Advancing glider-based acoustic measurements of underwater-radiated ship noise

Buoyancy-driven autonomous underwater gliders are a cost-effective means to measure data across near shore and open ocean regions (Testor , 2019; von Oppeln-Bronikowski , 2023), carrying a variety of sensing payloads over time spans approaching a year, depending on configuration (Schofield , 2007; Van Uffelen , 2023). With their quiet behaviour and ability to sample the ocean autonomously, underwater gliders have become an indispensable tool for oceanographic research and collecting in situ environmental data in passive acoustic monitoring (PAM) missions (Cauchy , 2023; Helal , 2024a). In the context of Atlantic Canada, in particular, this has led to a successful program of tracking and identification of North Atlantic Right whales (NARw) in their natural habitat without disturbing them (Gervaise , 2021; Klinck , 2012; Küsel , 2017).

Because of their typically low acoustic signature (Cauchy , 2023), gliders also show potential in applications related to quantifying the sources of noise such as ships (Haxel , 2019; Wang and Yuan, 2021). In 2024, a systematic review was published to present the promising ability to use gliders to measure and characterize the underwater radiated noise (URN) of vessels (Helal , 2024a). URN is a widely recognized parameter by ship classification societies and, therefore, an important metric in the science and engineering communities. Accurate glider-based assessment URN requires two main aspects: the knowledge of glider self-noise from the intermittent behaviour of the glider's internal mechanisms and sensors (Cauchy , 2023; Schofield , 2007) and the derivation of an ocean sound propagation loss (PL) model (Helal , 2024a). Recent studies indicate that employing a range-dependent water-column sound speed profile (SSP) for sound PL modeling is more advantageous than using a constant profile (Gervaise , 2021; McKenna , 2013; Vagle , 2021). The spatial mapping capabilities of gliders can be used to get a spatially varying field of SSP from glider salinity, temperature, and depth information to derive a range- and depth-dependent PL model (Jensen , 2011; MacGillivray , 2023). Range-dependent models can also take into account changes in the seabed bathymetry profiles, which might cause sound to travel differently than if the bathymetry was assumed to be flat, especially in shallow water areas (Luo , 2008; MacGillivray , 2023).

This study combined glider-measured ambient sound levels from an underwater buoyancy-driven glider and a glider-data-derived PL in deep (190+ m) and shallow (60+ m) water sites. We do this to (1) compare a glider, a mooring, and a surface drifting buoy PAM platform in characterizing vessel noise signature, and (2) assess the differences between developing PL models, including scenarios in which the SSP and bathymetry were held constant, either the SSP or bathymetry was range- and depth-dependent, and both variables were simultaneously range and depth dependent. Finally, (3) we use the glider platform to track and locate the sound source (vessel under test) that is directly relative to the glider. By evaluating the performance of gliders in characterizing and locating ship noise, researchers can determine their suitability for long-term monitoring of acoustic pollution from ships in different marine regions. Additionally, the data collected by gliders can help identify areas of high shipping activity and prioritize conservation efforts in those areas (Haxel , 2019; Kowarski , 2020; Matsumoto , 2011; Stinco , 2021).

The paper is organized as follows: Section II presents the background and methodology for conducting the experiments, whereas Sec. III presents and discusses the outcome of the results. Section IV discusses the glider's ability to detect the direction of arrival (DOA) of a marine sound source.

Two field experiments were conducted in Conception Bay, Newfoundland and Labrador, Canada. We used three separate platforms for URN measurements: a glider, a drifting buoy, and a bottom-mounted hydrophone. The first experiment in Holyrood Bay, located at the head of Conception Bay, was performed to quantify the acoustic self-noise signature of the glider. The site has very low shipping traffic, and the water depth was relatively shallow, ranging between 60 and 90 m. The second experiment was conducted in the middle of Conception Bay (see Fig. 1) close to the town of Brigus with deeper water depths of around 120–300 m. We refer to these sites as the shallow and deep water experiments in the rest of the text, noting that the reference to shallow and deep is relative to the selected sites.

The weather conditions on both trial days were calm (less than sea state 2; ANSI/ASA, 2009; ISO, 2016). Table I provides approximate oceanic and atmospheric conditions during the tests based on data provided by a Smart Atlantic buoy near the entrance to Holyrood Bay (Smart Atlantic, 2022).

The Slocum glider manufactured by Teledyne Webb Research is one of the most prolific buoyancy-driven underwater glider platforms in use by the ocean research community.¹ The glider science payload section was equipped with an RBR Legato Conductivity Temperature Depth (CTD) sensor² and the JASCO Ltd. manufactured Ocean Observer multi-channel PAM recording system³ with two High Tech, Inc. (HTI) wing-mounted hydrophones (High Tech, Inc., 2021). The tested glider featured a 350-m “high-displacement” (HD) oil pump (⁠ $450 cm3$⁠). All glider science payload sensor timestamps were synchronized to the glider science data-logging computer.

The glider was ballasted and prepared for deployment at the Memorial University glider facility in a large tank, following standard manufacturer recommendations. Baird (2007) provides a basic outline of the procedure. The drifting hydrophone array and bottom-mounted hydrophone were used as references to assess the glider's performance as a PAM system in comparison to these established fixed underwater noise measurement platforms. The bottom-mounted mooring was equipped with a single Ocean Sonics icListen RB9-HF hydrophone 45 m above the seafloor (Ocean Sonics, 2021a). The drifting buoy from Ocean Sonics Ltd. is equipped with an array of three icListen RB9-HF hydrophones set at fixed depths of 32, 63, and 94 m (Ocean Sonics, 2021b). The noise levels from each platform were evaluated and classified for each source individually.

