Acoustic twilight: A year‐long seafloor monitoring unveils phenological patterns in the abyssal soundscape

Despite the perpetual darkness of the deep sea, contrasting the sunlit epipelagic waters, many deep‐sea organisms exhibit rhythmic activities. To discern environmental cues that may serve as entrainment signals for deep‐sea organisms, this study investigated the soundscape of the abyssal plain south of Minamitorishima Island. Our analysis revealed clear diel and seasonal patterns, primarily driven by evening fish choruses and marine mammal vocalizations. These evening choruses, discernible above the background noise, likely serve as a circadian time cue for organisms capable of perceiving them within the aphotic depths. In addition, the frequent detection of whistles and echolocation clicks suggests this region functions as a foraging ground for marine mammals. These acoustic cues might guide organisms with auditory capabilities toward habitats rich in sinking food debris and whale falls. By elucidating the ecological processes shaping abyssal soundscape dynamics, these findings open new directions for further exploration in deep‐sea chronobiology.

The deep sea, due to its prevailing ecological conditions of perpetual darkness, low temperatures, high pressures, and sporadic food availability, has long been viewed as an environment lacking discernible rhythms (Grassle 1989;Beale et al. 2016).Contrary to this perspective, recent advances in deep-sea exploration have revealed rhythmic activities in deep-sea organisms inhabiting light-deprived depths.Examples include the diel vertical migration of planktonic organisms (van Haren and Compton 2013), the diel displacement of nektobenthic fishes (Doya et al. 2014), and diverse behavioral rhythms underpinning benthopelagic coupling (Aguzzi et al. 2011(Aguzzi et al. , 2015;;Modica et al. 2014).Although these rhythmic patterns suggest an underlying molecular clock mechanism (Häfker et al. 2017;Mat et al. 2020), the external environmental cues serving as entrainment signals for deep-sea organisms remain largely elusive.
Sound, a ubiquitous underwater cue that propagates effectively over long distances, serves as a vital sensory mechanism for a wide array of marine species (Johnson et al. 2004;Montgomery et al. 2006;Popper and Hawkins 2019;Solé et al. 2023).Coral larvae, mollusks, crustaceans, and fishes leverage this sensory modality for auditory navigation, utilizing it for orientation and settlement (Simpson et al. 2005;Vermeij et al. 2010;Eggleston et al. 2016).Although primarily studied in shallow-water species, auditory navigation is also hypothesized to be used by deep-sea larvae, possibly for dispersion between hydrothermal vents (Crone et al. 2006;Lin et al. 2019).Furthermore, deepdiving marine mammals such as sperm whales and beaked whales rely on sound for foraging, employing echolocation to locate prey within the mesopelagic and bathypelagic zones (Watwood et al. 2006;Visser et al. 2022).
Despite the growing recognition of the ecological importance of sound, the sonic environment in deep-sea ecosystems remains largely uncharted (Bolgan and Parmentier 2020;Havlik et al. 2022).To date, only a handful of studies have investigated soundscapes in the bathyal zone over several months (Melc on et al. 2012;Bohnenstiehl et al. 2013;Wiggins et al. 2016;Martin et al. 2019;Warren et al. 2021;Weiss et al. 2021), and even fewer have done so in the abyssal and hadal zones (Dziak et al. 2017;Chen et al. 2021).These pioneering studies reveal a deep-sea soundscape teeming with biological, environmental, and anthropogenic sounds.For example, Dziak et al. (2017) identified a noisy hadal environment at Challenger Deep-the Earth's deepest known locationresonating with sounds produced from storm-induced waves, marine mammals, shipping activities, and geophysical airgun surveys.Another case in point is the abyssal soundscape off Minamitorishima Island, which transforms from daytime quietude to a vibrant biological chorus after sunset (Chen et al. 2021).Furthermore, the soundscape recorded at mid-ocean ridges showed tidal modulations, suggesting pressure-induced variations at hydrothermal vents (Crone et al. 2006).These observations highlight the potential of deep-sea soundscapes as an environmental cue, reflecting ecological processes that span the entire depth range from the sea surface through the light-penetrated photic zone, all the way to the light-deprived aphotic zone.
Responding to the challenge of deep-sea exploration, recent advancements in underwater technology, which incorporate hydrophones or autonomous recorders into both mobile and stationary observation platforms, have broadened the scope and flexibility of monitoring (Bolgan and Parmentier 2020;Fregosi et al. 2020;Rountree et al. 2020).This progress marks a substantial stride toward a more comprehensive understanding of deep-sea soundscapes.Here, we present an analysis of the abyssal plain soundscape south of Minamitorishima Island, using a year's worth of audio data recorded by a deep-sea monitoring lander system.Our aim is to determine whether the abyssal plain soundscape exhibits discernible diurnal or seasonal patterns and identify the contributing sound sources to this phenology.By addressing these questions, we seek to expand our understanding of the ecological processes driving soundscape dynamics and discuss the potential role of underwater sound as an entrainment signal within deep-sea ecosystems.

