First Steps towards the Automated Detection of Underwater Vocalisations of Grey Seals (Halichoerus grypus) in the Blasket Islands, Southwest Ireland

Abstract

Underwater vocalisations of grey seals (Halichoerus grypus) were recorded by static acoustic monitoring (SM2M, Wildlife Acoustics) in the vicinity of a colony located at White Strand beach on Great Blasket Island, southwest Ireland during the pre-breeding and breeding seasons. Grey seal vocalisations were first classified across nine different categories based on aural and visual characteristics of the spectrograms, providing an acoustic repertoire for grey seals. This classification was further investigated by applying a classification tree analysis, resulting in five of the initial nine groups being selected. Furthermore, a comparison of two common approaches for the detection and extraction of vocalisations from acoustic files was done using the software Raven Pro and PAMGuard. The outputs of this study will present an essential first step towards the development of a protocol for underwater acoustic monitoring of grey seals in Irish waters and elsewhere.

1. Introduction

Acoustic techniques are widely applied in population monitoring programs of marine mammals and particularly of cetaceans [1,2]. In comparison with visual methods, acoustic monitoring provides many advantages, allowing the continuous collection of acoustic data over extensive periods of time and large areas, even under adverse environmental conditions and during darkness [1].

Although acoustic monitoring of cetaceans in Irish waters has been successfully conducted for several years [1,3,4,5,6,7,8], studies that use acoustic techniques to monitor pinniped species have not yet been conducted in Ireland.

Vocal behaviour is thought to have a key function in communication [9], mating strategies [10,11,12], and pup recognition [13] in some pinniped species. As vocal animals, pinnipeds have the potential to be subject to underwater acoustic monitoring programmes. However, the vocal activity of some species of pinnipeds is mostly restricted to the breeding period, as is the case for harbour seals (Phoca vitulina, [12]) and bearded seals (Erignathus barbatus, [14]). On the other hand, vocalisations of grey seals have been recorded outside the breeding season [15,16]. Although less common than acoustic studies of cetaceans, several acoustic monitoring studies of pinnipeds have also been conducted. Some examples are focused on the assessment of (1) the vocal repertoire of Antarctic pinnipeds during the breeding season [9]; (2) the occurrence of Ross seals (Ommatophoca rossii) and leopard seals (Hydrurga leptonyx, [17]); and (3) the temporal and spatial distribution of bearded seals (Erignathus barbatus, [18]) and ringed seals (Pusa hispida, [15]).

Under the EU Renewable Energy Directive (2018/2001) and the Offshore Renewable Energy Development Plan, a high increase in offshore wind farm development and ocean energy devices is foreseen in Irish waters, which will lead to an increase in anthropogenic activities in coastal and offshore areas and an associated increase in anthropogenic noise. Underwater noise is included as descriptor 11 in the Marine Strategy Framework Directive (008/56/EC), as it has the potential of adversely affecting marine organisms, including cetaceans and pinnipeds [19,20]. Acoustic monitoring is a powerful tool for assessing impacts of anthropogenic noise on these species and the implementation of mitigation measures [21]. Furthermore, underwater acoustic monitoring provides valuable information on the temporal and spatial distribution of marine mammals, including seasonal and migration patterns [2], although it should be noted that not all marine mammal species vocalise year-round [12,14] and this could be a potential restriction for PAM. This is key information to identify important habitats for these protected species and for the design and implementation of conservation and management plans for the species and their habitats, and will ultimately contribute to the designation process of Marine Protected Areas (MPAs).

In order to detect and identify pinniped species using acoustic monitoring techniques, their vocalisations must have been previously described and classified. Several studies have provided repertoires of vocalisations for different species of seals, including bearded seals [22,23], crabeater seals [24], harbour seals [25], hooded seals (Cystophora cristata, [26,27]), leopard seals (Hydrurga leptonyx, [24]), Mediterranean monk seals (Monachus, [28]), and Weddell seals (Leptonychotes weddellii, [29]).

