An overview of ambient sound using Ocean Observatories Initiative hydrophones Open Access

The Ocean Observatories Initiative (OOI) is an ocean observing network providing data from more than 800 instruments. The different sensor networks measure physical, chemical, geological, and acoustic data spanning from the air–sea interface to the seafloor. All of the data that are produced by OOI are available publicly. OOI acoustic data are well posed to contribute to the global, multi-purpose acoustic network envisioned by Howe et al. (2019). In this paper, we provide an overview of acoustic data that are available through OOI from 2015 to 2020.

Currently, there are 11 different hydrophones that are part of the OOI sensor network and have been used to record ambient sound from around the north-east Pacific Ocean since 2015. Five of the hydrophones are low frequency (LF) hydrophones with a sampling rate of 200 Hz. Six of the hydrophones are broadband (BB) hydrophones with a sampling rate of 64 kHz. Two of the BB hydrophones are located at depths of 200 m and are positioned in the sound fixing and ranging channel. The rest of the BB hydrophones and all of the LF hydrophones are located on the seafloor. The hydrophones are mounted on a tripod approximately 40 cm off of the seafloor, and no flow-shields are used. These long-term ambient sound recordings from the OOI network provide the opportunity for many data and experimentally driven advancements in the field of ocean acoustics. In this paper, we will present a long-term statistical analysis of the ambient sound recorded at nine of the 11 hydrophones, as well as highlight acoustic features that are present throughout the six years of data presented in this paper.

Monitoring long-term underwater ambient sound levels has received increasing attention in past years with more datasets emerging that cover ever larger temporal and spatial scales. Along with OOI, another ocean observatory that monitors the north-east pacific is the Ocean Networks Canada (ONC) NEPTUNE observatory (Barnes et al., 2007). Haver et al. (2018) have evaluated data from the Ocean Noise Reference Station network, consisting of 12 autonomous passive acoustic recorders distributed across various locations in the Atlantic and Pacific Ocean. Various datasets have also been used to compare LF sound spectral levels from the 1950s and 1960s to data recorded during the 1990s and 2000s to analyze the effect of increased shipping activity on underwater sound (Andrew et al., 2002; Andrew et al., 2011; McDonald et al., 2008; Širović et al., 2016). Results seem to be highly dependent on the measurement site but generally indicate that the increase in ambient noise due to shipping activity has slowed down in more recent years, compared to the 1950s and 1960s data reported by Ross (2005). These results were also confirmed by Chapman and Price (2011) and Farrokhrooz et al. (2017), who used hydrophone volume and line arrays in their measurement apparatus. Additionally, a statistical analysis of 2 yrs of ambient sound data investigating the shipping and whale sound contributions has been conducted in the Pacific (Curtis et al., 1999). With its 11 hydrophones recording ambient sound since 2015, the OOI provides an excellent complementary dataset to this area of ongoing research.

The OOI hydrophone dataset contains many different notable acoustic features. Some of these features are outlined in detail in Sec. III. A review of previous literature for some of the pertinent acoustic features is provided below.

Due to the difficult nature of directly observing information about marine mammals, passive acoustic monitoring is one of the primary tools used for learning more about marine mammal populations in the ocean. The OOI hydrophone network contains many marine mammal signals that offer the potential for further exploration (see Sec. III). Passive acoustic monitoring of marine mammals is a well explored field (Burtenshaw et al., 2004; Klinck et al., 2012; Menze et al., 2019; Wilcock, 2012) to which the OOI dataset can contribute. One of the primary biological features that is present in the OOI acoustic dataset is seasonal fin whale calls. Fin whale calls and the inter-pulse intervals of these calls have been extensively explored (Soule and Wilcock, 2013; Stafford et al., 2001; Thompson et al., 1992; Watkins et al., 1987; Širović et al., 2013). While there are several types of fin whale calls, the most common call consists of a downward swept signal lasting approximately 1 s, centered around 20 Hz. The presence of energy near 20 Hz can be used to monitor the migration patterns and seasonal fluctuations of fin whale vocalizations (McDonald et al., 1995; Nieukirk et al., 2004). Fin whale calls contain complex temporal and frequency patterns that indicate communication between multiple fin whales (McDonald et al., 1995; Soule and Wilcock, 2013).

Studying wind and rain levels and their effects on ambient sound in the ocean is a well-developed area of investigation (Curtis et al., 1999; Duennebier et al., 2012; Hildebrand et al., 2021; Knudsen et al., 1948; Ma et al., 2005; Wenz, 1962). Two of the six OOI BB hydrophones are accompanied by surface buoys that continuously record data at the air–sea interface, such as wind vectors and rain rates. Those in situ meteorological measurements with high temporal resolution (1 min) provide an excellent opportunity to use the OOI data for studying wind and rain noise in the northeast Pacific, which has recently been done by Schwock and Abadi (2021a) and Schwock and Abadi (2021b).

