Pilot whale stranding: acoustic analysis report

4. Methods

4.1 Data collection

MD-SEDD has been operating an array of passive acoustic monitoring (PAM) recorders across Scottish shelf waters since 2022 as part of the SPAN project^[1]. At each site within SPAN, PAM moorings comprising an acoustic release (VEMCO AR), a broadband sound recorder (Sylence-LP; RTSYS), and a cetacean click detector (F-POD; Chelonia Ltd.) have been deployed to collect data for up to six months at a time (Figure 2). These moorings are generally retrieved and serviced two to three times per year with the intention of collecting year round data at each site, however in practice this can be variable depending on battery life and memory card capacity, and potential losses of equipment.

Broadband sound recorders are programmed to record with a duty cycle of 10 minutes on and 20 minutes off to maximise battery life and memory capacity. They utilise a sample rate of either 64 or 128 kHz, facilitating the recording of sounds between 5 Hz and 32 kHz, or 5 Hz and 64 kHz, respectively (the highest frequency at which sounds can be recorded is half of the sample rate). The .wav files that are recorded contain all sounds within those frequencies, including those produced by animals, such as cetacean vocalisations, as well as human activity, such as shipping and construction. The detection range of these devices for different sound sources will vary depending on the local conditions, the sound pressure level at the sound source, and the frequency of the sound source (broadly speaking, lower frequency sounds can be detected at greater ranges than higher frequency ones).

The automated cetacean echolocation click detector, F-PODs, continuously log the occurrence of cetacean echolocation clicks. These devices are typically used for harbour porpoise (Phocoena phocoena) monitoring but are also capable of detecting sounds produced by other echolocating species, although it is difficult to identify these to species level. F-PODs have a maximum detection range of 400 m for harbour porpoises and 1 km for delphinids (Chelonia Ltd.). They do not record sounds but note each time they detect a sound that could be an echolocation click. The devices also log detections of some echosounders and sonar.

4.2 Analyses

Within the SPAN dataset, acoustic data collected at the Tolsta site were the closest to Traigh Mhòr, the location of the mass stranding event (Figure 1). PAM devices were deployed at this site in March 2023, and were serviced and redeployed as part of routine maintenance, in the days prior to the mass stranding. Following reports of the mass stranding, Marine Directorate vessels were tasked with recovering the Tolsta PAM mooring again and redeploying another PAM mooring in its place. This recovery took place on 21st July 2023 (Table 1). This means that data from two deployments cover the time period of relevance to the stranding. The first deployment covered the period from 5th March to 14th July 2023, and the second covered the period from 14th July to 21st July 2023. The moorings were deployed in a water depth of ~100 metres (chart datum) and the recorders were suspended on buoyed ropes at ~5 metres above the seabed.

For this report, acoustic data captured at Tolsta between 10^th and 16^th July were selected for analysis, to cover the day of, and days preceding the stranding, on 16^th July (Table 1.). The two broadband sound recorders used during these deployments were programmed to capture data using different sampling rates, due to the size of memory cards available for use at the time of the deployments. For the deployment up to 14^th July (767-Tolsta), the maximum recorded frequency was 32 kHz (sampling frequency of 64kHz), while for the deployment from 14^th July to 21^st July (868-Tolsta), the maximum frequency recorded was 64 kHz (sampling frequency of 128kHz). Unfortunately, the F-POD on the first mooring (767-Tolsta) had stopped recording eight days prior to the recovery on 14^th July due to battery depletion, so within the analysis time frame, F-POD data were only available from 14^th July onwards. For this report F-POD data between 14^th July to 16^th July (868-Tolsta) are presented.

Additionally, data from a broadband recorder and a click detector located at the Garenin site west of the Isle of Lewis were selected for the period covering 10^th to 16^th July 2023 to match the Tolsta site data analysed, through two separate deployments (deployment 747 and 864, see Table 1). The moorings were deployed in a water depth of ~80 metres (chart datum) and the recorders were suspended on buoyed ropes at ~5 metres from the seabed. PAM devices had been deployed at this site in January 2023 and were serviced and redeployed as part of routine maintenance on 13^th July 2023, prior to the mass stranding. For these deployments, the maximum frequency recorded was 64 kHz (sampling frequency of 128kHz). F-POD data from the first mooring (747-Garenin) stopped recording in April 2023 and had no data available for the period of interest. F-POD data from mooring 864-Garenin were analysed for the period 14^th to 16^th July 2023 to match Tolsta analysis.

Acoustic data were downloaded from the respective devices and quality checked.

