Measuring the Marine Soundscape of the Indian Ocean with Southern Elephant Seals Used as Acoustic Gliders of Opportunity
1. Introduction
1In the frequency band from a few tens of a hertz up to 50 kHz, the dominant sources of ambient noise in the ocean can be broadly divided into sounds resulting from geophony (i.e., sounds from natural physical processes, e.g., wind-driven waves, rainfall, seismicity, breaking waves, current), biophony (i.e., sounds from biological activities, e.g., whale vocalizations, snapping shrimp beds), and anthropophony (i.e., man-made sounds, e.g., commercial shipping, sonar, seismic prospecting, oil and gas surveys) (Knudsen et al. 1948; Wenz 1962). All these sources contribute conjointly to the noise spectrum characteristics (e.g., pressure level, spectral slope) in varying degrees, depending on their strength and conditions prevailing at the measurement location. Thus, the underwater ambient sound field contains quantifiable information about the physical and biological marine environment. To extract this information, passive acoustic systems have been used for monitoring, recording, and interpreting in a continuous and autonomous way the underwater acoustic signal, facilitating an all-weather and all-season ocean monitoring.
The study of ocean ambient noise plays a growing role in many different research fields. In biodiversity, it helps in preserving marine animal ecosystems by better understanding the impacts of human activities on their ecology (Sirovic et al. 2013). In meteorology, global climate models and local weather forecasts rely on field information about weather across oceans. Observations of rain and wind phenomena from underwater noise allow for better study of air–sea interactions, and increase greatly the spatiotemporal resolution provided by satellite (Vagle et al. 1990; Nystuen and Selsor 1997; Ma and Nystuen 2005; Pensieri et al. 2015). In cryogenics, the noise generated by glaciers allows for quantification of melting processes in the Arctic and is a good indicator of rapid climate processes (Urick 1971; Glowacki et al. 2015). In oceanography, measurement and characterization of ambient noise are essential to enhance the signal-to-noise ratio of acoustic-based underwater instruments (Rahmati et al. 2014). The need for better assessment of global change and its consequences have drawn attention and highlighted the need for an intense monitoring of underwater noise level and, consequently, for the development of innovative sensors and networks (Duennebier et al. 2002; Johnson and Tyack 2003; Aguzzi et al. 2011; Favali 2013) able to collect and analyze long-term underwater sound data.
Ambient noise studies mostly take place in the northeastern Pacific (Chapman and Price 2011) and Atlantic Oceans (Nieukirk et al. 2004). Previous studies in the Indian Ocean have focused on the northwestern (Wagstaff 2005) and the tropical regions of the Indian Ocean (Miksis-Olds et al. 2013; Hawkins et al. 2014; Tsang-Hin-Sun et al. 2015). Tournadre (2014) showed that the ship traffic has had a global increase in the Indian Ocean in the last two decades. In the Southern Ocean, a recent study highlighted the predominant role of icebergs in the Southern Hemisphere soundscape (Matsumoto et al. 2014). Miksis-Olds et al. (2013) reported that the observed sound floor increases are consistent with concurrent increases in shipping, wind speed, wave height, and blue whale abundance in the Indian Ocean. Nair et al. (2015) developed a semiempirical model to predict surface ambient sound spectra in 1–50 kHz for rainfall rates in 2–200 mm h−1 and wind speeds within 2–14 m s−1.
Most often, hydrophones used for marine soundscape studies are bottom mounted, shore terminated, and fixed at a certain depth. Other studies have also used mobile hydrophones, either dragged behind a drifting buoy or boat (Nystuen and Selsor 1997), or attached to vertical profiler float (Ward et al. 2011; Küsel et al. 2011; Barclay and Buckingham 2013). More recently, with the development of miniaturized electronic devices, hydrophones have been embedded in underwater gliders (Baumgartner et al. 2008; Matsumoto et al. 2011; Klinck et al. 2012). These gliders can survey a large area by autonomously navigating the defined area (Rogers et al. 2004; Rudnick et al. 2004), and they have shown great promise in monitoring marine mammals (Baumgartner et al. 2008; Klinck et al. 2012), oceanographic phenomena (Matsumoto et al. 2011), and meteorological surface conditions (Cauchy et al. 2015).
In the same line of technological innovation, the use of animalborne autonomous recording tags, called biologging, is becoming widespread (Ropert-Coudert and Wilson 2005), and allows for the acquisition of huge quantitative datasets for inferences on movement, ecology, physiology, and behavior of animals moving freely in their natural environment. Multichannel dataloggers are used, and data are sampled at high resolution over large temporal and spatial ranges, including geographical areas uncovered by satellite data. In addition to providing parameters related to the animal biological processes, environmental parameters (e.g., temperature, salinity, light, fluorescence) can also be continuously recorded.
In this project, southern elephant seals (SES) of the Kerguelen Islands are used as acoustic gliders of opportunity. SES are wide-ranging animals during their postbreeding and postmoulting migrations. Adult females (Mirounga leonina) from the Kerguelen Islands (49°200′S, 70°200′E) forage mainly in oceanic waters of the Antarctic and polar frontal zones (below 60°S) from October to February (Bailleul et al. 2010). Among top marine predators, air-breathing diving species such as SES are particularly well suited for biologging because their large size allows them to carry electronic devices with minimal disturbance. These devices are stuck on SES while they are on land in their breeding colonies. The strong east–west current speeds and the thick ice surface layers in this part of the austral ocean make the use of regular gliders very complicated, while these harsh environmental conditions are not a problem for SES.
So far, these biologged SES have been used to collect measurements of physical (Charrassin et al. 2008; Costa et al. 2008; Roquet et al. 2009) and biological (Guinet et al. 2013) oceanographic parameters, in often inaccessible regions. The SES and their closed-loop migratory route also provide the opportunity for using Acousondes1 (Acoustimetrics, Greeneridge Sciences, Inc., Felton, California) that can be retrieved at the end of their migration. Acoustic data have already been recorded and used to investigate behavioral and ecophysiological (breathing rate) parameters (Genin et al. 2015). While at sea, SES dive repeatedly to mesopelagic depths (300–500 m up to 2000 m) and tend to follow the diel vertical migration of their mesopelagic prey, diving generally deeper during the day (Guinet et al. 2014). SES regularly perform dives during which they spend a large proportion of time descending passively through the water column (Richard et al. 2014). In the following, this type of dive will be referred to as drift dives.