Study area and soundscape monitoring
The deep-sea soundscape recordings for this study were obtained from a location east of the Takuyo-Daigo Seamount (22 59.9 0 N, 154 24.5 0 E, depth 5552 m), approximately 150 km south of Minamitorishima Island (Fig. 1).This remotely situated area is characterized by minimal shipping traffic due to its relative uninhabitability.The surrounding abyssal plain is rich in manganese nodules (Machida et al. 2016), and a substantial reservoir of rare-earth element-enriched mud is found beneath the seafloor (Takaya et al. 2018).
The recording system utilized consists of a 6000 m rated icListen HF hydrophone (model SB60L-ETH, Ocean Sonics) and a Gordon Smart Recorder (Ocean Sonics).Both were integrated into an Edokko Mark I Type 365 (Okamoto Glass)-a free-fall standalone system for autonomous monitoring at depths of up to 8000 m for 1 yr (Kawagucci et al. 2020;Onishi et al. 2023).The icListen HF hydrophone was mounted on top of the Edokko Mark I and connected to two pressure-resistant glass spheres that housed the Gordon Smart Recorder and battery packs (see Onishi et al. 2023 for details).The hydrophone has an effective sensitivity of À170 dB re 1 V μPa À1 , with a frequency response spanning from 10 Hz to 200 kHz (AE6 dB bandwidth).No amplification or high-pass filtering was used.Uncompressed WAV files were made using a sampling rate of 512 kHz and a duty cycle that included a 2-min recording session every 4 h (12:00, 16:00, 20:00, 00:00, 04:00, 08:00 h, all time stamps are presented in UTC + 10:00).The Edokko Mark I was deployed on 13 March 2020, during the R/V KAIREI cruise KR20-E01C and was retrieved on 09 April 2021, during the R/V KAIREI cruise KR21-04C.

Soundscape analysis
The variability within the deep-sea soundscape was analyzed using a twofold approach.The first part employed a long-term spectrogram (LTS) to investigate biological choruses exhibiting pronounced diurnal and seasonal patterns.The second part sought to identify high-intensity transient sounds using short-time spectrograms.All acoustic analyses were conducted in the Python coding environment using the soundscape_IR toolbox (Sun et al. 2022).

Biological choruses
Biological choruses occur when numerous sounds from a large aggregation of soniferous animals form a continuous signal (Mooney et al. 2020).To detect such choruses, the technique of LTS was applied.LTS is an effective visualization tool for spectral and temporal variations in long-duration audio recordings, and its efficacy in delineating the heterogeneity of deep-sea soundscapes has been discussed (Chen et al. 2021).Following the methodologies of Chen et al. (2021), we generated a median power spectrum (10 Hz frequency resolution) for each 1-min fragment.Given the recording duty cycle, this yielded 12 median power spectra per day.These spectra were then chronologically assembled into an LTS.
The extraction of median power spectra ensures the spectral representation of continuous sound sources, including environmental sounds, biological choruses, and shipping noise (Chen et al. 2021).To enhance signal-to-noise ratios (SNRs) of biological choruses, we measured n 5 (f ): the 5 th percentile at each frequency bin f to represent the background noise level.This particular threshold of the 5 th percentile was selected because it was found to be below the level at which biological choruses are identifiable (Supporting Information Fig. S1).All log-scaled power spectra on the LTS were subsequently normalized by subtracting the n 5 (f ), with negative values set to 0.
Finally, we processed the LTS using periodicity-coded nonnegative matrix factorization (PC-NMF, Lin et al. 2017).PC-NMF is a machine learning tool designed specifically for soundscape analysis, it learns a set of spectral features and exploits their periodicity patterns to identify spectral features related to biological choruses in an unsupervised manner.After feature learning, the model can separate biological choruses from other sound sources, thereby enabling accurate prediction of their temporal patterns (Supporting Information Fig. S2).