The glider served as the primary PAM platform for the shallow water trial. This configuration, which did not include the drifting hydrophone array and bottom-mounted hydrophone that were employed in other trials, enabled the evaluation of the glider's acoustic performance without interference from other measurement systems. For the deep water trial, the hydrophone array and bottom-mounted hydrophone were deployed with the glider. We chartered a fiberglass vessel with a medium-speed diesel engine (see Fig. 2) as a moving shipping noise source. Table II summarizes the main characteristics of the vessel. The vessel was scheduled to navigate between two waypoints, back and forth, parallel to the three platforms' locations, as presented in Fig. 1. At engine speeds of 1500 and 2000 rpm, four runs of each speed were conducted, ensuring a constant vessel speed over the ground throughout each run (ISO, 2016). The vessel maintained a constant velocity throughout the testing, ensuring that equal time intervals translated to equal distances from the closest point of approach (CPA). This important consistency permitted precise measurement of the received noise level (RNL) over a specific time window, given by the vessel's speed and CPA, as discussed in (Helal , 2024b). The CPA is measured individually using an attached Global Positioning System (GPS) on the vessel and three platforms. The glider's velocity was nearly stationary compared to that of the vessel.

In both experiments, the glider executed a prescribed route between programmed waypoints and conducted several dives and climbs at regular intervals. The initial tests with the glider were performed in Holyrood Bay on July 29, 2022, near 47.4156 N and 53.1331 W. The glider completed six dive and climb cycle (one dive and climb cycle = 1 yo) profiles at this location at a water depth of 60–70 m, as shown in Fig. 3(a). In the subsequent deployment in Conception Bay on January 13, 2023 (47.5333 N and 53.1583 W), the glider completed four yo's diving to 190–200 m. The glider inflection dive depth changed with bathymetry (10 m above the seafloor) between yo's [Fig. 3(b)]. Using the total buoyancy drive available, Slocum gliders can typically achieve maximum forward speeds of 0.25 m/s or 22 km/day in still water. In the deep water site, ocean currents were negligible, and the glider achieved average forward speeds of 0.23 m/s. Using uncertainty estimates in Claus and Bachmayer (2015), we estimate that the glider position accuracy in the trials was precise to within 50 m as the average dive interval was less than 40 min and currents were less than 5 cm/s. In shallow water, the glider stayed on the surface for a long period of time after each dive cycle. Thus, Fig. 1 shows a drift before each dive. Although this drift could be a concern in some scenarios, it did not compromise our study. We focused on evaluating the glider's self-noise on a dive-by-dive basis.

The glider's 350-m-rated HD pump has a reduced oil column pumping time at depth compared to the more common 1000-m versions, reducing total time in the glider segment where pump noise contaminates the recordings. The shorter buoyancy pump operations result from physical design differences and variations in motor calibration constants, increasing oil flow speeds and thereby reducing power usage. As a result of faster oil flow speeds, the sound levels of the tested 350-m pump were somewhat higher during buoyancy pump operation than a 1000-m pump. Slocum gliders typically only activate the oil pump during inflection, going from a dive to a climb. The glider pump was, therefore, running only 10–30 s for each inflection interval, whereby the glider was transitioning from a dive to a climb. During surface inflections, the glider's internal vacuum draws back the oil volume without activating the pump motor, making this interval quieter.

The glider dive and climb angles were set to 26 deg for the shallow water trial and 20 deg for the trials in deep water. In both missions, the glider dynamically adjusted the battery pitch pack position to maintain even climb and dive angles, as well as the automatic buoyancy pump trim (autoballast) to maintain even vertical speeds of around 10–12 cm/s. The PORT and STBD (starboard) wing-mounted hydrophones were configured to record acoustic signals at a frequency of 32 kHz using 24-bit quantization, providing a dynamic range of 147 dB between the overload signal and quantization noise. The hydrophones on the drifting and moored platforms were set to sample at the same rate. The start-up time of the Ocean Observer on the glider is around 60–120 s, resulting in the glider recording being restricted to 10 m below the surface on the first glider dive segment only but continuously thereafter. In shallow water, the glider surfaced after every single dive and climb cycle, whereas in the Conception Bay deep site tests, the glider performed several dives and climbs between surfacings, as in the second and third yo's presented in Fig. 3(b). Due to the currents in Holyrood Bay, the glider was pushed backward between subsequent surfacings, leading to an adjacent set of glider profiles oriented alongside the central axis of the bay.

Spectral analysis can be used to isolate and determine the relative sound amplitudes of sound sources in acoustic recordings (Cauchy , 2023; Jiang , 2019). By looking at a spectrogram (Fig. 3) of the hydrophones, notes from the field tests and glider flight data, we are able to identify the time window of existing noise sources, e.g., vessel noise or biological sounds. In addition, through careful comparison of the glider flight record with the acoustic files' timestamp, we are able to identify the noise levels and timing of individual glider components for further analysis. The recorded windows with low noise levels were beneficial for assessing background noise during the experiments. The red boxes on the spectrogram (Fig. 3) show the data logging was paused either when the glider approached the surface or reached the surface. In Fig. 3, orange dots illustrate the noise levels produced by the glider's buoyancy engine, battery pack, and rudder (blue squares). Also, marine mammals (dolphins and minke whales) were sighted and identified during the deep experiment (highlighted with green boxes).