The main objective of this current study is to demonstrate that marine soundscape can be measured with biologged SES used as acoustic gliders of opportunity. However, the SES movements (e.g., depth/speed variations) impact the measured sound spectra. It is thus first required to identify, characterize, and if possible remove the acoustic noise produced by SES movements that corrupts the measured soundscape. Under the assumption that different sound sources have unique acoustic signatures, passive acoustics can be used to make a classification of the ambient sound field. In this paper, a multivariate acoustic response in a multivariate multiple linear regression (MMLR) framework is used to decompose the measured ambient noise into different acoustic sources (acting as predictors) related either to the environment or to the SES. The ambient noise is modeled as a multivariate response of spectral parameters, namely, sound pressure levels (SPLs) and spectral slopes (SS), in various frequency bands. These acoustic features have been widely used in studies on marine soundscape and acoustical meteorology (Nystuen and Selsor 1997; Ma and Nystuen 2005; Pensieri et al. 2015). Ancillary datasets on wind speed and on SES diving behavior are used to define the predictors. Once we have fully characterized the acoustic noise induced by the SES displacements, we present results on the measured soundscape, and we focus on the effect of wind speed on the ambient noise level. These measures are compared with those made in other ocean environments at comparable depths, and also with theoretical models.
This paper is organized as follows. Section 1 provides details on acoustic measurements and on the different variables used in the regression framework. Section 2 presents exploratory and qualitative results, with an emphasis on the differences between spectra measured during the drift and active swimming phases of the SES. Then, more quantitative results are presented, in particular to evaluate the accuracy of estimating wind speed from our passive acoustic recordings. Section 3 proposes two general discussions on the relations between the external influence variables and the measured spectra, and on the marine soundscape resulting from our acoustic glider of opportunity.
2. Methods
a. Materials
In the austral winter of 2012, five different postbreeding female SES of similar body conditions were captured and equipped with dataloggers on the Kerguelen Islands. These loggers included an Acousonde 3A device (already used in similar research studies; e.g., Burgess et al. 1998; Burgess 2000) glued on the back of the seal on the longitudinal axis, 10 cm behind the scapula. The Acousonde 3A recorded sound at a sampling frequency 
b. Acoustic database
Table 1 provides a global overview of the sound database. The five Acousondes deployed were activated on the field almost at the same time, making their recordings overlap in time. The smaller number of recordings in Acousondes A626021 and A626022 was caused by technological deficiency. To save onboard storage space, a duty cycle was set up in each Acousonde that automatically turned it on for 4 h every 24 h. Each 4-h recording was segmented into 10-s time windows (except for the long-term averaged spectrogram, where we used 30-s time windows), using non-overlapping Hamming windows. We removed all the sound files for which the SES depth was less than 10 m, which includes phases when the SES is on land and phases when the SES is at the surface at sea. All data processing and analysis were conducted using MATLAB. A total of 184.2 h of audio recordings have eventually been used, providing 66 296 observations of 10 s long.
Global overview of the passive acoustic recording database, with C as the cumulated duration. Each recording lasts approximately 4 h.
Figure 1 represents the migratory routes followed by the SES during which the Acousondes were active. The spatiotemporal coverage provided by SES routes ranges from 71° to 87° in longitude and from −46° to −52° in latitude.
Migratory routes of five different SES plotted on a map zoomed in on the Kerguelen Islands (49°200′S, 70°200′E), in the Indian Ocean. Details of individual SES are given in Table 1. These trajectories are based on GPS data from an Argos GPS satellite tag fixed on each SES. Only the segments of routes during which the Acousondes were active are drawn.
Citation: Journal of Atmospheric and Oceanic Technology 34, 1; 10.1175/JTECH-D-16-0124.1
c. Identification of drift phases
As already done in Dragon et al. (2012), pressure data were used to split the ascent and descent dive phases into two different categories based on diving behavior, namely, drift dives and active swimming dives. In our study, drift dives are important, as they are expected to offer the cleanest acoustic measures. Drift dive identification was processed in two steps. First, we used first the complete time–depth recorder (TDR), which allowed us to (i) identify drift dives and (ii) isolate the passive drift phases during those dives (Dragon et al. 2012). For each drift dive, a drift rate was determined as the slope coefficient of a linear regression between depth and time (Bailleul et al. 2010; Mitani et al. 2010). In the second step, we used accelerometer data to exclude phases of active swimming during drift phases assessed by the TDR-only data. Active swimming was considered to take place when lateral acceleration exceeded the −0.2 to 0.2 m s−2 range. It is noteworthy that a more detailed taxonomy could have been used [e.g., Richard et al. (2014) subdivided the active swimming dives into exploratory dives, shallow active dives, and deep active dives], but in this paper diving characteristics of the SES will be limited to this two-class behavior, for the sake of clarity and conciseness with our acoustical meteorology application.
Table 2 provides statistical information on three different dive phases of SES, namely, surface, drifting, and active swimming phases. The R package Biologging tools2 were used to estimate automatically the drift phases of SES. Drift phases have an average duration of 2 min. Overall, they cover only 6% of the total recording period, with a higher presence of more active swimming dives (83%). These estimated drift phases have already been used in previous biological studies (Vacquié-Garcia et al. 2012; Guinet et al. 2014; Richard et al. 2014; Genin et al. 2015).
Statistics on dive phases, with D as the mean phase duration, C as the cumulated duration, and 
d. Regression variables
Table 3 presents details on the different variables used in the regression analysis. They are detailed in the following.