Transient sounds
Typically, soniferous animals produce sounds that do not coalesce into a chorus, instead remaining as separate, individual signals.To capture these transient sounds, we employed the method of semi-supervised source separation proposed by Sun et al. (2022).We initially selected six recording files, including ambient sounds and anthropogenic noises such as impulsive machinery and echosounder signals.For these chosen files, we generated log-scaled spectrograms (20 ms time resolution, 250 Hz frequency resolution, 0-40 kHz frequency range) as training data for a nonnegative matrix factorization (NMF) model.Manual inspection confirmed congruence between the spectral features learned by the NMF model and the signals in the training data (Supporting Information Fig. S3).
Subsequently, the NMF model was employed to separate the pre-trained noise sources from all recordings.During the analysis of each 1-min fragment, the spectrogram was normalized by subtracting the n 85 (f ), with negative values set to 0. The selection of the 85 th percentile was based on empirical assessment, involving manual examination of the percentage of time bins occupied by transient sounds in each frequency bin (generally less than 10%, as illustrated in Supporting Information Fig. S4).In the semi-supervised separation procedure, an additional set of 10 spectral features were added into the NMF model to distinguish sounds not related to the pre-trained noise sources.This was done by fixing the parameters of the pre-trained spectral features while concurrently updating the new spectral features through an iterative process.This method ensured the effective extraction of high-intensity transient sounds with spectral features differing from those learned in the training data (Supporting Information Fig. S4).
The diversity encompassed within high-intensity transient sounds was investigated using uniform manifold approximation and projection (UMAP)-a technique widely used to project highdimensional data into a low-dimensional space (McInnes et al. 2020).By projecting all spectral features learned from semisupervised source separation to one-dimensional UMAP coordinates, we obtained a comprehensive view of the gradient structure encapsulating spectral heterogeneity among transient sounds (Sun et al. 2022).These transient sounds were then categorized based on their spectral characteristics and respective UMAP coordinates.

Results
This study analyzed 4260 min of audio recordings from 13 March 2020 to 06 March 2021.The analysis, using the median-based LTS method, showed that the abyssal plain soundscape was primarily shaped by sounds generated from biological activities, shipping, and storms, with a focus on frequencies below 10 kHz (Fig. 2a,b).
Our analysis revealed a clear diel variation in the abyssal plain soundscape, primarily driven by recordings obtained at 20:00 and 00:00 (Fig. 2a).During these periods, we observed spectral peaks resembling previously reported fish choruses (e.g., McCauley and Cato 2016;Lin et al. 2021).Further examination identified two distinct types of evening choruses: Type I, detected in the 1.45 to 2.1 kHz frequency range (Fig. 2c) and confined solely to the 20:00 recordings (Fig. 2d); and Type II, detectable in the frequency range of 0.45-0.65 kHz with relatively low SNRs (Fig. 2e), occurring within a slightly wider time window from 20:00 to 00:00 (Fig. 2f).These choruses exhibited different seasonal patterns, with Type I predominantly detected between December and June (Fig. 2d), and Type II between May and September (Fig. 2f).However, the separation of Type II chorus using PC-NMF was ineffective from 15 June 2020 to 28 July 2020 due to the interference of high-intensity environmental noise (Supporting Information Fig. S2).
The implementation of semi-supervised source separation facilitated the assessment of transient sound diversity, which was represented by 1-dimensional UMAP coordinates (Fig. 3a).Tonal sounds, comprising 14.3% of all signals, were found at UMAP coordinates below À11 (Fig. 3b).These tonal sounds were primarily detected at frequencies between 5 and 12 kHz, and their frequency modulated structure suggests an association with delphinids (Fig. 3c).The majority of signals, distributed at UMAP coordinates above À11, were broadband sounds (Fig. 3b).Upon visual inspection of spectrograms, these broadband sounds exhibited distinctive multi-pulses structures with varying inter-pulse intervals-an acoustic feature of odontocete echolocation clicks (Fig. 3c-e).Sorting the UMAP coordinates further revealed a multimodal distribution of peak frequencies within the clicks.Based on the UMAP distribution modes and corresponding spectral characteristics, we manually identified three distinct groups: low-frequency clicks with peak frequencies below 5.5 kHz (30.7%), midfrequency clicks between 5.5 and 20 kHz (51.9%), and highfrequency clicks above 20 kHz (3.1%).
Our data showed higher detection rates for mid-and highfrequency clicks during nighttime recordings compared to daytime (Fig. 3b).In addition to this diurnal variation, distinct