Sound from commercial ships has long been studied to assess its impact on the oceanic environment and marine life in particular (Hatch et al., 2008; Merchant et al., 2012; Merchant et al., 2014). It is also known that the acoustic signature of ships depend on a variety of factors, such as ship type, size, and velocity (Gassmann et al., 2017; McKenna et al., 2012; McKenna et al., 2013; Scrimger and Heitmeyer, 1991; Simard et al., 2016; Wales and Heitmeyer, 2002). Since some of the OOI hydrophones are located in the vicinity of major shipping lanes, a large number of ship passages can be found in the dataset. Ship types observed in the OOI dataset include merchant ships, research vessels, fishing vessels, and recreational boats. Very recently, Dahl et al. (2021) explored the effects of the COVID-19 pandemic on ship noise using the OOI data.

The remainder of this paper is outlined as follows: Sec. II describes the OOI hydrophones and data processing framework employed in this work. Section III presents spectrograms showing long-term sound spectral levels as well as specific acoustic features that can be found in the OOI hydrophone data. Sections IV–VII then analyze mean spectral level time series, sound distribution, noise floor time series, inter-hydrophone cross correlations, and power spectrum covariance matrices extracted from the long term spectrograms. Section VIII discusses possible areas of future investigation using the OOI dataset. Finally, Sec. IX summarizes the results of this research.

The OOI acoustic dataset consists of 11 hydrophones located off of the north-west coast of the continental United States. There are five low frequency (LF) HTI-90-U hydrophones (High Tech, Inc., MS) and six broadband (BB) icListen HF hydrophones (Ocean Sonics, Nova Scotia, Canada) with sampling frequencies of 200 Hz and 64 kHz, respectively. For the purpose of this paper, acoustic data analysis from all LF hydrophones and four of the six BB hydrophones will be presented. Two BB hydrophones are omitted because prolonged, strong interfering signals from other measurement instruments co-located with the hydrophones obscured the acoustic records. Figure 1(a) shows the geographic location of the nine hydrophones. Figure 1(b) shows the measured sound speed profiles for different water column profilers in the OOI dataset along with the depths of the different hydrophones. The profiles used to measure the sound speeds are often unreliable at large depths, which results in gaps in the data. The lighter traces for Axial Base and Oregon Slope are a linear regression from valid data below the thermoclines. Data availability for the nine hydrophones analyzed in this paper is shown in Fig. 2. Table I shows the depth and location of all nine hydrophones.

OOI has a web based data explorer (Ocean Observatories Initiative, 2021) that has many of the data products easily accessible. Unfortunately, due to the high file size of the acoustic datasets, hydrophone data are not available through the OOI Data Explorer and must be accessed through the OOI raw data server (BB hydrophones) or IRIS (Incorporated Research Institutions for Seismology, 2021) data server (LF hydrophones). These data servers save the data in mseed format, and it is not straightforward to use the raw data server for analysis.

In order to make the OOI acoustic dataset more easily accessible and to expedite the research process using OOI datasets, a public Python package was developed called OOIPy (Schwock et al., 2021). OOIPy handles accessing the data from the raw data servers and converts the data into formats that are within the Python scientific computing framework. OOIPy also has several data processing methods that aid in the statistical analysis of the acoustic dataset. The Welch mean and median power spectral density (PSD) estimates using calibrated hydrophone data can be calculated within OOIPy, and these PSDs can be grouped together to form spectrograms.

OOIPy also handles the calibration of hydrophone data. For the broadband (BB) hydrophones, the frequency responses of the devices were recorded by the manufacturer in a water tank for frequencies between 0 and 200 kHz. For the low frequency hydrophones, a single sensitivity value is reported which is valid between 2 and 90 Hz. Calibration data sheets for individual OOI hydrophones can be found at (Github, 2021). The 2 Hz cutoff for the LF hydrophones is the lowest frequency to which the hydrophones are calibrated; however, there does not seem to be a physical filter with this cutoff present in the data. The 90 Hz cutoff is due to the digital filtering required to down-sample the acoustic data to a sampling frequency of 200 Hz. This cutoff is characterized by the response files available for specific hydrophones (Incorporated Research Institutions for Seismology, 2021).

The description of ambient sound and analysis of long-term trends and seasonal patterns in this paper is based on the computation of PSD estimates. To do so, the multi-year acoustic time series obtained from each hydrophone are divided into blocks of 4096 samples for the BB and 512 samples for the LF hydrophones, whereby adjacent blocks overlap by 50%. Afterwards, each block is multiplied by a Hann window data taper to reduce spectral leakage and the magnitude square of the fast Fourier transform is computed. That is, for each block. a modified periodogram is estimated. A window length of 4096 samples for the BB data, combined with a median averaging as described below, was chosen. The window length was chosen as a trade-off between spectral resolution and robustness against interfering signals, particularly from acoustic Doppler current profiler (ADCP) pings.