Other acoustic research on pinnipeds has focused on assessing (1) diel and seasonal patterns in vocalisation rate and call types of bearded seals [30,31,32], crabeater seals (Lobodon carcinophagu, [33]), harbour seals (Phoca vitulina, [34]), ribbon and ringed seals [31], and Weddell seals (Leptonychotes weddellii, [35,36]); (2) vocal behaviour during breeding periods in harbour seals [12]; (3) differences in vocalisations between habitats in bearded seals [32], between areas in bearded [23] and cape fur seals [37], and between populations of ribbon seals [38]; (4) and the effects of anthropogenic noise on vocal behaviour of harbour and bearded seals [12,39].

There have not been many studies focusing on grey seal underwater vocalisations in recent years. Previous acoustic studies on grey seals have attempted to classify in-air [16,40] and underwater ([16,40,41,42,43,44];

Table 1. Summary of studies that have focused on adult grey seal underwater vocalisations.

Animals	Season	Location	Acoustic Repertoire	Findings	References
Captive	-	-	Clicks	-	[41]
Captive	-	-	Clicks Humming or moaning (corresponds with roar in [44])	Increase in vocalisations during breeding	[42]
Captive/ wild	-	Basque Island, Nova Scotia (Canada)	Clicks Wail (corresponds with roar in [44]) Jackhammer sound (corresponds with trots in [44] emitted by males	-	[43]
Wild	Breeding	Gulf of St. Lawrence (Canada)	7 call types: Mainly guttural rups, rupes (freq = 100–3000 Hz), and growls (freq = 100–500 Hz) Also clicks (3000 Hz), knocks, trots, and roars	Increase in call rate with social interaction throughout breeding period Variation in call types over breeding period and between day/night No differences in call rate between day/night	[44]
Wild	Breeding	Isle of May and Abertay Sands (Scotland)	10 call types Guttural rups, type 2, type 4, and guttural rupes accounting for 70% of calls	Variation in call types throughout the breeding period	[16]
Captive	-	-	3 groups of vocalisations: S1: tonal calls with harmonics S2: burst pulses S3: pulses with harmonics	-	[40]
Wild	Breeding	Farne Islands, England, UK	Percussive claps a	Produced by males, likely in male–female or male–male interactions	[45]
Wild	Breeding/non-breeding	Gulf of Riga	Based on [44]	Recorded and identified grey seals	[15,48]

^a Percussive claps are not vocalisations; previously mistakenly described as knocks by [44] and type 8 calls by [16].

) vocalisations. One study [44] provided the first thorough underwater repertoire of wild grey seals from the Gulf of St. Lawrence during the breeding season. The repertoire was composed of seven vocalisation types, with guttural “rups” and “rupes” being the most commonly recorded. Similarly, [16] provided an underwater repertoire of wild grey seals from the Isle of May and Abertay Sands in Scotland during the breeding season. Of the 10 underwater vocalisation types presented, five had already been described, with guttural rups and rupes the most frequent vocalisations recorded. In the same study, [16] also provided an in-air repertoire composed of six vocalisation types. One of the vocalisations described in both studies [16,44], knocks, was recently discovered to be a percussive sound produced by male grey seals with their forelimbs during the breeding season [45]. Recently, [40] focused on the in-air and underwater sound production mechanisms of captive grey seals and presented a classification of vocalisations based on acoustic and temporal parameters composed of three different groups: tonal calls with harmonics, burst pulses, and pulses with harmonics. Some research has also focussed on calls emitted by grey seal pups [13,46,47], revealing significant variation in calls between individual pups. More recently, [48] successfully recorded and identified grey seals in the Gulf of Riga, detecting three vocalisation types and percussive sounds. However, no studies have been published to date assessing the potential of available acoustic techniques for use in monitoring programmes for this species.

The aim of this study was to record and classify underwater vocalisations of grey seals in the Blasket Islands Special Area of Conservation (SAC) during pre-breeding and breeding seasons, providing a baseline acoustic repertoire for this species in Irish waters. The aim was also to compare two common approaches for the detection and extraction of vocalisations from acoustic files using the software Raven Pro 1.5 (Cornell Lab of Ornithology, Ithaca, NY, USA) and PAMGuard version 2.01.05 [49] in order to provide recommendations for grey seal acoustic monitoring in Irish waters.