For F-POD recordings this involved visual validation of click trains to examine any potential misclassification in detections. For the broadband recordings this involved checking that the length of the recorded files conformed to the duty cycle and sample rates specified. During this process, a continuous narrowband sound was detected on both broadband devices from Tolsta, as well as on one device from Garenin (864). On the broadband device deployed for 767-Tolsta, this sound had central frequencies of ~20 kHz (see Figure 3 for an example). The device used for deployment 868-Tolsta also had noise, at central frequencies of ~20 kHz, as well as harmonics at ~40 kHz and 60 kHz (see Figure 4 for an example). The device used for deployment at 864-Garenin had a similar strong continuous digital sound at frequencies of 24 kHz, 30 kHz, 47 kHz and 53kHz for the first 9.5s of every 10 minute recording, then changed to 20 kHz, 28kHz, 40 kHz and 49 kHz for the rest of all the recordings. The continuous nature of this sound, which occurs at a consistent amplitude for the duration of all captured recordings, suggests it was generated by the broadband devices themselves (i.e. self-noise), rather than originating from the environment. Additional tests were performed at the MD-SEDD Marine Laboratory in Aberdeen, where the broadband devices with suspected self-noise were set to record at the same settings used in the deployment, then left to record overnight in air. The resultant files were visually inspected, and the same pattern of noise was identified. This confirmed that the observed noise patterns present in the recordings while deployed at sea are not related to any noise source present in the ocean but are instead an artefact of the devices. As with the 864-Garenin device, for both broadband devices from Tolsta there appears to be a shift in frequency approximately 10 seconds from the beginning of each recording.

The broadband recorders were returned to the manufacturer to ascertain the source of this self-noise, which was identified as a component of the electronic board of each device which interfered with signal acquisition. This was not a sound that was produced or recorded by the device, but an artefact of the electronic processing. The frequency shift at the beginning of each recording was also queried with the manufacturer, concluding that the frequency after the shift, and therefore for the majority of the recording, is the correct frequency. Both broadband devices were also sent for calibration. Although the self-noise makes it difficult to estimate absolute underwater sound levels at the frequencies of the self-noise, it does not impact the ability to identify sounds produced by cetaceans or anthropogenic activities. It is also still possible to identify key features (e.g. frequency, call rate) of cetacean vocalisations to potentially identify the source of vocalisations to species level. Therefore, subsequent analyses focused on identifying the presence of cetaceans and anthropogenic sound within the dataset.

4.2.1 Broadband data analyses

Tolsta site

Analysis of the broadband data was performed manually since there are currently no readily available automated detectors for sounds specifically produced by pilot whales (Globicephala spp.), although classifiers could potentially be used to separate sounds produced by pilot whales and similar (e.g. killer whales) from smaller delphinid vocalisations. However, more generally, the aim of this analysis was to identify cetacean vocalisations and any potential anthropogenic impulsive sounds, not to classify cetacean vocalisations by species. Automated classifiers are currently only available for a few cetacean species, which do not cover the breadth of species that could be present around the deployed PAM devices analysed here. In particular, pilot whales have a highly complex and broad vocal repertoire of typical dolphin sounds (Vester et al. 2017) comprising of a variety of clicks and buzzes, broadband non-harmonic calls, and different types of whistles and pulsed calls. The vocal repertoire falls in different frequency ranges, from low-frequency clicks with main energy below 2 kHz to ultrasonic whistles above 60 kHz (Vester et al. 2017), with characteristics that overlap with other cetacean species vocalizations.

Analyses of the broadband acoustic data from the Tolsta site were carried out by MD-SEDD staff, using Raven Pro 1.6 software. Audio-visual inspection of each file was performed by computing spectrograms, which provide a visual representation of the frequency spectrum against time. Spectrograms were computed using a Hamming window with a window length of 1024 samples and 75% overlap. Spectrograms generated from each day (480 minutes over 48 .wav files) of data were displayed together using a window preset view of 180 seconds. This enabled scrolling through the long dataset by sliding a 180 second window, using an overlap to avoid missing any data. Cetacean vocalisations were identified by the presence of whistles, clicks or echolocation sequences. Whistles can be identified when inspecting spectrograms by their time-frequency contour, which can take different shapes including, but not limited to, upsweeps, downsweeps and inverted U-shapes. Echolocation clicks are short-duration, directional, broadband clicks typically used for foraging. These can also be visually identified in spectrograms (see Figure 6). All potential cetacean vocalisations or echolocation clicks were annotated by selecting the area of interest on a spectrogram (red bounding box in Figure 5) and saving the time of the sound of interest in a selection table. The same process was used to identify any potential anthropogenic impulsive sounds. When anthropogenic sounds were detected, the frequency, duration, and number of impulses were annotated. The absolute sound level of the sound of interest was not investigated due to the previously noted broadband self-noise recorder issues.

Garenin site

Analyses of the broadband acoustic data from the Garenin site were carried out by colleagues at SMRU, due to resourcing constraints within MD-SEDD. The analysis was undertaken using PAMGuard software (Gillespie & Macaulay, 2024). Initial inspection of the data identified cetacean clicks and whistles as well as other sounds of unknown source. In order to investigate the frequency characteristics and potential source of these sounds, a variety of sound processing models were applied at different sampling rates. First, the whole frequency range of the recordings (5 Hz to 64 kHz) was explored applying the following processing steps:

1) Spectrogram Display (SD64kHz) using an FFT length of 2048 bins, FFT hop size of 1024 (50% overlap) using a Hann window, resulting in a frequency resolution of 62.50 Hz and a time resolution of 16 ms.