The calculation of the URN uncertainty measured by the three platforms is conducted as a parameter to show the glider's performance as a PAM observer compared to the conventional stationary platforms. The study focused on the main sources of error that significantly impact the assessment of vessel noise signatures. They are listed in Table III. Each entry represents a source of measurement uncertainty, whether random or systematic. An overall standard error was calculated by combining each error for each platform individually (ISO/IEC, 2008; Taylor and Kuyatt, 1994).

In accordance with AQUO Consortium (2012) and Keizer (2022), we presented the uncertainty pertaining to the levels of vessel noise signatures measured by the glider relative to the two stationary PAM platforms. The sources of uncertainties for each platform were identified and assessed. This process included computing the standard uncertainty of the mean of the vessel's URN measurements for each platform. Measurement uncertainty refers to the built-in doubt surrounding measurement errors, which are unknown and should be minimized to increase the measurement quality. These uncertainties are converted to statistical quantities, termed the standard error. Evaluating standard errors in URN measurements allows for evaluating the effectiveness of different platforms in assessing the vessel noise signature across a wide frequency range.

The vessel was positioned about 2 km away from the PAM systems, and its engine was turned off. Our approach consisted of assessing the BNL from three observers simultaneously at the beginning of the trial to obtain three measurements. Afterward, the sound received by a particular platform is adjusted using the BNL measured by the same platform. For the background noise assessment, we used glider records that were not contaminated by vessel noise or glider pump motors and actuators. As a result, we selected the last climb in the Holyrood Bay trial and the first dive in the Conception Bay trial. The glider's sound levels were averaged throughout the water column, starting with the initial dive phase (after the oil pump turned off) and ending before the reactivation of the oil pump for ascent.

Several methods of self-noise assessment exist. Tesei (2019) measured the self-noise of the Slocum glider in a water tank using one hydrophone at the noise. The water tank might overestimate the self-noise levels caused by the wall reflections (Tesei , 2019). Cauchy (2023), Haxel (2019), and Jiang (2019) also used mounted hydrophones attached to the glider at the front and/or tail in the open-field ocean. The mission flight data files were used to identify dive phases, transitions, and time spent in each phase. Time-frequency analysis is used to characterize different types of self-noise, specifically ambient noise, rudder noise, and pump motor noise. These characteristics are measured during an ordinary dive (Cauchy , 2023; Haxel , 2019; Jiang , 2019). We followed the same procedures to determine the self-noise of the Slocum glider.

The glider records continuously when each internal component, such as buoyancy pumps, rudders, and the pitch motor, is active. By superimposing the flight data with the corresponding acoustic files, we can derive time-series signal windows that effectively represent the occurrence of each noise source. Using the same time windows, the acoustic file segments were analyzed to estimate the acoustic signature for every state of the different motors in the glider. This was performed for the PORT and STBD hydrophones to account for the directional characteristics of the self-noise. A comparative analysis, in narrowband and one-third octave bands, was conducted between the resultant noise levels and ambient noise of the ocean spectral analysis techniques to identify the characteristic frequencies and intensity of noise produced by each component.

The ocean was subject to contamination from diverse sources of acoustic interference in the Holyrood site. Thus, the sixth dive was selected for the purpose of analyzing the acoustic signature of different motors. In Fig. 3(b), the first dive in the deep water trial had less background noise than the other dives; it was preferred for the self-noise signature investigation. In addition, the self-noise analysis was focused on the period when the glider reached its maximum depth and began ascending. Audio data when the glider surfaced were excluded, as discussed in Sec. II C.

The empirical PL values, which describe how the sound intensity attenuates due to traveling through the water, are not available in this study to precisely characterize the vessel noise signature. Therefore, the paper also aims to show the help of using range-dependent (R/D) oceanography data (temperature, salinity, and density) of the ocean in characterizing the PL of sound compared to use of range-independent (R/I) parameters. How can the changes in SSP get different PL values, particularly in shallow water? It is well-known that there is no standardized method of determining the PL that accounts for the environment uncertainty. Because of that, we used several existing models from previous studies for our comparison.

PL is commonly estimated using simple spreading laws [ $PL=N log10(R)$]. R represents the distance from the noise source in meters while N is a scaling factor. N ranges from 10 to 20, depending on the source and receiver distance. This approach is limited in constructing accurate predictions in complex environments. It is most effective in scenarios in which the environmental properties are range independent. Spreading law models may result in significant errors when applied to complex coastal and inland water environments (Jensen , 2011). Therefore, the study focused on frequency-dependent PL models, which can include marine environmental parameters (Chion , 2019; MacGillivray , 2019; MacGillivray , 2023).

PL numerical models, which aim to represent the complex ocean environment accurately, predominantly depend on parameters such as the average sound speed in the ocean, depth of the noise source, and characteristics of sediment absorption on the seafloor (Ballard and Lee, 2017; Bureau Veritas, 2018; Lloyd's Register, 2018; Yang , 2018). However, the accuracy of PL models can be enhanced by integrating range- and depth-dependent SSPs of a volume of the ocean. In addition, the seafloor topology may also impact the estimated values by a numerical PL model. Various numerical methods, such as wavenumber integration (WNI) and the parabolic equation (PE) approach, employ SSP to effectively simulate sound propagation in the ocean (Jensen , 2011; MacGillivray , 2023).