The resulting periodograms for each hydrophone are stacked together to obtain long term spectrograms. As the resulting amount of data is generally intractable for post-processing and evaluation, different temporal scales can now be achieved by averaging a certain number of periodograms together to achieve a single (compressed) PSD estimate. The averaging time we applied varies depending on whether a “global” representation of the data or a detailed view on a specific acoustic feature was desired. For the long-term spectrogram in Figs. 3 and 4, for example, averaging was conducted over 15 min periods, which is equivalent to 28 124 and 702 periodograms for the BB and LF hydrophones, respectively. Furthermore, we always employ median averaging instead of mean averaging to mitigate the effect of outliers (Schwock and Abadi, 2021c), which are particularly strong for the BB hydrophones due to ADCP pings. The resulting spectral estimates can therefore be regarded as Welch median PSD estimates (Schwock and Abadi, 2021c; Welch, 1967). It is noted that the frequency dependent sensitivity correction and the median averaging over multiple periodograms is conveniently implemented in OOIPy.

Due to the size of the BB raw dataset, cloud computing was employed to compute the BB long-term spectrograms. Thereby, a separate Microsoft Azure virtual machine (Standard A2m V2 with two CPUs and 16 GB of RAM) has been used for processing a 3 month chunk of data. In total, 18 machines were running concurrently and the processing time for each machine was between 2 and 5 days, depending on the data coverage within the respective 3 month time span.

Figure 3 shows the long-term spectrograms for the four broadband (BB) hydrophones investigated in this paper. A 15 min Welch-median power spectral density (PSD) estimate is calculated with a Hann window, 4096 FFT (fast Fourier transform) points and 50% overlap. Figure 4 shows the long-term spectrograms for the low frequency (LF) hydrophones. A 15 min Welch-median PSD estimate is calculated with a Hann window, 512 FFT points, and 50% overlap.

For all BB, long-term spectrograms shown in Fig. 3, spectral levels decrease with increasing frequency. Temporal patterns of the ambient sound, however, clearly differ between locations and also depend on the frequency band. Further investigation into the time dependence of spectral levels is presented in Sec. IV.

The LF spectrograms shown in Fig. 4 display several notable features. Around 1 Hz, there is a spike of energy that is due to microseisms (Webb, 1992). At 20 Hz, seasonal vocalizations of fin whales can clearly be seen at all five hydrophone locations. The high prevalence of fin whale signals in the LF hydrophones could lend itself to further studies of fin whale migration using passive acoustic monitoring. In April 2015, the Axial Seamount Volcano erupted. This eruption can be seen in the Axial Base, Central Caldera, and Eastern Caldera spectrograms as a BB spike. In July and August 2019, there was a seismic reflection survey that was conducted directly over the Axial Seamount Volcano (Arnulf et al., 2019). A spike in energy can be seen at all five locations, with the effect being much clearer for the Axial Base, Central Caldera, and Eastern Caldera locations.

Besides analyzing long-term patterns in ambient sound, many different specific acoustic features can also be observed in the OOI hydrophone dataset. Some examples of resolvable features include rain, wind, marine mammals, ships, earthquakes, and volcanic activity. This rich collection of many different types of acoustic features lends itself to many acoustic monitoring applications. Examples of some prominent acoustic features are shown in Fig. 5. Color scales for individual spectrograms are different.

An example of a ship passing directly over the Axial Base LF and BB hydrophones is shown in Figs. 5(a) and 5(b). Using AIS data, this ship was determined to be a cargo ship of length 189 m that passes within 4.839 km of the hydrophone with an approximate speed of 12 knots. The V-shape in the ship passage spectrogram described in McKenna et al. (2012), can be clearly seen in Fig. 5(b). At the time of nearest arrival, the recorded sound intensity at 50 Hz is 94.731 dB relative to $μPa2/Hz$⁠. Using a simple spherical spreading model, the range of the ship at nearest arrival, and the depth of the hydrophone, the source level of this ship is estimated to be 169.54 dB relative to $μPa2/Hz$ at 50 Hz.

A spectrogram of four air-gun shots recorded by the Axial Base BB hydrophone from the 2019 seismic reflection survey (Arnulf et al., 2019) is shown in Fig. 5(c). The research vessel is 14.5 km from the Axial Base hydrophone during the recording. The source level of the air-gun shots is approximately 225 dB at 1 m relative to $μPa2/Hz$⁠.