2. Materials and Methods

2.1. Study Area

This study was conducted in the area surrounding White Strand beach on Great Blasket Island, which is located within the Blasket Islands SAC, County Kerry, southwest Ireland (Figure 1). This archipelago hosts one of the most important grey seal breeding colonies in Ireland [50] and is one of the ten designated SACs for this species in Ireland [51], as grey seals are a protected species listed under Annex II of the European Union’s Habitats Directive (92/43/EEC).

2.2. Acoustic Data Collection

Underwater vocalisations of grey seals were recorded with a song meter SM2M marine recorder (Wildlife Acoustics, Inc.), deploying at an approximate distance of 100 m from a grey seal colony between White Strand beach on Great Blasket Island and the island of Beginish (at 52°6′31.398″ N; 10°30′46.722″ W, Figure 1). The SM2M was placed at approximately 5 m below the surface, and attached to a weight and a buoy to keep it in the same location and increase buoyancy. Underwater acoustic recordings were obtained at a sampling rate of 384 kHz and 16-bit resolution (hydrophone sensitivity of −165 dB re 1 V/µPa and frequency response from 2 Hz to 30 kHz +/− 2dB of rated sensitivity) during pre-breeding and breeding seasons (defined as in [52]), from 10 to 14 June 2019 (5 days) and from 16 to 19 September 2019 (4 days). The SM2M was programmed prior to deployments to record during the first 30 min of every hour, i.e., at intervals of 30 min. Recordings were obtained during day and night and amounted to a total of 34 h and 30 min in June and 9 h and 50 min in September. Problems regarding sealing of the device during the second deployment led to unpredictable recording times and irregular recording durations.

2.3. Qualitative Classification of Grey Seal Calls

Underwater recordings as .wav files were imported into the software Raven Pro 1.5 (Cornell Lab of Ornithology, Ithaca, NY, USA). Software settings were as follows: page size 150 s, page increment 90 percent, and step increment 10 percent. The sample rate of the spectrogram was 384 kHz. Spectrogram brightness, contrast and window size were set to 57, 70, and 4000, respectively. The type of window selected was Default 1.3 Power. Time resolution of spectrograms was set at 0.02 s and frequency resolution was set at 3000 Hz, but it was modified when needed. Time and frequency resolution were selected based on previous studies that reported grey seal vocalisations [16,44], and the rest of the applied settings were selected based on own judgment after some preliminary analyses. After a preliminary examination of the spectrograms, acoustic data were processed and grey seal vocalisations were selected based on aural and visual features such as vocalisation shape, as in previous studies aimed at classifying vocalisations of pinnipeds [16,23,25,27,44]. A total of 19,523 grey seal vocalisations (16,745 in June and 2778 in September) were extracted and manually classified by the same person (MPT) into nine different categories (Figure 2,

Table 2. Classification of underwater vocalisations of grey seals recorded at the Blasket Islands (SAC). Vocalisation types, description, and comparison with previous studies are presented.

Call Type	Description	Comparison with Previous Studies
1	Multiple element vocalisation. Mostly emitted in series ranging from 1 to 31 elements. Relatively constant frequency, presenting a frequency upsweep at the end on some occasions.	Rup [44] Type 1, guttural rupe [16] Rup [48]
2	Presents the lowest frequency of all vocalisation types identified. Relatively constant frequency. Presence of harmonics on some occasions.	Growl [44] Type 9, growl and/or type 10 [16]
3	Two-part vocalisation, both audible: First component presents a relatively constant frequency. Second component presents a sharp frequency upsweep at the end of the call. Similar to type 4 vocalisations. Similar sound to type 1 vocalisations.	Type 6 [16] Second part of the vocalisation similar to rup type [44] S3 [40]
4	Ascending shape. Sharp frequency upsweep at the end of the vocalisation.	Type 4 [16] Rup [44] S3 [40]
5	Similar to type 1 vocalisations. Multiple element vocalisation. Most times emitted in series ranging from 1 to 15 elements. Relatively constant frequency with an upsweep at the end of each element on some occasions. Can present slightly variable shapes. Main differences with type 1 vocalisations are that vocalisation presents longer duration of each element and presents a lower minimum frequency.	Rup [44] Type 1, guttural rupe [16] S3 [40]
6	Series of short pulses. Variable in shape, frequency, and duration. Sometimes similar to type 4 vocalisations, presenting a sharp frequency upsweep at the end of the call, but different in sound.	Trrots, pulses, clicks [44] Type 3 [16] S2 [40]
7	Opposite to type 4 vocalisations. Descending shape. Presents a frequency downsweep at the end of the call.	This vocalisation type has not been described before.
8	Two-part vocalisation, both audible: First component presents a sharp frequency upsweep at the end of the call. Sometimes similar to type 4 vocalisation. Second component presents a longer frequency downsweep at the end of the call. Similar to type 7 vocalisations, but harmonics are often present in the second element.	Guttural rupe [44] Type 5, rupe [16] Rupe [48] S3 [40]
9	Tonal calls, moans. Sometimes similar to type 2 calls, but different in sound. Presents the highest frequency of all call types identified due to the presence of harmonics, up to >4000 Hz.	Mumming or moaning [42] Wail [43] Roar [44] Type 7, moan [16] S1 [40] Moan [48] S1 [40]
Claps a	Short-duration sounds Often in series of 1–5 [16,44,45] High frequencies, up to 10–15 kHz [16,44,45] Higher frequencies recorded during this study, up to >50 kHz	Knocks [44] Percussive claps [45]