2) Click Detector (CD64kHz) with an 8 dB threshold, a high-pass Butterworth digital filter with a 50 Hz cutoff frequency, and a trigger filter with a high-pass 50 Hz cutoff frequency.

3) Whistle and Moan Detector (WM64kHz) to search for frequency modulated signals with a minimum frequency of 100 Hz, a maximum frequency of 24 kHz, and a detection threshold of 8 dB.

4) Noise Band Monitor to inspect 1/3 octave band noise measurements between the 111 and 57 kHz frequency range, measuring sound energy every 10 seconds.

5) Long-Term Spectral Average (LTSA) averaging spectral energy every 1 second, using the FFT data.

Then, the frequency range of 5 Hz – 4 kHz was further explored given the identification of “knocking sounds” below 1 kHz (see results section below). A decimator model was employed to resample the audio to a lower sampling rate of 8 kHz, and acoustic processing models were applied to the new decimated audio data. The following data modules were configured:

1) Spectrogram Display (SD4kHz) using an FFT engine at the decimated data with an FFT length of 256 bins and an FFT hop size of 128 (50% overlap) using a Hanning window, resulting in a frequency resolution of 31.25 Hz and a time resolution of 31 ms.

2) Whistle and Moan detector (WM4kHz) with a minimum frequency of 30 Hz, a maximum frequency of 4 kHz, and a detection threshold of 8 dB.

3) Click Detector (CD4kHz) with a trigger threshold of 6 dB above the background noise level, incorporating a digital high-pass Butterworth filter with a 20 Hz cutoff frequency, and a trigger band-pass Butterworth filter with a frequency response of 40 Hz (high-pass) to 1 kHz (low-pass).

4) Click classifier using an energy detector comparing the energy within the test frequency band of 50 Hz to 500 Hz to a control frequency band of 600 Hz to 1000 Hz, applying a minimum energy difference of 2 dB between the test and control bands. The classifier was further configured to search for a peak frequency within the 40 to 400 Hz frequency range.

After manually examining the sound processing module results in PAMGuard Viewer, to explore all potential sound sources for both the full frequency range (5 Hz–64 kHz) and the decimated data (5 Hz–4 kHz), it was observed that no relevant signals were present at higher frequencies. Consequently, it was decided that further investigations on the detected <1 kHz knocking sounds would be based on the decimated data and their sound processing modules.

4.2.2 F-POD data analyses

Analyses of the F-POD data from the Tolsta and Garenin sites were carried out by MD-SEDD staff. F-POD data were downloaded and processed for detection of local presence of delphinid species and for sonar presence. The F-PODs deployed in January and March 2023 stopped recording prior to the 10^th July 2023 and therefore only data from the F-PODs deployed on 14^th July (868-Tolsta and 864-Garenin) could be used for data analysis (Table 1). F-POD data were processed using the manufacturer’s F-POD.exe software (V 1.1) and KERNO-F classifier (Chelonia Ltd., 2022). This classifier is an algorithm that systematically seeks click “trains” in the series of clicks logged in each minute, processes them according to several features and allocates them to “quality” classes of high, moderate or low which represent the confidence of the classification. Visual validation of the click trains around the time of interest was carried out to examine any potential misclassification in detections and to verify the classifications assigned.

The different classes are:

• Narrow Band High Frequency Clicks (NBHF; in range 110-155 kHz), which are associated with harbour porpoises.
• Other cetacean clicks (Other Cet; in range 30-160 kHz), which are associated with delphinid clicks, including but not limited to long-finned pilot whales.
• Sonar trains

We report on detections that have been classified as high or moderate quality by the manufacturer’s software. Low confidence detections have not been included due to uncertainty around their true source.

4.2.3 Other information sources

In addition to the acoustic data collected by MD-SEDD, this report also draws upon three other information sources.

Firstly, AIS (Automatic Identification System) and VMS (Vessel Monitoring System) data, which provide information on vessel movements, were investigated. Most large ships are required to carry AIS transmitters, and fishing vessels over 12 metres in length are required to carry VMS transmitters. Other vessels may carry AIS transmitters by choice. The Marine Directorate has access to these data, and as part of this project, these data were visually reviewed for the time periods where sounds of interest were identified in the acoustic data.

Secondly, information was requested on activities undertaken by the Ministry of Defence (MOD), within 100 nautical miles of the stranding, in the week preceding this.

Thirdly, Marine Directorate Licensing Operations Team (MD-LOT) which is responsible for marine licensing in Scottish waters, and Offshore Petroleum Regulator for Environment and Decommissioning (OPRED) which is responsible for oil and gas licensing were asked to provide information on licensed activities within the region in the week preceding the stranding.