Figure 5(d) shows a sequence of individual fin whale calls and is recorded by the Axial Base LF hydrophone in December 2016. For a single call, the initial call and subsequent echoes are visible in the spectrogram. In McDonald et al. (1995), a study of fin whale vocalizations reveals that different fin whales with different frequency signatures of their calls are likely interacting with each other. This could be an explanation for the alternating frequencies of the calls seen in Fig. 5(d). As seen in Fig. 4, 20 Hz energy from seasonal Fin whale vocalizations are present at all LF hydrophones locations. Additionally, Ragland et al. (2022) found that using noise interferometry between the Eastern Caldera and Central Caldera LF hydrophones, a fin whale chorus can be detected with directionality pointed toward the Bearing Sea. Figure 5(e) shows an example of unidentified marine mammal vocalizations recorded by the Oregon Offshore BB hydrophone in January 2019. As of the publishing of this article, presence of marine mammal vocalizations in the OOI BB hydrophone data remains unexplored. Previous studies have investigated the effects of air-gun experiments on marine mammal populations (Haver et al., 2018; McDonald et al., 1995; Nieukirk et al., 2004). Since there is an air-gun experiment conducted over OOI hydrophones and a large collection of marine mammal vocalizations, the OOI dataset provides the opportunity for further investigation into the effects of air-guns on marine mammals.

A rain event is shown in Fig. 5(f). The maximum rain rate during this spectrogram is 25 mm/h. Three consecutive events of strong wind are shown in Fig. 5(g). Peak wind speeds of those wind events reach values between approximately 13–15 m/s. Rain rates and wind speeds were obtained from surface buoys located in the vicinity of the hydrophones. Schwock and Abadi (2021a) and Schwock and Abadi (2021b) explore effects of rain and wind on the ambient sound using the OOI dataset.

Figure 5(h) shows the eruption of the Axial Seamount volcano in April of 2015. Studies about the 2015 Axial Seamount eruption using OOI data include Caplan–Auerbach et al. (2017), Nooner and Chadwick (2016), and Wilcock et al. (2016). The average spectral density for April 24 between 07:30 and 10:00 UTC and between 1 and 90 Hz is 96.739 dB relative to $μPa2/Hz$⁠. Using spherical spreading and the location of the fissures as reported by Nooner and Chadwick (2016), the source level of the volcano for this frequency band and time segment is estimated to be 184.9 dB relative to $μPa2/Hz$⁠. Figure 5(i) shows a 6.2 magnitude earthquake that occurred 319 km from the Central Caldera hydrophone (data are available from IRIS Earthquake Browser). The earthquake was located at (43°38′41.64″ N and 127°36′ 11.16″ W) and occurred at 9:31:47 UTC.

The OOI hydrophone dataset contains many diverse acoustic features that are observable due to its wide frequency range and spatial distribution. Future investigations that could take advantage of the features outlined here are discussed in more detail in Sec. VIII.

In order to investigate long-term trends in the acoustic environment sampled by the OOI hydrophones, long-term median time-series for different spectral bands were investigated. Figure 6 shows the times series of the monthly median spectral levels in the 100 Hz, 500 Hz, and 5 kHz one-third octave bands for the four broadband (BB) hydrophones. Figure 7 shows the time series of the monthly median spectral levels in the 20 and 50 Hz one-third octave bands for all five low frequency (LF) hydrophones. Median averaging is used for all time series plots to mitigate the effects of outliers, such as the seismic reflection survey or axial seamount volcanic eruption. Additionally, 25th and 75th percentiles are plotted as shaded regions to communicate the range of possible values for a given month.

None of the BB median spectral level time series show a significant seasonal pattern. Linear trend lines were not computed for the BB hydrophone median time-series, because there is currently not enough data for an accurate estimate. Additionally, the large changes in spectral levels at the lower frequencies are concerning and potentially point to calibration issues with the hydrophones. Issues with calibration are discussed further in Sec. VIII. Most previous studies analyzing long-term ocean sound time series focus on frequencies well below 100 Hz. The few studies that have compared ambient sound levels at higher frequencies (typically 100–500 Hz) over multiple decades found that spectral levels in this frequency range remain largely constant over time (Chapman and Price, 2011; McDonald et al., 2006), which is consistent with our observations at 5 kHz.

In the LF time series plots, the large seasonal fluctuations in the 20 Hz band are due to the seasonal vocalizations of local fin whales. This seasonal variation was previously observed in Andrew et al. (2011). Linear trend lines for the two spectral bands are reported. It should be noted that, due to the variation of ambient sound visualized by the 25th and 75th percentiles and the shorter time of six years that is analyzed, the slopes reported contain error. Small increases in sound levels in the 20 Hz band for the Axial Base, Central Caldera, and Southern Hydrate locations are seen over the six years with a linear slope of +0.34, +0.63, and +0.64 dB/yr, respectively. In the Eastern Caldera and Oregon Slope hydrophone locations, large changes are seen in linear trends of +1.8 and –1.1 dB/yr, respectively.

Energy in the 50 Hz band has contributions primarily from shipping noise (Andrew et al., 2011; Wenz, 1962), and might also contain some fin whale vocalizations (Širović et al., 2013). The 50 Hz trend line slopes for the Axial Base, Central Caldera, Eastern Caldera, Oregon Slope, and Southern Hydrate are -0.57, –0.055, +0.35, –1.1, and +0.15 dB/yr. Andrew et al. (2011) reports a decade long trend in the 50 Hz band of -0.26 dB/yr, for a hydrophone near 45 N (hydrophone “h” in their study). This trend was recorded by a hydrophone located on the continental shelf off of the coast of Oregon.