Details of variables used in the regression analysis, classified into four categories: acoustics, SES displacements, SES status dive, and environment. All center frequencies 
1) Acoustic variables
2) SES displacements
Histograms of (a) SES swimming speed, (b) acceleration, (c) depth, and (d) wind speed data matched with the analyzed passive acoustic recordings. Each histogram has been individually unitary normalized. Each observation is labeled either as a drift phase (blue) or an active swimming phase (red).
Citation: Journal of Atmospheric and Oceanic Technology 34, 1; 10.1175/JTECH-D-16-0124.1
3) Environment
The dataset on wind speed was extracted from the Advanced Scatterometer (ASCAT) retrievals3 and provided gridded daily-averaged wind and wind stress fields over global oceans (Bentamy and Croize-Fillon 2012). The calculation of daily estimates uses ascending and descending available and valid retrievals. The objective method aims to provide daily-averaged gridded wind speed, zonal component, meridional component, wind stress, and the corresponding components at global scale. The error associated with each parameter, related to the sampling impact and wind space and time variability, is provided too. When compared with buoy measurements, this error for wind speed is below 2 m s−1 (Bentamy and Fillon 2015). More details about data, the objective method, and computation algorithm can be found in Bentamy and Croize-Fillon (2012). The European Centre for Medium-Range Weather Forecasts (ECMWF) analyses have been used as a temporal interpolation basis of ASCAT retrievals. The resulting fields are 10 m high above the ocean surface, and have spatial resolutions of 0.25° in longitude and latitude, and a 3-h temporal resolution.
An acoustic measurement is a spatially integrated measure, with an area of measurement at the ocean surface (also called the listening area in literature) that depends on the hydrophone depth and on the beam pattern of the ambient noise field. The listening area of a 500-m-depth hydrophone is around 8 km2 (Nystuen et al. 2015), which is well included in the satellite pixel resolution (around 30 km2 at the SES latitudes).
Figure 2d represents the histogram of wind speed data matched with the analyzed passive acoustic recordings. From operational point of view, it is more significant to provide the correct classification of wind speed class than wind speed estimates. We thus identified from these data four different classes of wind speeds, labeled 
e. Multivariate multiple linear regression
1) Model formulation
Parametric nonlinear models also exist, and represent the relationship between a continuous response variable and one or more continuous predictor variables in the form 
In regression analysis terms, two noncorrelated predictors used in combination would predict unique variance in a response, while two more correlated predictors tend to predict shared variance, and so are less efficient. Prior to MMLR, a principal components analysis (PCA) can then be used to decorrelate predictors. PCA is a dimension reduction method that constructs independent new variables that are linear combinations of the original variables, by reducing redundancy (i.e., increasing standard deviation) between all variable dimensions. PCA was performed to see how our different SES and environmental variables were structured within the dataset, and whether the acoustic variables were responding to this structure. PCA also avoids multicollinearity problems in regression analysis (Zhang et al. 2006). Such problems make regression coefficients become unstable when highly correlated predictors are present. This further justifies the use of PCA in our study.
In this project, two different setups of the MMLR method were used. We first studied the contributions of SES and environment variables (i.e., the predictors 
Also, all variables were standardized (i.e., zero mean and unitary variance), which allows a better comparison of regression weights between each predictor (especially with variables having different range values, as for depth and wind speed in our study), as the unstandardized weights are a function of the variance of the predictor variables. Standardization removes most of the correlation between linear and higher-order terms, which reduces the chance of adding these terms unnecessarily.
2) Evaluation metrics
3. Results
a. Exploratory data analysis
Figure 3 shows a typical long-term spectral average computed over four acoustic recordings (i.e., a total of 16 h of recordings from four different days) from the SES individual A626019. SES variables and wind speed values are superimposed onto this spectrogram. It can be seen that the drift and active swimming phases are clearly identified through the acoustic energy in low-frequency bands that correlate well with the SES speed. Also, acoustic energy in higher-frequency bands appear to be reinforced with higher wind speed, independently from other SES-related variables. In other words, two frequency regions of the spectra are mainly impacted, showing first strong evidence of correlations: a first region in frequency subbands below 2.5 kHz, modulated accordingly to SES speed, and a second region above this frequency that is more dominated by wind speed.
Long-term spectrogram averaged over four acoustic recordings (i.e., a total of 16 h of recordings from four different days) from SES A626019. It was generated using 2048-point FFTs, Hamming windows, and no overlap, and averaged every 30 s. The four variables V, 
Citation: Journal of Atmospheric and Oceanic Technology 34, 1; 10.1175/JTECH-D-16-0124.1
To provide more details on the shape of spectra measured during drift phases, Fig. 4 represents an averaged spectrum computed during temporal phases with values of SES speed and acceleration belonging to their 5th percentiles. In the following, these phases will be referred to as extreme drift phases. Also, this averaged spectrum will be associated with our measured soundscape. Regarding median levels across this acoustic dataset, they ranged between 47 and 70 dB ref 1 μPa2 Hz−1. The spectrum slope falls off quickly from its highest level at 50 Hz at a rate of about −5 dB per octave up to 2 kHz, where it becomes flatter up to 5000 Hz, assuming a slope of −5 dB per octave.
Averaged spectrum computed during extreme drift phases, i.e., with values of SES speed and acceleration belonging to their 5th percentile. Two representative theoretical SS (dB per octave) are superimposed onto the spectrum. The approximate signs mean that these slopes are only locally correct in frequency (i.e., the first slope is a decrease of −5 dB within the octave 500–1000 Hz, and the second slope is a decrease of −2 dB within the octave 2500–5000 Hz).
