig. 4B) and bug (Fig. 4D) were found in MCE groups only. “Trill” and “AM grunt” were heard in both MC and MCE groups. “Trill”, “door”, “whine”, and “bug” are long calls ranging from 1 to 5 s duration with a peak frequency ranging from 100 to 400 Hz. They were produced in bouts of random durations. “AM grunt” was short and, while commonly produced alone, it was sometimes associated with LF calls such as “gru” or “snort”. “Whine”, “trill”, and “bug” were also found in humpback whale songs recorded around the study site.

Pulsed sounds (PS)

PS was the least represented category (1%, N = 21/2,033). Four social calls were classified in this category: “fry”, “bubble sound”, “moped”, and “gloop” (Fig. 5). These calls consisted of a repetition of very short, low-frequency sounds, and they were pretty uncommon. “Fry” (Fig. 5A) was heard in two groups and the remaining PS calls were heard in only one group each. “Fry” was heard in both MC and MCE groups. “Bubble sound” (Fig. 5B) was heard in a MC group. “Moped” (Fig. 5C) and “gloop” (Fig. 5D) were only found in MCE groups.

Aural-visual classification validity

For the validation of our classification RFs, 757 social calls representing 17 call types were used for the analyses (social calls for which Dur, Fmax, Q25, Q50, Q75, and PR were all measured). These analyzed social calls exclude call types for which we had less than six exemplars. The analyzed call types were: 100 Hz, bass, boom, gru, snort, burp, thowp, wop, downsweep, woohoo, trumpet, heek, whoop, squeak, ascending shriek, trill, and fry. The RF showed a global accuracy of prediction of 93% for classifying the calls into the five defined main categories (OOB error rate = 7%, Table 3). The acoustic variables showing the highest importance for the classification were Q25, Fmax, and PR (Gini index: 10.24, 10.18, and 6.93 respectively, Table 3). Most call categories showed low individual classification error rates. For classifying the calls by types, the RF showed a global accuracy of prediction of 77% (OOB error rate = 23%, Table 4). The acoustic variables showing the highest importance for classification were Fmax, Q25, and Dur (Gini index: 32.10, 31.78, and 26.94 respectively, Table 4). Of the 17 calls included in the RF analysis, only four call types showed exceptionally high error rates (≥50%): gru, snort, thowp, and wop. These call types may share features with other call types (short, harmonic, and low-to-medium frequency calls).

Received level

Five call types were selected for the calculation of the received level: three LF (“100 Hz”, “bass”, and “boom”), one AM (“trill”), and one MF sound (“heek”). The received level ranged from 132 to 154 dB re 1μPa RMS with an average of 141 dB re 1μPa RMS (N = 50, 10 per call type; Table 5).