Our work suggests that, between 2015 and 2020, the linear trends of spectral levels in the 50 Hz band vary from location to location, but generally are around ±0.5 dB/yr. The noted exception is the Oregon Slope location, where a large change of –1.1 dB/yr is seen. The spectral levels reports by Andrew et al. (2011) seem to be in a similar range for the different hydrophones that they study.

The distribution of ambient sound spectral levels can be described by spectral probability density functions (SPDFs) as initially proposed by Curtis et al. (1999) and formally defined by Merchant et al. (2013). SPDFs are obtained from the long-term spectrograms of the broadband (BB) (Fig. 3) and low frequency (LF) (Fig. 4) hydrophones by computing histograms for each frequency bin. Plotting the histogram values as a function of frequency and spectral level gives the desired SPDFs, which are shown in Figs. 8 and 9 for the BB and LF hydrophones, respectively.

The SPDFs for the Oregon Shelf, Oregon Offshore, Oregon Slope, and Axial Base locations are shown in Fig. 8 along with various percentiles (black lines). A feature that all locations have in common is that spectral levels decrease with increasing frequency and the spread in spectral level (i.e., the difference between lowest and highest spectral level at a given frequency) is typically below 40 dB. However, the exact shape of the SPDFs and the trajectories of the corresponding percentiles differ significantly between the locations. The Oregon Shelf and Oregon Offshore percentiles have a similar trajectory as the wind noise spectral levels in Fig. 4 of Schwock and Abadi (2021b). This suggests that wind is the dominant factor in the general ambient sound. Other sound sources, such as marine mammals and rain events, can also impact the ambient sound distribution. However, as the SPDFs are generated from 15 min Welch median PSD estimates requiring a source to be present for at least 7.5 min to significantly affect the PSD estimate, we speculate that those sound sources have a rather small effect on the distribution.

For lower frequencies (below approximately 1 kHz), we assume that nearby ship passages and distant shipping activity have a significant effect on the SPDFs. While (infrequent) nearby ship passages would mainly affect the higher percentiles, the acoustic signature from distant shipping can also increase the spectral levels of the lower percentiles. This is most notable for the Oregon Offshore location where the 1st percentile spectral level around 50 Hz is about 15–20 dB higher compared to the other locations. Such a low frequency (LF) peak has also been reported in other studies and is thought be a result of sound from distant shipping and high latitude winds that travels large distances via the deep sound channel (Anderson, 1979; Bannister, 1986; Wagstaff, 1981; Wenz, 1962). Therefore, a peak at low frequencies indicates that the location has access to the deep sound channel. While the Axial Base and Oregon Slope sound spectral levels show significant LF peaks only in the higher percentiles, the Oregon Offshore hydrophone exhibits a LF peak also in the lowest percentiles. This suggests that only Oregon Offshore has access to the deep sound channel year around, which is likely a result of the shape of the sound speed profile at this location.

Figure 9 shows the SPDFs for all six LF hydrophones along with various percentiles (black lines). Except for frequencies around 20 Hz, the SPDFs for all hydrophones are concentrated around 60–70 dB with a spread (i.e., the difference between lowest and highest spectral level at a given frequency) of about 20 dB for most locations and frequencies. The more spread out densities around 20 Hz are caused by the seasonal fluctuations of local fin whale calls that can also be seen in the long-term spectrogram plots in Fig. 4. Furthermore, in July and August 2019, there was a seismic reflection survey conducted directly over the Eastern Caldera, Central Caldera, and Axial Base hydrophones. This survey is also visible in Fig. 4. The seismic reflection survey is the cause of the 99th percentile having much larger values for these three hydrophones. If the seismic reflection survey is removed from the data, the 99th percentile for Axial Base, Central Caldera, and Eastern Caldera has similar values to the other two LF hydrophones.

It is noted that LF spectral levels at Oregon Slope are approximately 3 dB higher than BB spectral levels at the same location and comparable frequencies. We speculate that (1) differences in geometry of the ocean bottom and (2) the behavior of the Welch median estimator in the presence of outliers are responsible for this phenomenon. It can be shown that the Welch median estimator becomes increasingly biased when a larger percentage of periodograms are affected by outliers (Schwock, 2021). As the number of periodograms for each 15 min PSD estimate differs for the LF and BB data (see Sec. II), a larger percentage of outliers in the LF PSD estimates could result in an overestimation of the LF spectral level.