Citation: Journal of Atmospheric and Oceanic Technology 34, 1; 10.1175/JTECH-D-16-0124.1
We then quantified the averaged deformation of spectra due to SES displacements relative to this extreme drifting behavior. In Fig. 5, we computed the differences between averaged logarithmic power spectral density (PSD) with values of SES speed and acceleration belonging to their 20th, 60th, 95th percentiles (dot-plus-circle, dashed–dotted, and solid lines, respectively), and their 5th percentile. From Fig. 5a, we can observe that SES speed strongly impacts the acoustic data for frequencies lower than 2000 Hz, with variations that can reach 40 dB. The three acceleration components contribute to an increase in the measured noise below 800 Hz, with gains from −30 to −10 dB. From Fig. 5b, it can be observed from the dashed–dotted curves that the acoustic distortions from our marine soundscape spectrum above (see Fig. 4) are quite minimal, that is, in the order of magnitude of 5 dB with the SES speed and less than 1 dB with the acceleration variables. Also, we did not find any significant influences of depth on the acoustic features (see Fig. A1).
Differences between averaged logarithmic PSD with values of SES speed and acceleration belonging to their 20th, 60th, and 95th percentiles (dot-plus-circle, dashed–dotted, and solid lines, respectively), and their 5th percentile. (a) The full spectrum, (b) a zoomed-in view of the frequency range [10; 2000] Hz and level range [−8; 8] dB.
Citation: Journal of Atmospheric and Oceanic Technology 34, 1; 10.1175/JTECH-D-16-0124.1
b. Correlation analysis
To provide a more global insight into the correlation structure of our variables, we performed a PCA on the complete variable dataset. Figure 6 shows both the orthonormal principal component coefficients of each variable on the first two principal axes and the principal component scores for each observation (i.e., the coordinates of the original data in the new coordinate system defined by the principal components). In the space of the principal components of the acoustic/SES/environmental variables, the first two principal components distinctly separate the different acoustic features. Indeed, these features move from the first to the second principal axis as their center frequencies increase. In other words, low-frequency SPL observations remain in the bottom-right half-space, while the higher-frequency SPL observations are in the top-right half-space. We can now see how SES and wind speed variables distribute on the fan of SPL values. SES speed V is among low-frequency SPLs with a high score on the first PCA axis and a negative score on the second axis (
Orthonormal PCA for each variable (blue lines), and the principal component scores for each observation resulting from the PCA in the first two PCA axes. For the sake of clarity, all the SPL acoustic features are located at their respective black bullet in increasing order, as illustrated by the black arrow. Also, observations are classified into four different classes of wind speeds, 
Citation: Journal of Atmospheric and Oceanic Technology 34, 1; 10.1175/JTECH-D-16-0124.1
These correlation tendencies were already highlighted through individual pairwise correlations between the acoustic, SES, and environmental features (see appendix B). To further compare the contributions of SES and environmental variables to the measured spectra, a multivariate regression analysis was eventually conducted. Previous correlation analysis (section 1) showed that our acoustic responses are at least moderately correlated, which is necessary for the multivariate regression analysis to make sense. Also, our acoustic database is assumed to be large enough so this analysis is reliable. We then modeled the multivariate ten-dimensional acoustic response with an MMLR model, and used SES and environment parameters as independent predictors. Table 4 shows the details of this analysis, with the regression coefficients 
Results of the MMLR model for the multivariate acoustic response. SES and environmental variables are taken as predictors. Results are reported in terms of regression coefficients, p values, ordinary 
SES speed has coefficients at least around 10 times as superior as the other predictors in the response from SPL(0.05) to SPL(0.2). Wind speed becomes prevalent in the higher-frequency SPL responses, which is consistent with the correlation analysis results above. Above SPL(3.2), the coefficients of wind speed are at least around 4 times as superior as the other variables, and as high as 10 times for SPL(4). The acoustic response SPL(4) is predicted by the equation 
c. Wind dependence of ambient noise
1) Qualitative observations
Representation of four averaged ambient noise spectra corresponding to the different wind speed classes, 
Citation: Journal of Atmospheric and Oceanic Technology 34, 1; 10.1175/JTECH-D-16-0124.1
2) Quantitative validation
In a further experiment, a sample fitting verification was performed to testify the effectiveness of different regression classifiers to predict wind speed. These classifiers were tested on a drift-only dataset and on the complete dataset. A fivefold cross validation with a stratified procedure has been applied to each dataset, meaning that each fold contains roughly the same proportions of observations from the different wind speed classes.
As listed in the first column of Table 5, we first tested a simple linear regression model, using the 
Performance of different regression classifiers to estimate wind speed. Performance is reported in terms of the wind speed estimation error 
As an evaluation metric, we use the rms difference 
An example of wind speed prediction is represented in Fig. 8a using the multiple linear regression model superimposed onto the ground truth. Figure 8b represents the fitting plot of the adjusted model, showing that the model as a whole is significant (i.e., a horizontal line does not fit between the confidence bounds). The slope of this line is the slope of a fit to the predictors projected onto their best-fitting direction—in other words, the norm of the coefficient vector. Wind speed values are displayed in unstandardized (natural) units.
Predictions of wind speeds obtained with an MMLR model, taking acoustic features as predictors. (a) The ground truth (blue) and the rms error of the model (orange) are superimposed. (b) The fitting plot of the adjusted model is represented.
Citation: Journal of Atmospheric and Oceanic Technology 34, 1; 10.1175/JTECH-D-16-0124.1
Table 5 displays for each regression model the average error 
4. Discussion
As SES speed and acceleration are highly reduced during drift phases (a strong negative correlation coefficient between drift phases and V, r = −0.95, P < 0.0001), the self-noise due to pressure fluctuations around the hydrophone is also reduced (Burgess et al. 1998). This tendency has already been documented in different moving hydrophone situations, like with vertical profilers (Barclay and Buckingham 2013). These drift phases then provide clean experimental conditions for acoustic analysis, with a domination of environment-related acoustic events over SES variables.