As discussed in Helal (2024b), the WNI techniques are commonly used for simulating sound propagation in stratified media and have shown effectiveness in modeling sound wave propagation in shallow waters. This method allows for accurate modeling of intricate environments, capturing the interactions between sound waves and the layered ocean structure at frequencies below 4 kHz. The study (MacGillivray , 2023) highlighted inaccuracies of the WNI method in estimating source levels below 1 kHz at various CPAs. On the other hand, the PE method models ocean sound propagation below 1 kHz by approximating the Helmholtz equation, effectively addressing range-dependent issues. This method effectively simulates low-frequency sound waves, considering bending phenomena caused by the range and depth dependency of the SSP while disregarding backscattered sound as insignificant according to standard ocean-acoustic models. Its accuracy decreases at higher frequencies due to assumptions that may not hold, making ray tracing more appropriate for shorter wavelengths (Collins, 1993; Helal , 2024b).

The AcTup toolbox (Duncan and Maggi, 2005) was used to implement PE and WNI models for frequency ranges of 10–160 Hz and 200 Hz–4 kHz, respectively, as discussed in Helal (2024b). The numerical solution had a boundary condition starting at one wavelength to avoid a singularity at the source location. The depth resolution, a function of the frequency, is equal to 0.25 × wavelength. The horizontal range resolution was 1 m step. The vessel draft (source depth), SSP, seafloor bathymetry profile, and acoustic characteristics of the seafloor sediment were used as inputs to the models. The source depth and seabed acoustical characteristics remain the same, whereas the SSP and seabed bathymetry profile change by altering the location where the PL was calculated (R/D or R-B/D model).

Sound waves in shallow water environments often interact with the seafloor because of the limited depth (MacGillivray , 2023). This interaction results in reflection, refraction, and absorption of the sound waves that are impacted by the composition of the sediment. For example, sandy sediments generally have higher sound speeds than muddy sediments. The speed of sound in sediments varies with frequency, impacting the strength of reflected waves. Furthermore, sediment density is an important factor in determining acoustic impedance and attenuation coefficients (Dall'Osto and Tang, 2022). The limitation in the PL calculation was the ability to determine the sediment layer thickness or how many layers were under the seafloor. We treated the seafloor as a half space, which is characterized by constant acoustic properties and might have a negative impact on lower frequencies below 20 Hz in 60 m water depth. In deep water, the interaction of the seabed is less important; however, the sediment acoustic characteristics are used to retain the SSP as the only changing parameter. As stated in the Introduction, the SSP presented in this study is a range-dependent profile (from the glider) and a range-independent profile. This method helped us understand how the glider contributes to the performance of the PL models.

Given the challenges in accurately estimating seabed acoustic properties at our site from the acoustic, SSP, and bathymetry, we relied on verified data from the scientific community to ensure robust results in our numerical PL models. Table IV displays the acoustical properties of the sediment found on the seabed, containing a central layer of mud and a dispersion of sand and gravel (Ballard and Lee, 2017; Dall'Osto and Tang, 2022; Shaw and Potter, 2015).

Alternatively, previous research has presented simple, explicit formulas applicable for determining the PL in shallow and deep waters. These methods do not require extensive numerical computation. Among these techniques are the ISO 17208-2 standard (ISO, 2019), the Meyer and Audoly (M-A) method (Meyer and Audoly, 2020), the Seabed Critical Angle (SCA) method (MacGillivray , 2023), and the ECHO Certification Alignment (ECA) method (Ainslie , 2022), which are simplified equations that can estimate the PL better than a simple spreading law. The methods of ECA and ISO 17208-2 primarily focus on surface reflection, but the SCA and M-A approaches take into consideration surface and seafloor reflections. They are sensitive to the accuracy of seabed properties. These models require less computation effort compared to the numerical PL models.

The glider covered distances of 550 m in the shallow water site and 1500 m in the deep water site. Using an onboard CTD sensor, the glider collected 1.0 Hz sampled conductivity, temperature, and pressure data from which we derived in situ underwater SSP (McDougall and Barker, 2011). Figure 4 illustrates the SSP for each of the glider dives and the average value.

In total, six PL models were created for the two sites, using the ISO, PE, WNI, ECA, SCA, and M-A methods for the shallow and deep water trials. A comprehensive comparison and analysis of the discrepancies between the models' outcomes can be accessed in supplementary material Figs. S.3 and S.4. An envelope between the PL model was implemented to show the frequency bands sensitive to the model type. The models addressed 100 and 500 m ranges, including multiple receiver points at depths of 20, 30, 40, 50, 60, and 70 m for the shallow water site, and depths of 30, 60, 90, 120, and 150 m for the deep water site. The sound source was consistently positioned at a depth of 2 m and applied to all of the models. This approach allows for a detailed study of PL variations in different water depths with dramatic SSP changes, which is important for precise acoustic modeling and analysis.