^a Percussive claps are not vocalisations, previously mistakenly described as knocks by [44] and type 8 calls by [16].

) according to the classifications made by [16,44]. The general approach was that, after having established different categories, vocalisations were then selected. Multiple element vocalisations (e.g., types 1 and 5; Figure 2; Table 2) were selected as one single call, provided that all elements presented similar aural and visual characteristics. For two-part calls (e.g., 3 and 8; Figure 2; Table 2), each selected call was composed of only two elements.

2.4. Quantitative Classification of Grey Seal Calls

For each selected vocalisation, parameters including total call duration, low frequency, high frequency, peak frequency (frequency at which peak power occurs), frequency range, average power density, and peak power density (

Table 3. Acoustic variables extracted for grey seal underwater vocalisations (n = 19,523) in the Blasket Islands (SAC). Means ± SD are presented for low, high, and peak frequency; frequency range, and duration for each vocalisation type are also presented. Proportion of vocalisations for each type is also presented.

Vocalisation Type	Low Frequency (Hz)	High Frequency (Hz)	Peak Frequency (Hz)	Frequency Range (Hz)	Call Duration (s)	N (%)
1	187.90 ± 57.77	1324.88 ± 235.22	878.65 ± 262.16	993.85 ± 268.06	0.59 ± 0.74	30.96
2	196.91 ± 71.62	921.27 ± 189.33	703.91 ± 184.40	727.90 ± 176.13	0.76 ± 0.87	7.51
3	127.91 ± 37.86	1337.69 ± 222.62	851.03 ± 260.07	1090.21 ± 290.45	0.41 ± 0.12	4.87
4	138.31 ± 49.96	1235.51 ± 211.26	831.54 ± 220.43	1000.65 ± 245.73	0.26 ± 0.14	5.96
5	121.52 ± 28.81	1308.41 ± 240.51	877.96 ± 264.41	1045.25 ± 296.37	0.59 ± 0.42	20.50
6	148.55 ± 50.71	1225.10 ± 301.58	843.37 ± 250.57	966.93 ± 304.20	0.45 ± 0.41	10.59
7	158.33 ± 52.04	1112.04 ± 276.76	789.73 ± 289.65	875.11 ± 231.05	0.28 ± 0.14	2.10
8	157.64 ± 55.10	1237.30 ± 248.69	841.62 ± 215.27	975.67 ± 278.54	0.43 ± 0.19	17.03
9	163.24 ± 122.15	4062.14 ± 1723.55	1125.29 ± 1061.03	3491.69 ± 2119.47	0.93 ± 1.01	0.49

) were extracted into selection tables for each acoustic file. Most of these parameters have been used in previous studies aimed at classifying marine mammal sounds, e.g., [23,44].

A further quantitative analysis was performed to validate the perceptual classification. This consisted of a non-parametric classification tree analysis using the package rpart [53] in RStudio, v. 3.6.2 [54]. This analysis has advantages over similar methods such as principal component analysis or discriminant function analysis, including that it can be applied to abnormally distributed data or dependent data [55]. Although it was not possible to estimate the number of seals in the water within the hydrophone range at the time acoustic files were recorded, meaning that different vocalisations could have been emitted by the same individual, this classification tree analysis allows for potential pseudo-replication [27].