For frequencies between 1 and 800 Hz and depths of approximately 1–5 km, noise floor spectral levels (i.e., spectral levels in the absence of ships and marine life and when wind is low) in the eastern North Pacific have been recently analyzed by Berger et al. (2018). Instead of focusing on the spectral trajectory of the noise floor, here. we analyze how the noise floor spectral levels change over time. To do so, noise floor spectral levels are computed separately for each month during the measurement period that has a data coverage of at least 50%. For this study, we define the noise floor to be the 5th percentile of the SPDF, which was defined in Sec. V. That is, for each month, we first compute a SPDF and extract the 5th percentile (i.e., noise floor) PSD. Afterwards, the noise floor PSDs were averaged into one-third octave bands and concatenated to obtain time series of the noise floor. The results for the broadband (BB) and low frequency (LF) hydrophones are shown in Figs. 10 and 11, respectively. The 1st and 15th percentiles are shaded.

Figure 10 shows the noise floor time series for the 100 Hz, 500 Hz, and 5 kHz one-third octave bands at all four BB locations. One can observe that the trajectory of the noise floor time series generally follows the trajectory of the mean spectral level time series in Fig. 6. That is, the noise floor remains approximately constant for Oregon Offshore and the 5 kHz band at Oregon Shelf, Oregon Slope, and Axial Base. On the other hand, an increase (decrease) in noise floor can be observed for 100 and 500 Hz at Oregon Shelf (Oregon Slope and Axial Base). Furthermore, no clear seasonal pattern, especially with a period of one year, can be observed. On average, noise floor spectral levels are lower than median spectral levels by about 4.4, 7.9, and 11 dB, at 100 Hz, 500 Hz, and 5 kHz, respectively. While those results can serve as a starting point to analyze the noise floor in the northeast Pacific Ocean, more data are necessary to draw better conclusions. Fortunately, the OOI will provide more BB hydrophone data in the upcoming years, which, together with the data shown in this work, can be used to obtain more definitive results.

Figure 11 shows the noise floor time series for the LF hydrophones. The 20 Hz band still contains the primary fluctuations due to seasonal fin whale vocalizations. This indicates that when fin wales are present, they dominate the ambient sound by vocalizing throughout the time. The noise floor in the 50 Hz one-third octave band on the other hand, stays roughly constant around 60 dB with only weak seasonal fluctuations. Furthermore, some locations show weak seasonal trends in the 20 and 50 Hz one-third octave band. Those trends generally follow the pattern of the median spectral level time series in Fig. 7. That is, while spectral levels increase over time for the Axial Base, Central Caldera, Eastern Caldera, and Southern Hydrate, they decrease at the Oregon Slope.

Time series extracted from the long-term spectrograms in Figs. 3 and 4 are used to compute cross correlations between different hydrophones. To do so, a time series at a single frequency bin from one hydrophone is correlated with the time series at the same frequency of another hydrophone. This process is repeated for all frequency bins and hydrophone combinations. The resulting frequency dependent cross correlations averaged into one-third octave bands (using median averaging to mitigate the effect of outliers) are shown in Fig. 12. The distances between hydrophone pairs are given in parentheses. Additionally, power spectrum cross-covariance matrices are shown for the LF hydrophones locations in Fig. 13.

Figure 12(a) shows that the correlation of ambient sound levels between the different broadband (BB) hydrophones is highly frequency dependent. With the exception of the Oregon Slope–Axial Base combination, correlation values are usually highest between 1 and 10 kHz and increase with decreasing distance between the hydrophones. That is, the highest correlation for frequencies above 1 kHz can be observed between the Oregon Slope and Oregon Offshore hydrophone, which are separated by only 38 km. On the other hand, sound between Oregon Offshore and Axial Base, as well as Oregon Shelf and Axial Base, which are separated by 411 and 447 km, respectively, is less correlated. The question of which factors contribute the most to the cross correlation curves in Fig. 12(a) is beyond the scope of this article. However, we speculate that a main contributor could be the (dis)similarity of wind conditions at different locations at the same time. It is known from Schwock and Abadi (2021b) that spectral levels between 1 and 10 kHz are highly correlated to the wind speed and neighboring locations are more likely to have similar wind speeds at the same time.

For frequencies below 1 kHz, correlations are typically much lower than in the high frequency range. The only exception are the Oregon Slope and Axial Base hydrophone, whose correlation increases with decreasing frequency and assumes values above 0.6 for frequencies below 100 Hz, despite a separation of more than 370 km. A possible explanation is that low frequency sound is able to propagate between these two locations without experiencing significant attenuation. As both hydrophones are located in depths of more than 2500 m, there could exist propagation paths that do not interact with the sea surface or only do so in a limited way, therefore resulting in less sound attenuation. Furthermore, the geometry of the seafloor at and between the two hydrophones may also crucially influence the sound propagation. However, investigating this behavior in more detail is beyond the scope of this article and left for future research.