Applying the time-frequency analysis to the glider acoustic data revealed that the brushless direct current (DC) motor of the oil pump is the dominant source of self-noise for the glider, where its levels are notably higher than those of the rudder and background noise across the measured frequency spectrum. In the shallow water trial, it emitted a tonal peak at 16 Hz (960 rpm) along with harmonics up to 1 kHz [Figs. 5(a) and 5(b)]. The pitch motor noise was included in the oil pump analysis as both occurred at the same time; see supplementary material Fig. S.2. The comparison between the PORT side [Fig. 5(a)] and STBD side [Fig. 5(b)] indicates a slightly higher intensity of noise on the STBD, especially in the low-frequency range, hinting at possible asymmetries in glider noise emissions between the two sides. The STBD hydrophone recorded a peak noise level of 135 dB at 48 Hz while the PORT hydrophone recorded 125 dB. According to Table V, there was a difference in broadband noise levels of about 9 dB, with the STBD side being noisier. The Slocum glider's 350-m HD pump pumps oil faster than its 1000-m HD pump. This may cause differences in received PORT and STBD sound levels when the glider pump is activated. There is no off-centered moving mechanical component that is responsible for this effect. Because this is only a feature while the pump is on, it does not affect the rest of our analysis and only changes the uncertainty in self-noise when comparing the wing-mounted hydrophones.

In contrast, the rudder noise is evenly distributed across frequencies with moderate peaks and consistently goes above the BNL, which remains steady across frequencies. The rudder's noise added 6–13 dB over the ambient noise across a wide frequency range from 25 Hz to 10 kHz. This is consistent with the findings of Jiang et al. (2019). The noise levels recorded by both hydrophones were found to be comparable during the rudder motion as the difference is less than 1 dB, as shown in Table V.

During the deep water trial, BNLs were 8 dB higher compared to shallow water (see Table V). The oil pump noise profile exhibited decreased sound levels compared to the shallow water trial. The oil pump's noise signature was characterized by a tonal peak at 14 Hz (840 rpm) with a maximum level of approximately 127 dB at 44 Hz [Figs. 5(c) and 5(d)]. Thus, STBD hydrophone noise levels exceeded the port side, and the self-noise is characterized as directional noise.

The rudder noise was observed several times during the glider's descent and ascent phases; see the flight computer data in Fig. S.2. The rudder produced more frequent movement, resulting in a wideband noise generating a broad-spectrum signal with amplified harmonic energy between 100 Hz and 10 kHz, which is consistent with the findings of Liu (2018) and Haxel (2019). Its noise remains recognizable from the background noise with fewer peaks and a smoother distribution across the frequency spectrum. The STBD side [Fig. 5(d)] has slightly higher noise output compared to the port side [Fig. 5(c)], which could indicate consistent side-dependent self-noise characteristics.

We found that the oil pump ran at a higher speed in the shallow water experiment (960 rpm) compared to that (840 rpm) in the deep water test, leading to an increased level of self-noise, as shown in Table V. This observation suggests that the glider's buoyancy pump system operates differently at different water depths. Furthermore, the rudder displayed more significant variability in its movements in deep water compared to the consistent and predictable patterns observed in shallow water; see Fig. S.1. These inconsistencies in deep water led to shorter, less detectable rudder noise durations compared to the longer, more apparent durations observed in shallow water trials. This suggests that the glider's self-noise emissions were affected by the ocean environment in which the rudder tried to keep the glider heading on track.

The BNLs measured by the three platforms are presented in Fig. 6. The RNL of the vessel showed disagreement between the platforms resulting from the variation in the CPA between the vessel and each observer in each trial. The BNL was below the sound received by the hydrophones, making the measurement reliable and comparable.

The calculation of URN was facilitated by applying the geometric spreading law, which is primarily dependent on the spatial separation between the vessel and hydrophone array, thus, simplifying the acoustic propagation model (ANSI/ASA, 2009; ISO, 2016). Maintaining a consistent distance between the vessel's path and all platforms was challenging because of operational constraints. Consequently, for every measurement taken on each platform, a geometrical spreading loss correction was applied, referencing the actual latitudes and longitudes recorded from the vessel and respective platform (ISO, 2016; Jiang , 2020).

We found that the vessel URN measured by all three platforms was generally similar across wide bands but with some variations at specific frequencies in the 1500 rpm range trial. Figure 7(a) shows that at the engine speed of 2000 rpm, the URN levels from all platforms displayed prominent peaks in the lower frequency range, followed by a gradual decrease as the frequency increased. In contrast, Fig. 7(b) illustrates the 1500 rpm engine speed trial, which exhibits a less noticeable peak in the low-frequency range and gradual decrease in the noise levels. The peaks below 100 Hz represent the fundamental frequencies of the propeller and engine of the vessel.

Comparing the two experiments, the total vessel noise signature across most frequencies from all observer platforms was higher in the 2000 rpm vessel engine speed trial than in the 1500 rpm trial. Furthermore, the 2000 rpm trial typically has a more significant standard error. The URN changed from 157 to 165 dB as an overall noise by increasing the engine speed. The noise results were aligned with the findings of small vessels studied in Wladichuk (2019). The buoy and glider measured similar URN levels across the frequency range except at 2 kHz, where the noise levels tended to lower during the 1500 rpm experiment. The moored hydrophone was almost located in the deep sound channel. The deep sound channel acts as a waveguide, which means that the moored hydrophone recordings include sound waves from far away (Gassmann , 2017). Therefore, the URN levels were the highest over the glider and drifting buoy.