The analysis consisted of 3 main steps: (1) creation of a classification tree—all the acoustic parameters were included in the analysis, and data were split at each node using only one variable and based on the Gini index in order to minimise misclassification (Breiman et al., 1984); (2) validation of results using cross-validated error values and the complexity parameter (Appendix A: Figure A1;

Table A1. Cross-validation results obtained for the classification tree analysis.

	CP	nsplit	rel Error	x Error	xstd
1	0.13	0	1	1	0.005
2	0.04	1	0.87	0.87	0.005
3	0.02	2	0.83	0.83	0.005
4	0.01	6	0.74	0.75	0.005
5	0.01	7	0.73	0.75	0.005

); and (3) pruning was applied to avoid model overfitting and to obtain the best possible tree.

The method of analysis applied here has been used widely in previous research that aimed to classify calls emitted by different pinnipeds [23,25,27] and cetaceans [56,57].

2.5. Automatic Detection of Grey Seal Vocalisations

In order to automatically detect grey seal vocalisations, a section of the underwater acoustic recordings as .wav files collected in the Blasket Islands was imported into the software PAMGuard [49], available via http://www.pamguard.org. Acoustic files in which a high number of grey seal vocalisations had been previously detected manually using the software Raven Pro 1.5 (

Table 4. Underwater acoustic recordings from the Blasket Islands selected for PAMGuard analysis.

Date	Time	Manual Detections a	Season	Diel	Tide	Recording Time (min)
11 June 2019	23:00	607	Pre-breeding	Night	Flood	30
12 June 2019	21:00	675	Pre-breeding	Day	Flood	30
13 June 2019	18:00	732	Pre-breeding	Day	Ebb	30
13 June 2019	19:00	705	Pre-breeding	Day	Slack low	30

^a Vocalisations detected manually using the software Raven Pro 1.5.

) were selected.

To process the acoustic data for the presence of grey seal vocalisations, different modules were added: a decimator to down-sample the raw data, a low frequency fast Fourier-transformation (FFT) module with an FFT length of 512 and FFT hop of 256, and a whistle and moan detector set to a maximum frequency of 5 kHz. Afterwards, data were explored in PAMGuard viewer mode to verify detections.

To manually verify these detections and calculate the number of false positives, the first 10% of vocalisations of each processed file was selected, i.e., files were manually inspected (Figure 3) until the time when 10% of detections had been verified. Subsequently, these were compared with manual detections from Raven Pro 1.5 for the same periods of time.

3. Results

3.1. Underwater Acoustic Repertoire

A total of 19,523 vocalisations (16,745 recorded in June and 2778 in September 2019) were manually selected and classified into nine different vocalisation types (Figure 2; Table 2) based on aural and visual spectrogram characteristics.

3.2. Classification Tree Analysis

All the acoustic parameters extracted (Table 3) were included in the classification tree analysis. In order of importance, chosen automatically by the method, the variables used in tree construction (i.e., in the classification process) were low frequency, high frequency, peak power density, and total call duration. The output of this classification analysis showed five different vocalisation types (Figure 4). These were types 1, 2, 4, 5, and 8, while vocalisation types 3, 6, 7, and 9, identified by manual analysis, did not appear in the output, suggesting that some vocalisation types could be included as subtypes in the same group.

3.3. Assessment of Acoustic Techniques for Grey Seal Detection

The four selected files to be processed in PAMGuard accounted for 2 h of recordings (Table 4). The number of total automatic detections made by the software for these files were 1962, 982, 437, and 1493. After 10% of these detections had been verified, mean proportion (± SD) of correct detections was 64.89 ± 8.47% (

Table 5. Comparison of the number of underwater acoustic vocalisations manually detected using Raven Pro 1.5, with the automatic detections obtained from the PAMGuard analysis.