The SES depth d was seen to have much minor effects on the measured spectra, with regression coefficients inferior to 0.4 (P < 0.0001) over the different acoustic descriptors. This result can be explained easily. Indeed, the SES explanatory variables impacting the spectra are directly linked to the SES, and so independent from its depth. Also, our main explanatory variable for the environment, the wind speed, is assumed to be a uniformly distributed surface sound source, with an underwater propagation that is at near-vertical angle, as for most geophysical sound sources (wind/rain/drizzle). Refraction by the sound speed profile of the associated acoustic energy is then negligible, and is consequently independent of depth if absorption is neglected. In our data, absorption effects were not significant in comparison to the other explanatory variables of SES speed and wind speed. This property of depth independence has also been studied by Barclay and Buckingham (2013), validating experimentally Cron and Sherman’s (1962) theory that predicts a constant (vertical) directional density function for wind-driven sources.
Our results from the marine soundscape part will be the focus now. To capture a marine soundscape, a full noncorrupted measured spectrum (typically from 10 to 10000 Hz) is needed. According to our correlation analysis, the most corrupted frequency bands are below 1 kHz, with major contributions from V. This ensures us that for low V values, corruption from nonenvironmental sources should be canceled out for these SPL responses. Following this idea, we performed a measure of our marine soundscape using only on the 5th percentile values of the SES speed and acceleration (see Fig. 4), referring to these dives as extreme drifts, when the SES is very slow. These V values range from 0.2 to 0.6 m s−1 and induce variations strictly inferior to 5 dB in the SPL(0.05) (from Fig. 5), which are negligible regarding SPL variations in the same frequency band from classical acoustic sources present in a marine soundscape [e.g., Tsang-Hin-Sun et al. (2015) reported, for the same geographical area, daily variations in the sound pressure level in the order of 20 dB at 10 Hz and 15 dB at 50 Hz, which are associated with tectonic and shipping noise, respectively]. The spectrum shape obtained during these drift phases is quite similar from standard deep-ocean ambient noise spectra. The deep-water Atlantic curve falls off rapidly from its highest level at 50 Hz at a rate of about −8 dB per octave. It is flat from 200 to 400 Hz and then assumes a slope of −4 dB to 1000 Hz (Perrone 1969; Marshall 2005). Our spectrum shape can also be well approximated by the standard Knudsen–Alford–Emling curves, which are straight lines of constant negative slope of 5 dB per octave over the full frequency range (Knudsen et al. 1948). Considering the median levels of our averaged spectrum, in the Kerguelen Islands region, they were the same order of magnitude as those recently reported in the Indian Ocean by Tsang-Hin-Sun et al. (2015). For example, our SPLs ranged generally from 65 to 75 dB ref 1 μPa2 Hz−1 in the frequency band [50–100] Hz, whereas levels across the sites studied in Tsang-Hin-Sun et al. (2015) have been reported to be 75–85 dB ref and 65–70 dB ref 1 μPa2 Hz−1 (with the hydrophone WKER1, moored at 500 m below the sea surface) at 50 and 100 Hz, respectively. These levels are quite coherent with a measure done above the sound fixing and ranging (SOFAR) channel.
Regarding the measurement of wind speed, studies in acoustical meteorology (Vagle et al. 1990; Nystuen and Selsor 1997; Ma and Nystuen 2005; Pensieri et al. 2015) have shown that surface wind speed impacts the linearly ocean ambient noise level in such a way that relatively simple computational approaches (typically a constant threshold set to a one-third octave subband-averaged SPL) are sufficient to predict wind speed from underwater acoustics. Selecting the proper frequency bands is often the critical issue in such studies. It is known that SPL spectra associated with wind result from resonant acoustic radiation from bubbles generated by breaking waves (Medwin and Beaky 1989). The population size distribution of the bubbles defines the shape of the SPL spectra, and this shape of distribution is invariant with wind speed. However, as wind speed increases, the total bubble concentration increases because the fractional area coverage of breaking waves increases. The increase in bubble concentration leads to a concomitant increase in the SPL across all frequencies. This uniform increase as a function of frequency allows the SPL at a single frequency to be used to estimate wind speed. For example, the wind speed algorithm Wind Observations Through Ambient Noise (WOTAN) from Vagle et al. (1990) is based on the SPL at 8 kHz. More recently, Cauchy et al. (2015) applied the WOTAN algorithm to acoustic passive recordings from an underwater glider, with an error of 2 
Another interesting finding of our study is that wind speed estimation can be performed quite independently of SES diving behavior. Indeed, the error estimations between the drift and active swimming phases do not show discrepancies statistically significant, with average errors of 
5. Conclusions
Measuring and interpreting marine soundscape require passive acoustic recordings that are free from nonenvironmental noise. This current study explored the possibility of measuring soundscape with recordings from biologged southern elephant seals (SES). This technological approach offers valuable advantages, such as a high temporal resolution in acoustic measurements (in comparison to satellite-based measures). Also, SES allow the exploration of regions that are inaccessible to other technologies (e.g., underwater gliders). Indeed, the strong east–west current speeds and the rough sea state in the Southern Ocean make the use of regular gliders very complicated, while these harsh environmental conditions are not a problem for SES.
In this study, a multivariate multiple linear regression framework was essentially used to discriminate the acoustic contributions between the processes related to SES and the ocean environment. Our results showed that with passive acoustic recordings from a tagged free-ranging SES, minimal sound corruption from the SES could be obtained during extreme drift phases, allowing for the analysis of marine soundscape. Also, in frequency bands higher than 2.5 kHz, wind speed could be estimated using simple classification approaches (such as linear regression models), independently from the SES biological processes.
Future studies should use a more complete recording dataset, including additional environmental (e.g., rain, current, aquatic seisms), anthropological (e.g., ship traffic), and biological (e.g., whale vocalizations) variables that were not significantly present in our current dataset. This will be needed to fully validate our system as an operational measuring system to measure marine soundscape and to explain its characteristics with a full set of acoustic sources. Also, the ability to associate mobile acoustic data with estimates of surface weather conditions allows for a novel approach to studying air–sea interactions, which will need to be tested at a larger scale, that is, using different types of mobile platforms (e.g., gliders, profiling floats), in different ocean environments.