The glider exhibited comparable or lower URN measurement uncertainty to the other recording platforms, reaching ± 2 dB at frequencies between 250 and 630 Hz, with the exceptions of peaks at 100 and 125 Hz (due to the rudder noise), as shown in Fig. 7(a). The hydrodynamic noise during the glider's flight may also introduce some uncertainties, but it was not distinguishable in the trials. At engine speeds of 1500 rpm, the glider was 18 m below the sea surface and within the surface mixed layer. A notable URN discrepancy was observed, particularly at frequencies above 2 kHz compared to the other two platforms [Fig. 7(b)]. At that depth, the propagation of the high frequencies near the surface attenuated significantly, as occurred between 2 and 5 kHz as a result of the positive gradient of the SSP (Garrett, 2020). Thus, above 2 kHz, the recorded sound integrity of the glider warrants further investigation. The drifting buoy showed promising uncertainty values that match ISO 17208-1 (ISO, 2016). The results emphasize the glider's potential as a dependable instrument for reliable URN measurements from vessels.

It will be important to further assess and validate these results against longer deployments in different ocean conditions. The absence of long cables and floats compared to the mooring and drifting buoy contributed to the reduced sound measurement uncertainties of the glider. In contrast, previous research showed that bottom-moored systems generally have lower levels of uncertainty. We acknowledge that some circumstances may have contributed to the glider's impressive performance: (1) depth variability, where the glider's capacity to change depth enabled sampling of various places within the water column, potentially mitigating noise fluctuations that depend on depth; (2) location of the moored system, in which the hydrophone is anchored in a single location near the deep sound channel, recorded distant sources of noise during trials with a particular impact on low frequencies; and (3) cable-induced noise, where the vibrations and chafing caused by the cables created low-frequency self-noise that was unique to our particular configuration. Although cable-induced noise is not present in all moored systems, it helped the glider perform with less uncertainty across most frequency bands than the single-moored hydrophone platform.

Using the methodology described in Sec. II G, we applied six PL models to determine PL values at 100 and 500 m from the sound source in shallow and deep water environments. The Holyrood Bay trial in shallow water showed that the simplified models ECA, SCA, and M-A produced similar estimations of PL at a distance of 100 m from the sound source. However, as we went deeper, the differences between these models became increasingly apparent. The PE model showed a significant deviation from the others below 200 Hz, leading to higher PL values, as shown in Fig. 8. When considering the seafloor bathymetry profile (R-B/D), where the glider passed over during the missions into the PE model, PL values increased, especially in frequencies below 100 Hz. The centered frequency of 250 Hz was where all of the models almost matched; as shown, the gray envelope thickness decreased until the water depth was half. This highlights the important role of bathymetry in shallow water situations. When the sound wavelength is much larger than the ocean depth, the ocean acts as a high-pass filter for sound waves.

Increasing the horizontal distance between the measurement point and sound source to 500 m revealed expected behaviour within the simplified models due to more interactions between the sound and environment boundaries. The width of the envelope (gray shading in Fig. 8) increased, which demonstrates an unpromising representation of the PL using the simplified models. Figures 8(c) and 8(d) show the convergence of the ISO 17208-2 and ECA models with the PE model, regardless of flat (R/D) or actual bathymetry (R-B/D), particularly below 100 Hz. However, the models differed significantly as water depth increased for frequencies above 250 Hz. Supplementary material Figs. S.3 and S.4 visually show the inconsistencies or absolute errors between ISO standard 17208-2 and the other models. The significant deviation of the numerical PE model at frequencies below 250 Hz, compared to other models, is noticeable. The error surged up by going near the seafloor. The other models had a good agreement at low frequencies with ISO standard 17208-2, which means those models had difficulty predicting PL values in this frequency range, which may lead to the same result of underestimating the contribution of vessel noise signature at low frequencies. Overall, discrepancies between PE (R/D) and R/I, assuming flat bottom, and PE (R-B/D), actual bathymetry, in Fig. 8 suggested that the bathymetry is critical for predicting PL at the shallow water site, which implies the importance of the seabed geoacoustic properties to the PL prediction.

In deep water trial, we observed a significant deviation with the M-A model, which was above 200 Hz. Therefore, we chose not to include the M-A model in our envelope (see previous definition) to represent the critical frequency bands while estimating the PL values. The PE model showed a similar pattern, particularly in representing high PL values between 30 and 150 Hz. Actual bathymetry (R-B/D) data had a more minor effect on the PE model than the flat assumption (R/D), as shown in Fig. 9. The gray envelope narrowed near the water's surface at 250 Hz and shifted toward 200 Hz closer to the seafloor. In supplementary material Fig. S.4, from rows 4 to 8, all simplified models, including the WNI model, showed significant alignment below 500 Hz at the 100 m range.

At 500 m from the sound source, simplified models showed more noticeable differences in estimating the PL values near the water's surface (30 m) and all agreed by going deeper (120 m), as shown in Figs. 9(c) and 9(d). However, the PE model consistently agreed with the simplified models in estimating the PL values throughout the water column except between 30 and 100 Hz, where high PL values were observed. The WNI model demonstrates strong agreement with simpler models at frequencies above 250 Hz in shallow and deep water for short ranges. Figure 1 highlights the notable difference between the PE model and ISO 17208-2 in estimating PL values at low frequencies. The gray envelope in Fig. 9 showed that the simplified equations (e.g., ISO 17208-2) accurately estimated the PL values in deep water as it had small thicknesses across most frequency bands. However, the envelope thickness near the surface suggests that predicting the PL values near the water surface using simplified equations is complex. These findings provide valuable insights into the complexities of underwater sound propagation by highlighting the intricate variances in model behaviour and their sensitivity to water depth.