Date	Time	Manual Detections a	Detections PAMGuard b	Correct Detections (%)	False Positives (%)	Detections PAMGuard c	PAMGuard Efficiency (%) d
11 June 2019	23:00	134	196	58.16	41.84	72	53.73
12 June 2019	21:00	146	98	76.53	23.47	47	32.19
13 June 2019	18:00	112	44	59.09	40.91	20	17.86
13 June 2019	19:00	230	149	65.77	34.23	71	30.87

^a Number of vocalisations manually detected in Raven Pro 1.5 until the time when 10% of PAMGuard detections had been reached. ^b 10% of automatic detections made by PAMGuard. ^c Number of correct PAMGuard automatic detections after subtracting the number of vocalisations detected more than once, as well as vocalisations that had not been selected by Raven. ^d PAMGuard efficiency (%) was calculated as (detections PAMGuard ^c/manual detections ^a) × 100.

), with 35.11 ± 8.47% (mean ± SD) accounting for false-positive detections (Table 5). False-positive detections were mainly due to noise caused by the mooring system and water movements.

In some cases, vocalisations were not clear enough to be manually classified into one of the vocalisation types using Raven Pro 1.5. Therefore, some automatic detections made by PAMGuard had not been previously selected manually in Raven. Furthermore, some individual vocalisations were automatically detected more than once by PAMGuard. This was particularly observed for the second part of vocalisation type 8, due to the presence of harmonics (see an example in Figure 3). This was taken into account when calculating PAMGuard detection efficiency, obtaining a mean value of 33.66 ± 14.86% (Table 5).

Number of vocalisations for each vocalisation type manually extracted using Raven and automatically detected by PAMGuard are presented in Appendix B (

Table A2. Number of vocalisations detected using Raven for each vocalisation type for the selected acoustic files until 10% of detections were verified.

		Vocalisation Types Detected in Raven
Date	Time	1	2	3	4	5	6	7	8
11 June 2019	23:00	18	7	18	1	23	19	1	47
12 June 2019	21:00	55	8	7	7	46	6	1	16
13 June 2019	18:00	47	8	3	6	29	6	1	12
13 June 2019	19:00	50	7	6	8	88	5	4	62
Total		170	30	34	22	186	36	7	137
%		27.33	4.82	5.47	3.54	29.90	5.79	1.13	22.03

and

Table A3. Number of vocalisations detected by PAMGuard for each vocalisation type for the selected acoustic files (corresponding to 10% of detections).

		Vocalisation Types Detected by PAMGuard
Date	Time	1	2	3	4	5	6	7	8
11 June 2019	23:00	5	13	12	4	7	20	0	39 a
12 June 2019	21:00	8	16	5	2	13	3	0	8 b
13 June 2019	18:00	0	7	1	0	14	0	1	3
13 June 2019	19:00	7	8	0	6	48	0	1	12
Total		20	44	18	12	82	23	2	62
%		7.60	16.73	6.84	4.56	31.18	8.75	0.76	23.57

^a Of the total, PAMGuard detected only the first part of 4 vocalisations and only the second part in 15 vocalisations, type 8. ^b Of the total, PAMGuard detected only the second part of 3 vocalisations, type 8.

). The types of vocalisations most commonly detected by PAMGuard were types 5 (31.18%), 8 (23.57%), and 2 (16.73%;

Table 6. Proportion of each vocalisation type detected using Raven and automatically detected by PAMGuard for the selected acoustic files (corresponding to 10% of PAMGuard detections).

		Vocalisation Types
		1	2	3	4	5	6	7	8
% Detected in Raven		27.33	4.82	5.47	3.54	29.90	5.79	1.13	22.03
% Detected by PAMGuard a		7.60	16.73	6.84	4.56	31.18	8.75	0.76	23.57

^a Total vocalisations automatically detected by PAMGuard that could be manually classified accounted for 82.13%. The rest could not be classified.

), while for the same acoustic files, the types most commonly detected using Raven were 5 (29.90%), 1 (27.33%), and 8 (22.03%).

The proportion of vocalisation types 3, 4, 5, 6, 7, and 8 detected by PAMGuard were similar to the proportions detected using Raven (Table 6). The proportion of type 1 vocalisations detected manually was higher (27.33%) than that automatically detected by PAMGuard (7.60%; Table 6). In contrast, the proportion of type 2 vocalisations detected by PAMGuard was higher (16.73%) than the proportion detected manually (4.82%; Table 6). When the first 10% of vocalisations of each processed file had been verified, we could identify these type 2 calls selected by PAMGuard, but due to high background noise, they had not been previously selected by Raven.