The cross correlations for the low frequency (LF) hydrophones are shown in Fig. 12(b). Generally, the correlations increase with decreasing inter-hydrophone distance. High correlation values around 20 Hz can be traced back to the seasonal migration patterns of fin whales, which are also visible in the long-term spectrograms in Fig. 4 for all five LF locations. For frequencies at and below 1 Hz, the high correlation values can be attributed to the occurrence of microseisms that often affect the sound levels at multiple hydrophones at the same time. Between those frequency bands with high inter-hydrophone correlation, the correlation values dip and reach a minimum around 3–4 Hz for most hydrophone combinations. This suggests that there are no ubiquitous sound sources in this frequency range. As for the BB hydrophones, a more thorough analysis of the inter-hydrophone sound correlation remains a topic for future investigation.

In order to investigate the correlation between different frequency bins in the power spectral densities (PSDs) over the six years of data analyzed, the power spectrum density covariance matrices are computed for the LF hydrophones and shown in Fig. 13. There is a large correlation between 15 and 25 Hz, which is expected due to fin whale vocalizations. Additionally, other frequencies do not seem to be correlated to the 20 Hz frequency band. This is likely explained by fin whale vocalizations dominating this frequency band and not vocalizing in other frequency bands. Frequencies between 40 and 90 Hz seem to have higher correlation which is expected due to shipping noise. Frequencies between 2 and 15 Hz have different levels of correlation depending on the location. The correlation in this frequency band could be caused by seismic or volcanic activity. Notably, the correlations in this frequency band are different for the Axial Base, Central Caldera, and Eastern Caldera hydrophones which are located on the active Axial Seamount volcano, as compared to the Oregon Slope hydrophone and the Southern Hydrate hydrophones. There are several other notable features, which are not easily explained. There appear to be “rays” of correlation with a slope not equal to one. These “rays” are present in all hydrophones but are easiest to see in Figs. 13(d) and 13(e). These “rays” indicate that sound levels at a single frequency are correlated with sound levels at a single, lower frequency. The correlation between single frequencies scales linearly, but not one to one. Additionally, there are diagonal lines just above and below the major diagonal for frequencies above 30 Hz. Further investigation into the power spectrum covariance matrices for different LF hydrophone locations is saved for future work.

There are still many potential areas of research that can utilize the OOI data resources. Using some of the dataset features highlighted throughout this paper, we discuss several areas of potential future investigation, and propose several improvements that could be made to the OOI acoustic dataset.

The OOI hydrophone network provides a great opportunity for future development in the field of ocean acoustics. The authors would like to note several areas that would further the usefulness of this dataset for future consideration. ADCP pings contaminate a significant portion of the broadband data, and moving hydrophones to be farther away from these measurement devices could result in better quality data. Another major area of improvement that is needed from OOI is the calibration of the hydrophones. Documentation for hydrophone calibration is not easily available, and specific calibration sheets for individual LF hydrophones are not available. Additionally, large fluctuations in the BB spectral levels (as seen in Fig. 6) raise concerns about the calibration of the BB hydrophones. Future calibrations could benefit long-term ambient sound research. The LF hydrophones are not calibrated below 2 Hz, but there appears to be useful information in the data that is present below 2 Hz. Additionally, re-calibrating the hydrophones over time could help long-term analysis compensate for calibration drift with time. Last, the addition of even more acoustic sensor types, such as directional arrays or distributed acoustic sensing, would allow for further scientific development using the OOI dataset.

In combination with ancillary datasets, such as from the Automatic Identification System (AIS), the OOI low frequency (LF) and broadband (BB) hydrophone data can be used to study sound from commercial ships in greater detail. For example, the hydrophone data, along with modern techniques from statistical and deep learning, can be used to develop algorithms for ship classification and path tracking as shown by Niu et al. (2017). Furthermore, by comparing LF spectral levels associated with commercial shipping and other man made activity with past measurements, such as from Andrew et al. (2011), the continuing impact of anthropogenic sources on underwater sound can be assessed and predictions for the future can be made. Additionally, the prevalence of ship passes in the data could provide the opportunity for further investigation into the feasibility of using ship passes for tomographic inversion.

The OOI BB hydrophone data can be used to detect rain events and estimate wind speeds and rain rates from passive underwater acoustic recordings. Respective algorithms have been proposed by Ma et al. (2005) and Vagle et al. (1990), as well as more recent examples using machine learning (Taylor et al., 2021; Trucco et al., 2022). However, deep learning techniques have not yet been fully exploited for such tasks. In situ wind and rain measurements at Oregon Shelf and Oregon Offshore can be used to evaluate existing algorithms developed using data at different locations, as well as to develop new deep learning algorithms for remote sensing of wind and rain as well as weather prediction applications.