Acknowledgments
The authors are indebted to several researchers of the CEBC (Chizé, France), in particular Gaetan Richard, for assistance in sample preparation, data collection, processing algorithms, as well as numerous exchanges on SES biology. This study was conducted as part of the Institut Polaire program 109 (ecology of seabirds and marine mammals, P. I. H. Weimerskirch). The authors also would like to thank the Direction Générale de l’Armement (DGA, France) for supporting this work, the Fondation Total for funding the Acousonde as part of the Sea Bio-Sound project, the CNES-TOSCA, and the French Polar Institute for logistical and financial support, as well as all the Kerguelen fieldworkers for collecting data.
APPENDIX A
Influence of Depth on the Acoustic Features
Figure A1 shows the effect of depth on the SPL descriptors. We computed the differences between the SPLs above 100 m and at successive depths down to 800 m, averaged over 200 descending dives of the SES. We made the assumption that during the duration of these dives, the ocean ambient noise does not vary significantly. We can observe that the two lowest-frequency SPLs, 
Effects of depth on the SPL descriptors. The SES speed has also been superimposed onto these curves. Median values have been computed every 100 m in depth d.
Citation: Journal of Atmospheric and Oceanic Technology 34, 1; 10.1175/JTECH-D-16-0124.1
APPENDIX B
Pairwise Correlation and Scattering Plot of Variables
The pairwise correlation between the acoustic, SES, and environmental features was performed using standardized values. Resulting correlation coefficients are provided in Fig. B1 along with scattering plots. Strong correlation relations appear between specific parameters, such as SES speed with SPL(0.05), SPL(0.2), and SPL(0.5) (
Scattering plots and pairwise correlations between acoustic variables and SES and environment variables, using standardized values. Pearson correlation coefficients higher than 0.4, and with a P value below 0.001 (red).
Citation: Journal of Atmospheric and Oceanic Technology 34, 1; 10.1175/JTECH-D-16-0124.1
REFERENCES
Aguzzi, J., and Coauthors, 2011: The new seafloor observatory (OBSEA) for remote and long-term coastal ecosystem monitoring. Sensors, 11, 5850–5872, doi:10.3390/s110605850.
Aoki, K., and Coauthors, 2011: Northern elephant seals adjust gliding and stroking patterns with changes in buoyancy: Validation of at-sea metrics of body density. J. Exp. Biol., 214, 2973–2987, doi:10.1242/jeb.055137.
Bailleul, F., M. Authier, S. Ducatez, F. Roquet, J. B. Charrassin, Y. Cherel, and C. Guinet, 2010: Looking at the unseen: Combining animal bio-logging and stable isotopes to reveal a shift in the ecological niche of a deep-diving predator. Ecography, 33, 709–719, doi:10.1111/j.1600-0587.2009.06034.x.
Barclay, D. R., and M. J. Buckingham, 2013: Depth dependence of wind-driven, broadband ambient noise in the Philippine Sea. J. Acoust. Soc. Amer., 133, 62–71, doi:10.1121/1.4768885.
Baumgartner, M. F., S. M. Van Parijs, F. W. Wenzel, C. J. Tremblay, H. C. Esch, and A. M. Warde, 2008: Low frequency vocalizations attributed to sei whales (Balaenoptera borealis). J. Acoust. Soc. Amer., 124, 1339–1349, doi:10.1121/1.2945155.
Bentamy, A., and D. Croize-Fillon, 2012: Gridded surface wind fields from Metop/ASCAT measurements. Int. J. Remote Sens., 33, 1729–1754, doi:10.1080/01431161.2011.600348.
Bentamy, A., and D. Croize-Fillon, 2015: Daily ASCAT surface wind fields. IFREMER Tech. Rep., 15 pp.
Breiman, L., and J. H. Friedman, 1997: Predicting multivariate responses in multiple linear regression. J. Roy. Stat. Soc., 59B, 3–54, doi:10.1111/1467-9868.00054.
Burgess, W. C., 2000: The bioacoustic probe: A general-purpose acoustic recording tag. J. Acoust. Soc. Amer., 108, 2583–2583, doi:10.1121/1.4743598.
Burgess, W. C., P. Tyack, B. Le Boeuf, and D. P. Costa, 1998: A programmable acoustic recording tag and first results from free-ranging northern elephant seals. Deep-Sea Res. II, 45, 1327–1351, doi:10.1016/S0967-0645(98)00032-0.
Cauchy, P., T. Pierre, L. Mortier, and B. Marie-Noelle, 2015: Passive acoustics embedded on gliders—Weather observation through ambient noise. Proc. Third Underwater Acoustics Conf. and Exhibition (UACE2015), Platanias, Crete, Greece, European Acoustics Association, 565–570.
Chapman, N. R., and A. Price, 2011: Low frequency deep ocean ambient noise trend in the Northeast Pacific Ocean. J. Acoust. Soc. Amer., 129, EL161, doi:10.1121/1.3567084.
Charrassin, J. B., and Coauthors, 2008: Southern Ocean frontal structure and sea-ice formation rates revealed by elephant seals. Proc. Natl. Acad. Sci. USA, 105, 11 634–11 639, doi:10.1073/pnas.0800790105.
Costa, D. P., J. M. Klinck, E. E. Hofmann, M. S. Dinniman, and J. M. Burns, 2008: Upper ocean variability in west Antarctic Peninsula continental shelf waters as measured using instrumented seals. Deep-Sea Res. II, 55, 323–337, doi:10.1016/j.dsr2.2007.11.003.
Cron, B. F., and C. H. Sherman, 1962: Spatial-correlation functions for various noise models. J. Acoust. Soc. Amer., 34, 1732–1736, doi:10.1121/1.1909110.
Dragon, A.-C., A. Bar-Hen, P. Monestiez, and C. Guinet, 2012: Horizontal and vertical movements as predictors of foraging success in a marine predator. Mar. Ecol. Prog. Ser., 447, 243–257, doi:10.3354/meps09498.