4. Discussion

Underwater vocalisations from grey seals recorded during pre-breeding and breeding seasons in the Blasket Islands SAC were firstly classified into nine different vocalisation types. An underwater acoustic repertoire for this species is provided in the present study, complementing previous studies conducted by [16,44], on which our classification was based.

Vocalisation types were described based on aural and visual spectrogram characteristics, and are similar to types previously described (Figure 2; Table 2). Type 1 has multiple elements, with relatively constant frequency, although presenting a sharp frequency upsweep in some vocalisations recorded. This type has been previously described as rups [44] and as type 1 [16]. Similarly to previous studies [16,44,48], this was the most commonly recorded call, accounting for 30.96% of all vocalisations. Initially, it had been suggested that this vocalisation type was produced by females during interactions with other individuals during the breeding season [44], although [16] observed that this vocalisation type was produced by males. In this study, this vocalisation was recorded during pre-breeding and breeding seasons.

Type 2 presented the lowest maximum-frequency component of all vocalisation types, accounting for 7.51% of all vocalisations recorded. It has been previously classified as growls [44] and as type 9 or type 10 [16], with both authors suggesting this vocalisation as being emitted only by males during the display of agonistic behaviours and being recorded during the breeding season. In this study, this vocalisation type was recorded during both pre-breeding and breeding seasons and could suggest the presence of breeding males.

Type 3 is a two-part vocalisation, where the first component presents relatively constant frequency and the second component a sharp frequency upsweep at the end. The second element is similar to type 4 vocalisations. It accounted for 4.87% of the vocalisations recorded and has been previously described only by [16] as type 6. Type 4 presents an ascending shape. It accounted for 5.96% of vocalisations. It has been previously described as rups [44] and type 4 [16].

Type 5 is similar to type 1 vocalisations. It accounted for 20.50% of vocalisations, and together with type 1, accounted for more than 50% of vocalisations recorded during this study. Types 1 and 5 could be classified as subtypes of the same vocalisation group, as type 5 is similar to rups and type 1 described in previous studies [16,44]. Vocalisation type 6 is described as a series of pulses or clicks. It accounted for 10.59% of all vocalisations and has previously been described as “troots” [44] and type 3 [16]. Type 7 presents a descending shape and accounted for 2.10% of all vocalisations. It has not been previously described as a call, although it is similar to the second component of type 8.

Type 8 is a two-part call. The first component is sometimes similar to type 4, as it presents an ascending shape, while the second component presents a frequency downsweep, similar to vocalisation type 7. It has a longer duration than the first component and harmonic components are often present. It accounted for 17.03% of all vocalisations and has been previously described as guttural rupes [44] and as type 5 by [16], and it has been suggested that it is emitted by females during interactions with other individuals [16,44].

Type 9 was the least commonly recorded vocalisation on the Blasket Islands. This was a tonal call, with harmonics up to a frequency higher than 4000 Hz. It has been widely reported and previously classified as moaning [42], wail [43], roar [44], and type 7 [16]. Together with types 1 and 8, this vocalisation type has been identified in the Gulf of Riga using acoustic monitoring [48]. One study [48] found that the type 9 vocalisation was the most commonly recorded during the summer months outside the breeding period, and [16] that this vocalisation type tended to become less common as the breeding season progressed. However, the type 9 vocalisation has been reported as being emitted by both females and males during social interactions and by wild and captive individuals [16,43]. On the Blasket Islands, grey seals were observed to be involved in social interactions, such as mating and fighting [52].

Recently, [45] recorded for the first time male grey seals in the wild during the breeding season making a percussive sound with their forelimbs. This sound had been previously recorded [16,44] and mistakenly classified as knocks, an underwater vocalisation type [45]. Like [44] in the Gulf of St. Lawrence, [45] also detected this sound during the grey seal breeding season in the Gulf of Riga. These percussive clap sounds were recorded during both pre-breeding and breeding seasons in the Blasket Islands during this study, where seals were observed hauled out in the area in large numbers during both periods, while [45] suggested that this behaviour could be related to mating strategies, which could be the case for our study area, as it is an important breeding site for grey seals.