These findings reveal the complex relationship between model choice, SSP, and bathymetry in shaping our understanding of underwater sound propagation. The study found that simplified models, such as ECA, SCA, and M-A, show close agreement in predicting PL values at a 100-m range in shallow water. However, as the depth increases, the deviations between the models become more noticeable. The inclusion of bathymetry data in the PE model results in higher PL values below 200 Hz, highlighting the importance of bathymetry in the filtration of sound waves by shallow waters. At a distance of 500 m, the accuracy of the simplified models is compromised across a wide frequency range in deeper waters. The variability observed in the gray envelope suggests that using simplified models for estimating PL may be challenging because of underlying environmental complexities. Therefore, the numerical PE model with the range-dependent SSP and bathymetry profile is crucial to determining PL values. Accordingly, the glider added more understanding about how a source of noise contributes to the ambient noise of the ocean by accurately estimating the sound PL value using the range-dependent SSP. In deep water, the apparent agreement among the vessel's URN obtained from the three platforms depicted in Fig. 7 indicates that the geometrical spreading model law is acceptable for deep water conditions because of minimal seabed interaction.

The monopole source level (MSL) represents the vessel in 1500 and 2000 rpm trials, and the URN is presented in Fig. 10. The MSL below 100 Hz, which was picked up by all three platforms, went up after the environmentally driven PL correction that was found using the PE (R-B/D) model at each CPA point was applied. Ideally, this correction would result in the same MSL being recorded across the different recording platforms. This close alignment suggests that the model accurately compensates for environmental factors, confirming the PL model's reliability and overall good performance.

One key motivation for PAM-type glider missions is not only to measure sound amplitudes or identify the source but also to localize them (Jiang , 2019; Kowarski , 2020; Stinco , 2021; Wang and Yuan, 2021). In general, information on the time delay between at least three hydrophones is required to identify an omnidirectional sound source. Processing these time delays can be used to calculate the relative angle of the sound source to the receiver (hydrophones), which is referred to in the literature as DOA (Li , 2019; Tokgöz , 2020; Zhang and Rao, 2009). For the case of only two hydrophones, as is on our glider, the time delay information produces a mirror result allowing two possible sets of angles, also known as the forward-backward limitation on a coordinate reference frame (Li , 2019; Reeder , 2004). There has been some work using sound data to compute DOA from a set of only two hydrophones to a moving sound target (Li , 2019). Here, we attempt a different approach, using the directional sensitivity of the glider's hydrophones and showing accurate estimation of DOA to a moving sound source (vessel) relative to the glider in two cases: (i) vessel and glider have the same heading and (ii) glider and vessel have opposite heading. This approach is simple and can be used on platforms where both hydrophones are pointed in the same direction. We also discuss the limitations of this approach.

A 360-deg view around the glider was considered for determining the vessel's position relative to the glider. The sign of the time delay was crucial in determining the vessel's position within a semicircle. Our analysis showed a negative sign, indicating that the vessel was located on the port side of the glider in both trials, and this agrees with the actual recorded geographic data during the sea trials. After identifying the semicircle, the direction angle was reduced from 360 deg to a range of 180 deg. An essential question at this point pertained to the vessel's movement, precisely, whether it was aligned with or opposed to the glider's heading direction. The received SPL was used to address the question. Consequently, the received sound level became lower when the vessel was positioned behind the glider but increased once the vessel had passed the glider. This comparison is feasible when the selected points are equidistant from the CPA. The glider's pitch angle had a minimal variation during diving and climbing, less than $±1.2°$⁠, which had negligible impact on the sound arrival to the hydrophones.

Therefore, six points were selected prior to and following the CPA between the vessel and glider. The points displayed in Fig. 12 were approximately equal distances from the CPA (point “0”). Received sound levels were measured at all 12 vessel's locations. The vessel's direction was determined by comparing the sound levels received before and after the vessel passed the glider at corresponding points. An increase in the average mean of received sound after the vessel passes the glider suggests that the vessel is likely following a similar heading angle as the glider. In contrast, a decrease in the average mean after passing the glider indicates that the vessel is moving in the opposite direction of the glider. This allowed for precise identification of the vessel's movement in relation to the glider's trajectory.

The glider's wing-mounted HTI hydrophones have consistent sensitivity from 50 to 10 000 Hz, with a minimal uncertainty of ± 0.1 dB. Figures 12(b) and 12(d) display the mean and standard deviation values, respectively, for the received sound levels measured at six equidistant points before and after the vessel passage of the glider. We use standard deviation to show uncertainty in DOA as it is a simple metric of variance in sound levels received during the vessel's approach and departure from the glider. This variation is caused by the changing distance between the sound source and receiver as the sound transmission is significantly affected by this distance.