Duennebier, F. K., and Coauthors, 2002: HUGO: The Hawaii Undersea Geo-Observatory. IEEE J. Oceanic Eng., 27, 218–227, doi:10.1109/JOE.2002.1002476.
Favali, P., 2013: NEMO-SN1 abyssal cabled observatory in the Western Ionian Sea. IEEE J. Oceanic Eng., 38, 358–374, doi:10.1109/JOE.2012.2224536.
Genin, A., G. Richard, J. Jouma’a, B. Picard, N. El Ksabi, J. Vacquié-Garcia, and C. Guinet, 2015: Characterization of postdive recovery using sound recordings and its relationship to dive duration, exertion and foraging effort of southern elephant seals (Mirounga leonina). Mar. Mammal Sci., 31, 1452–1470, doi:10.1111/mms.12235.
Glowacki, O., G. B. Deane, M. Moskalik, Ph. Blondel, J. Tegowski, and M. Blaszczyk, 2015: Underwater acoustic signatures of glacier calving. Geophys. Res. Lett., 42, 804–812, doi:10.1002/2014GL062859.
Guinet, C., and Coauthors, 2013: Calibration procedures and first dataset of Southern Ocean chlorophyll a profiles collected by elephant seals equipped with a newly developed CTD-fluorescence tags. Earth Syst. Sci. Data, 5, 15–29, doi:10.5194/essd-5-15-2013.
Guinet, C., and Coauthors, 2014: Southern elephant seal foraging success in relation to temperature and light conditions: Insight into prey distribution. Mar. Ecol. Prog. Ser., 499, 285–301, doi:10.3354/meps10660.
Hawkins, R., J. Miksis-Olds, D. L. Bradley, and C. Smith, 2014: Periodicity in ambient noise and variation based on different temporal units of analysis. Proc. Meet. Acoust., 17, 070035, doi:10.1121/1.4772729.
Johnson, M., and P. Tyack, 2003: A digital acoustic recording tag for measuring the response of wild marine mammals to sound. IEEE J. Oceanic Eng., 28, 3–12, doi:10.1109/JOE.2002.808212.
Klinck, H., and Coauthors, 2012: Correction: Near-real-time acoustic monitoring of beaked whales and other cetaceans using a Seaglider. PLoS One, 7, 1–8, doi:10.1371/annotation/57ad0b82-87c4-472d-b90b-b9c6f84947f8.
Knudsen, V., R. Alford, and J. Emling, 1948: Underwater ambient noise. J. Mar. Res., 7, 410–429.
Küsel, E. T., D. K. Mellinger, L. Thomas, T. A. Marques, D. Moretti, and J. Ward, 2011: Cetacean population density estimation from single fixed sensors using passive acoustics. J. Acoust. Soc. Amer., 129, 3610–3622, doi:10.1121/1.3583504.
Ma, B. B., and J. A. Nystuen, 2005: Passive acoustic detection and measurement of rainfall at sea. J. Atmos. Oceanic Technol., 22, 1225–1248, doi:10.1175/JTECH1773.1.
Marshall, S., 2005: Depth dependence of ambient noise. IEEE J. Oceanic Eng., 30, 275–281, doi:10.1109/JOE.2005.850876.
Matsumoto, H., J. H. Haxel, R. P. Dziak, D. Bohnenstiehl, and R. W. Embley, 2011: Mapping the sound field of an erupting submarine volcano using an acoustic glider. J. Acoust. Soc. Amer., 129, EL94, doi:10.1121/1.3547720.
Matsumoto, H., D. R. Bohnenstiehl, J. Tournadre, R. P. Dziak, J. H. Haxel, T.-K. A. Lau, M. J. Fowler, and S. A. Salo, 2014: Antarctic icebergs: A significant natural ocean sound source in the Southern Hemisphere. Geochem. Geophys. Geosyst., 15, 3448–3458, doi:10.1002/2014GC005454.
Medwin, H., and M. M. Beaky, 1989: Bubble sources of the Knudsen sea noise spectra. J. Acoust. Soc. Amer., 86, 1124–1130, doi:10.1121/1.398104.
Miksis-Olds, J. L., D. L. Bradley, and X. M. Niu, 2013: Decadal trends in Indian Ocean ambient sound. J. Acoust. Soc. Amer., 134, 3464–3475, doi:10.1121/1.4821537.
Miller, P. J. O., M. P. Johnson, P. L. Tyack, and E. A. Terray, 2004: Swimming gaits, passive drag and buoyancy of diving sperm whales Physeter macrocephalus. J. Exp. Biol., 207, 1953–1967, doi:10.1242/jeb.00993.
Mitani, Y., R. D. Andrews, K. Sato, A. Kato, Y. Naito, and D. P. Costa, 2010: Three-dimensional resting behaviour of northern elephant seals: Drifting like a falling leaf. Biol. Lett., 6, 163–166, doi:10.1098/rsbl.2009.0719.
Nair, N. R., N. Elizabeth Shani, R. Raju, and S. Satheeshkumar, 2015: Underwater ambient noise variability from satellite data—An Indian Ocean perspective. 2015 IEEE Underwater Technology (UT), 4 pp., doi:10.1109/UT.2015.7108285.
Nieukirk, S. L., K. M. Stafford, D. K. Mellinger, R. P. Dziak, and C. G. Fox, 2004: Low-frequency whale and seismic airgun sounds recorded in the mid-Atlantic Ocean. J. Acoust. Soc. Amer., 115, 1832–1843, doi:10.1121/1.1675816.
Nystuen, J. A., and H. D. Selsor, 1997: Weather classification using passive acoustic drifters. J. Atmos. Oceanic Technol., 14, 656–666, doi:10.1175/1520-0426(1997)014<0656:WCUPAD>2.0.CO;2.
Nystuen, J. A., M. A. Anagnostou, E. M. Anagnostou, and A. Papadopoulos, 2015: Monitoring Greek seas using passive underwater acoustics. J. Atmos. Oceanic Technol., 32, 334–349, doi:10.1175/JTECH-D-13-00264.1.