As the OOI hydrophones will continue collecting data over the next years, the mean spectral levels and noise floor time series shown in this article can soon be extended to better assess long-term changes in ambient sound and noise floor levels. The amount of data in the OOI dataset can also be used for tasks, such as time series forecasting, with the goal of predicting future ambient noise levels. This has previously been studied by Alvarez et al. (2001) using genetic algorithms on a small dataset collected in the Strait of Sicily, Italy. Expanding this approach to a larger scale using LF and BB OOI hydrophone data is a promising direction for future research. The low-frequency hydrophone calibration information is only available above 2 Hz. In future works, relative analysis, similar to the analysis of Curtis et al. (1999), in combination with seismometers co-located with the LF hydrophones, could be used to further study long-term trends of low frequency ambient sound that is below 2 Hz.

While we focus on hydrophone data in this article, the OOI provides a variety of other data products collected at the air-sea interface, water column, and seafloor. Therefore, the effects of environmental parameters, such as temperature, salinity, conductivity, and others, on the acoustic data can be studied using OOI data. Additionally, OOI acoustic data could be used in tandem with other ocean observatories data, such as the Ocean Networks Canada (ONC) Neptune observatory (Barnes et al., 2007), which also studies the north-east Pacific.

Ambient noise interferometry is the method of using ambient sound to estimate the time domain Green's function between two points. Ragland et al. (2022) demonstrate the viability of using ambient noise interferometry between the Eastern Caldera and Central Caldera LF hydrophones to resolve direct and multi-path acoustic arrivals between the two hydrophones. Further developments could include using ambient noise interferometry, along with other OOI sensors, to develop ways to passively estimate ocean variables, such as sound speed, using only ambient sound. The viability of ambient noise interferometry is very dependent on the specific acoustic environment of the two hydrophones (Skarsoulis and Cornuelle, 2019). Since the OOI hydrophone network contains many different hydrophones with differing environments, the effects of the environment on ambient noise interferometry could be experimentally explored in greater detail using the OOI dataset.

As highlighted in Sec. III A and also in Figs. 4 and 5, marine mammal vocalizations are present throughout the OOI dataset. Weirathmueller et al. (2017) have already used seismic data to study fin whale vocalizations in the region. The OOI dataset could be utilized to further investigate the behaviors of different marine mammal species in the north-east Pacific. Given the size of the OOI acoustic dataset, and the presence of marine mammal vocalizations throughout, this dataset could also lend itself to machine learning methods for learning more about marine mammal populations and the meanings of their vocalizations.

In Sec. VII, the inter-hydrophone cross correlation for different spectral levels is explored. Explanations for the different correlations that are observed are still largely unknown. Further investigation could help explain some of the correlations that are observed. Moreover, by computing power-spectrum covariance matrices for each individual hydrophones, it would be possible to study the frequency ranges of characteristic noise sources, such as ships or marine mammals. Additionally, analyzing spectral correlations in time could potentially reveal time-delayed correlations of different spectral bands. This technique could possibly be used to study fin whale migration patterns.

As machine learning tools become more popular and more powerful, the need for large amounts of data continues to grow. OOI provides a significant opportunity for developing machine learning tools within the field of ocean acoustics. Many specific areas to which machine learning could be applied have been mentioned in Secs. VIII A–VIII H. However, one of the largest limiting factors for machine learning in ocean acoustics is acquiring labeled datasets. Oftentimes, labelling data in ocean acoustics is non-trivial and requires expert knowledge. A potential future area of investigation in machine learning for ocean acoustics is using unsupervised machine learning to learn dimension reductions for ambient sound, which could be used to learn complex patterns within data. This technique could then be used for applications, such as ambient sound annotation, or remote sensing. This endeavor requires a large amount of data in order to be successful. Given the large size of the OOI acoustic dataset, it could potentially lend itself well to this type of investigation.

In this paper, we have presented an overview of the publicly available OOI dataset. We have highlighted prominent features that are within the dataset. Long-term trends in the 50 Hz band and median spectral levels were found to vary from hydrophone to hydrophone. The 20 Hz band showed strong seasonal patterns as previously observed by Andrew et al. (2011) and also had varying linear trends depending on the hydrophone site. Long-term trends in the 100 Hz, 500 Hz, and 5 kHz were found to vary significantly between hydrophone locations as well. Spectral probability density functions were used to show that ambient sound distributions highly depend on frequency band and location. Long-term noise floor analysis reveals that the trends observed in the median spectral level time series are similar to the trends for the 5th percentile in the spectral bands that are investigated. Inter-hydrophone spectral correlations reveal that for broadband hydrophones, spectral levels between 1 kHz and 10 kHz show the highest level of correlation between hydrophones. Low frequency hydrophones see the highest levels of inter-hydrophone correlations near 2 and at 20 Hz due to microseisms and fin whale vocalizations, respectively. In general, correlation between hydrophones is stronger for shorter distances. Last, we discuss potential future areas of investigation using the OOI dataset given the acoustic features presented throughout this work.

This research was supported by the Office of Naval Research Grant No. N00014-19-1–2644. The authors would like to thank the Ocean Observatories Initiative and all people involved therein for deploying and maintaining the instruments and making the collected data publicly available.