The DOA values were corrected after determining the vessel's relative movement direction to the glider. Figure 13(a) represents a trial where the glider and vessel (9 kn) had the same heading, and the estimated DOA obviously pointed toward the direction of the vessel at each point of motion. Figure 13(b) shows the opposite heading between the glider and vessel. The vessel was sailing at a lower speed of 6 kn. Thus, the glider with two hydrophones successfully estimated the azimuth angle of any sound source emitting a frequency above 800 Hz. However, the system is not sufficient in getting the elevation angle because of one-dimensional spatial information using only two hydrophones. Additional experiments are necessary to improve our understanding of the glider's ability to detect DOA under different operating conditions and in lower SNR scenarios, such as sea trials during sea state 3 or more. These experiments will provide insight into the glider's reliability in practical scenarios. It is important to acknowledge that the glider's ability to determine the distance of the source was limited by the number of hydrophones used. A third hydrophone would need to be attached to estimate the distance from the glider to the source. This would also further improve and verify the method employed to estimate DOA. In future trials with this glider, we intend to equip the glider with an additional hydrophone to evaluate this method further.

We evaluated the ability of a buoyancy-driven underwater Slocum G3 glider to conduct passive acoustic measurements of URN in two experiments in shallow and deep water sites. The glider's ability to fly quietly is notably beneficial for applications requiring a high SNR such as anthropogenic noise monitoring and directional sound source recognition. The objective was to evaluate the performance of a glider compared to a fixed-moored hydrophone and a drifting buoy with an array of three hydrophones in quantifying the URN of a targeted vessel. This study used hydrographic data from the glider's CTD sensor to enhance the modeling of underwater sound propagation in shallow and deep water. In addition, we use the glider recordings during different phases of the flight to assess the acoustic self-noise signature and quantify its capabilities in PAM applications.

The study found that the oil pump was the main cause of glider self-noise. The noise levels were significantly higher than those of the rudder and background noise in the shallow water, and there was a distinct tonal peak at 16 Hz. In deep water, the oil pump's noise decreased at 14 Hz while the rudder noise became broader and exhibited more harmonic energy between 100 Hz and 10 kHz. These findings indicate that the glider's noise characteristics change with depth and water depth affects the glider's self-noise profiles.

The glider hydrophone SPL recordings aligned closely with the drifting hydrophone arrays and were better than the moored deep hydrophone in monitoring the URN of an under way vessel. The glider was able to capture acoustic data with a high SNR despite broadband noise from the glider's buoyancy engine and mechanical noise from the fin. However, the sound recordings close to the water's surface and at the glider's deepest point must be discarded prior to URN analysis. The glider pitch motor assembly in our deployments did not present as a significant separate source of self-noise as the motor was only activated for a very short duration at the beginning of the dive or climb, coinciding with the activation of the pump. Under different operational scenarios, the impact of the pitch-motor assembly on the glider's self-noise must be further assessed. Moreover, resulting from the increase in glider position uncertainty, acoustic gliders are unsuitable for quantifying the source level of ship noise in harsh environmental conditions despite their potential for URN monitoring and identification.

To ascertain the use of the glider CTD data in getting a better PL model of underwater sound, we investigated six PL models. We examined sound PL in shallow and deep water environments, considering factors such as range-dependent SSPs and bathymetry profiles. The study found that the numerical model parabolic equation is robust in estimating PL loss, especially at frequencies below 250 Hz. The simplified mathematical models were more effective at higher frequencies above 250 Hz. The range-dependent SSP and bathymetry profile impacted the PL values significantly in the shallow water environment compared to deeper water. The SSP profile had a higher impact with large ranges between the source and receiver. The PE model using range-dependent bathymetry was 6–10 dB higher than the range-independent model below 100 Hz. Assumptions and approximations become inaccurate at higher frequencies, making them less effective for handling shorter wavelengths in varying environmental conditions. The deep PE model shows minor changes when using range-dependent bathymetry compared to the range-independent approach. Therefore, using range-dependent oceanographic data, the glider provides additional information in shallow water coastal environments, significantly impacting the understanding of sound source PL. This is important for applications in which gliders are used such as to monitor shipping zones for anthropogenic noise pollution in coastal environments or identifying marine mammals.

We used the two wing-mounted hydrophones on the glider to identify the relative DOA or bearing from the glider to the target vessel. The SNR, hydrophone spacing, and acoustic characteristics of the moving noise source most significantly impact the effectiveness of the calculated DOA. Our analysis established an effective frequency range for DOA estimation, which is critical to avoid spatial aliasing and enhance the accuracy of source localization. At two different speeds, the glider successfully detected the vessel's DOA, allowing us to estimate the vessel's heading relative to the glider under certain conditions. In the future, adding a third hydrophone could improve the accuracy of DOA estimates by estimating the elevation and providing ranging information to better localize sound sources.

See the supplementary material for plots of the PL error between the PE, WNI, ECA, SCA, and M-A compared to the ISO-17028-2 method as a reference for shallow and deep water; plots of the glider's internal mechanisms operation with respect to the glider depth; plots of the glider's internal mechanisms operation with respect to the glider depth; for a description of the glider flight during the two sea trails; a description of the glider flight during the two sea trails; plots of the glider's internal mechanisms operation with respect to the glider depth; oceanography data collected as function of depth and range; plots of the PL error between the PE, WNI, ECA, SCA, and M-A compared to the ISO-17028-2 method as a reference for shallow and deep water; and plots of the PL error between the PE, WNI, ECA, SCA, and M-A compared to the ISO-17028-2 method as reference for shallow and deep water.

This research received funding from the Ocean Frontier Institute projects Future Ocean and Coastal Infrastructures (FOCI) and Transforming Ocean Observations. Research funding was provided by the Ocean Frontier Institute, through an award from the Canada First Research Excellence Fund.

The authors have no conflicts to declare.

The data that support the findings of this study are available from the corresponding author upon reasonable request.