The acoustic parameters low frequency, high frequency, peak power density, and total call duration were included in the classification tree analysis carried out in this study. Of all the vocalisation types described above, only types 1, 2, 4, 5, and 8 were selected by the classification tree analysis, based on their low frequency, high frequency, peak power density, and total call duration. The rest of the vocalisation types, i.e., 3, 6, 7, and 9, were excluded in the classification process. Although type 7 has not been previously reported and could be part of another vocalisation type, as it is similar to the second element in type 8, vocalisations 3, 6, and 9 are very distinctive and have been previously described (e.g., [16,42,44]). As the technique is refined across more acoustic datasets, this will allow for testing the accuracy of the categories selected. Differences in the classification of vocalisations presented could be due to differences between populations [23], although movements of grey seals have been recorded between Ireland and Scotland [58]. Local factors such as the presence of ecotourism activities when the study was conducted could also have had an impact on seasonal and diel patterns in vocalisations.

This study had some limitations, such as the short recording times of just a few days, the second recording period being conducted at the beginning of the breeding season, as well as the malfunction of the acoustic equipment during that period. Furthermore, the sex of individuals that produced vocalisations could not be determined, and it was also not possible to determine whether recorded vocalisations were emitted by a single or multiple individuals. However, given the grey seal numbers recorded in the area throughout the recording periods, it is very likely that overall, the vocalisations extracted are from different individuals.

Verifying acoustic signals such as clicks and whistles after using automatic detectors is often conducted in research studies focused on cetaceans [59,60]. However, assessing the efficiency of available non-invasive acoustic techniques for the detection and extraction of grey seal vocalisations has never been done before. In this study, a comparison between grey seal vocalisations manually detected using the software Raven Pro 1.5 and vocalisations automatically detected by the software PAMGuard was performed for a subset of the data recorded in the Blasket Islands. The efficiency of PAMGuard in detecting grey seal vocalisations was higher than 30%. PAMGuard performed well in detecting vocalisation types 3, 4, 5, 6, 7, and 8, as the proportion of these vocalisation types was similar to the ones obtained manually. PAMGuard also detected a high proportion of type 2 vocalisations. This is a positive first result in the automatic detection of grey seal underwater vocalisations, since vocalisation types 5 and 8 were the second and third most commonly recorded during this study. Although PAMGuard did not perform as well when detecting type 1 vocalisations, which were the most commonly recorded in this study, further research could focus on developing parameters within the whistle and moan detector to be more accurate on this specific vocalisation and improve its detection.

Moving forward, although these techniques need to be refined to automate the process of detection of grey seal vocalisations, we consider that the application of PAMGuard for grey seal acoustic monitoring together with a subsequent manual verification could provide information about the presence of this species. Therefore, acoustic monitoring of grey seal populations could start to be integrated in the already established monitoring programmes for cetaceans, helping to maximise existent large acoustic datasets within a conservation framework.

Ireland has committed to net-zero carbon emissions by 2050 (Climate Action and Low Carbon Development Bill, March 2021). In order to achieve the targets, renewable energy developments are planned along the Irish coast. Some examples include what will be Ireland’s largest offshore windfarm, Codling Wind Park, with a projected starting date of construction in 2024/2025. Looking at the west coast, in April 2021, the Electricity Supplied Board (ESB) announced that Moneypoint in County Clare was to become a major base for renewable energy and that a partnership with Equinor (formerly Statoil) will see a 1.4 GW offshore wind farm off Counties Clare and Kerry by 2030 with the capacity to power 1.5 million households. Therefore, acoustic data collected for these areas of all marine mammal species will be essential for planning decisions, and the assessment of archived acoustic datasets for grey seal presence will now be possible.

The outputs of this study will help to lay the foundations for acoustic monitoring of grey seals in Irish waters, and its application would bring additional advantages, providing information about the occurrence, distribution, and habitat use of this species, which could contribute, by providing policy advice, to the designation of Marine Protected Areas. Overall, the outputs of this study will help meet requirements under European legislation regarding the monitoring, management, and conservation of this protected species.