Pensieri, S., R. Bozzano, M. Anagnostou, E. Anagnostou, R. Bechini, and J. Nystuen, 2013: Monitoring the oceanic environment through passive underwater acoustics. 2013 MTS/IEEE OCEANS—Bergen, 10 pp., doi:10.1109/OCEANS-Bergen.2013.6607995.
Pensieri, S., R. Bozzano, J. A. Nystuen, E. N. Anagnostou, M. N. Anagnostou, and R. Bechini, 2015: Underwater acoustic measurements to estimate wind and rainfall in the Mediterranean Sea. Adv. Meteor., 2015, 612512, doi:10.1155/2015/612512.
Perrone, A. J., 1969: Deep-ocean ambient-noise spectra in the northwest Atlantic. J. Acoust. Soc. Amer., 46, 762, doi:10.1121/1.1911759.
Piggott, C. L., 1964: Ambient sea noise at low frequencies in shallow water of the Scotian Shelf. J. Acoust. Soc. Amer., 36, 2152, doi:10.1121/1.1919337.
Rahmati, M., P. Pandey, and D. Pompili, 2014: Separation and classification of underwater acoustic sources. 2014 Underwater Communications and Networking, Sestri Levante, Italy, CMRE, doi:10.1109/UComms.2014.7017145.
Richard, G., J. Vacquié-Garcia, J. Jouma’a, B. Picard, A. Génin, J. P. Y. Arnould, F. Bailleul, and C. Guinet, 2014: Variation in body condition during the post-moult foraging trip of southern elephant seals and its consequences on diving behaviour. J. Exp. Biol., 217, 2609–2619, doi:10.1242/jeb.088542.
Rogers, E. O., J. G. Genderson, W. S. Smith, G. F. Denny, and P. J. Farely, 2004: Underwater acoustic glider. 2004 IEEE International Geoscience and Remote Sensing Symposium Proceedings, Vol. 3, IEEE, 2241–2244, doi:10.1109/IGARSS.2004.1370808.
Ropert-Coudert, Y., and R. P. Wilson, 2005: Trends and perspectives in animal-attached remote sensing. Front. Ecol. Environ., 3, 437–444, doi:10.1890/1540-9295(2005)003[0437:TAPIAR]2.0.CO;2.
Roquet, F., Y.-H. Park, C. Guinet, F. Bailleul, and J.-B. Charrassin, 2009: Observations of the Fawn Trough Current over the Kerguelen Plateau from instrumented elephant seals. J. Mar. Syst., 78, 377–393, doi:10.1016/j.jmarsys.2008.11.017.
Rudnick, D. L., R. E. Davis, C. C. Eriksen, D. M. Fratantoni, and M. J. Perry, 2004: Underwater gliders for ocean research. Mar. Technol. Soc. J., 34, 73–84, doi:10.4031/002533204787522703.
Sato, K., Y. Mitani, M. F. Cameron, D. B. Siniff, and Y. Naito, 2003: Factors affecting stroking patterns and body angle in diving Weddell seals under natural conditions. J. Exp. Biol., 206, 1461–1470, doi:10.1242/jeb.00265.
Sirovic, A., S. M. Wiggins, and E. M. Oleson, 2013: Ocean noise in the tropical and subtropical Pacific Ocean. J. Acoust. Soc. Amer., 134, 2681–2689, doi:10.1121/1.4820884.
Tennekes, H., and J. L. Lumley, 1972: A First Course in Turbulence. MIT Press, 300 pp.
Tournadre, J., 2014: Anthropogenic pressure on the open ocean: The growth of ship traffic revealed by altimeter data analysis. Geophys. Res. Lett., 41, 7924–7932, doi:10.1002/2014GL061786.
Tsang-Hin-Sun, E., J.-Y. Royer, and E. C. Leroy, 2015: Low-frequency sound level in the southern Indian Ocean. J. Acoust. Soc. Amer., 138, 3439–3446, doi:10.1121/1.4936855.
Urick, R., 1971: Noise of melting icebergs. J. Acoust. Soc. Amer., 50, 337–341, doi:10.1121/1.1912637.
Vacquié-Garcia, J., F. Royer, A.-C. Dragon, M. Viviant, F. Bailleul, and C. Guinet, 2012: Foraging in the darkness of the Southern Ocean: Influence of bioluminescence on a deep diving predator. PLoS One, 8, e43565, doi:10.1371/journal.pone.0043565.
Vagle, S., W. G. Large, and D. M. Farmer, 1990: An evaluation of the WOTAN technique for inferring oceanic wind from underwater sound. J. Atmos. Oceanic Technol., 7, 576–595, doi:10.1175/1520-0426(1990)007<0576:AEOTWT>2.0.CO;2.
Van Dyke, M., 1982: An Album of Fluid Motion. 14th ed. Parabolic Press, 176 pp.
Wagstaff, R., 2005: An ambient noise model for the northeast Pacific Ocean basin. IEEE J. Oceanic Eng., 30, 286–294, doi:10.1109/JOE.2004.836993.
Ward, J., and Coauthors, 2011: Beaked whale (Mesoplodaon densirostris) passive acoustic detection in increasing ambient noise. J. Acoust. Soc. Amer., 129, 662–669, doi:10.1121/1.3531844.
Wenz, G. M., 1962: Acoustic ambient noise in the ocean: Spectra and sources. J. Acoust. Soc. Amer., 34, 1936–1956, doi:10.1121/1.1909155.
Zhang, Z., J. Neubauer, and D. A. Berry, 2006: Aerodynamically and acoustically driven modes of vibration in a physical model of the vocal folds. J. Acoust. Soc. Amer., 120, 2841–2849, doi:10.1121/1.2354025.
Acousondes are miniature, self-contained, autonomous acoustic recorder designed for underwater applications.
Developed by Yves le Bras (CEBC-CNRS-UMR 7372 Université de La Rochelle, Chizé, France) and available at https://github.com/SESman/rbl.
Available at