Lin and Kawagucci
Acoustic twilight in the abyss seasonal patterns were observed for the four groups of transient sounds (Fig. 4a).Whistles were primarily detected from March 2020 to September 2020, while the detection of lowand mid-frequency clicks showed several seasonal peaks, including May 2020 to August 2020 and December 2020 to January 2021.Furthermore, clicks occurring at different UMAP coordinate ranges displayed slight variations in their occurrence patterns (Fig. 4b).

Discussion
Our findings unveil the existence of diel and seasonal variations within the abyssal plain soundscape off Minamitorishima Island, adding to the growing body of evidence for rhythmic phenology in the deep sea.This soundscape phenological pattern, characterized by biological choruses that reflect atmospheric twilight, is distinct from the 12.4-and 24.8-h tidal rhythms observed in deep-sea hydrothermal ecosystems (Cuvelier et al. 2017;Mat et al. 2020).Therefore, it highlights the significance of the evening choruses as a unique 24-h circadian time cue in the aphotic zone.
The inability to isolate individual signals from the evening choruses suggests that these choruses originate from soniferous animals located at considerable distances from our recorder, rather than from those in close proximity.The Type I chorus identified in this study parallels the chorus described in Chen et al. ( 2021), whose study site was approximately 65 km away from our recording location, suggesting that these choruses likely represent a bio-oceanographic feature that extends over large spatial scales.Comparable observations of evening choruses in the deep sea have also been reported globally, including in the western Pacific Ocean (Lin et al. 2021) and the eastern Indian Ocean (Cato 1978).Furthermore, these evening choruses have been detected across a wide range of depths, ranging from 277 m (Lin et al. 2021), and430-490 m (McCauley andCato 2016), to 1000 m (Cato 1978), and even down to the deepest recorded depth of 5500 m (Chen et al. 2021).In the Perth Canyon, western Australia, an evening chorus detected in the frequency range of 1.89 to 2.67 kHz was inferred to be associated with mesopelagic fishes of the family Myctophidae (McCauley and Cato 2016).Myctophidae fishes are known to perform diel vertical migrations between mesopelagic and epipelagic depths and play a major role in deep scattering layers (Catul et al. 2011).Therefore, it is plausible that mid-water fishes capable of sensing diel light changes may contribute to these evening choruses, emphasizing the role of underwater sound as a signaling vector in the interaction between the shallow and deep-water ecosystems.
Many deep-sea fish species possess specialized soundgenerating anatomical structures and adaptations that augment hearing (Buran et al. 2005;Ladich and Schulz-Mirbach 2016).As a result, auditory perception has been acknowledged as an important sensory function in the life of deep-sea fishes (Bolgan and Parmentier 2020).Even though there remains a lack of studies that specifically delineate the hearing and vocal behavior of deep-sea fishes, our analysis provides evidence that evening choruses can be distinguished above the ambient noise floor within the surveyed abyssal plain.Although not all sounds recorded may be perceptible to all marine organisms, our finding raises the intriguing possibility that evening choruses could act as potential entrainment signals for deep-sea organisms capable of auditory perception.
The occurrence of numerous whistles and echolocation clicks in our recordings suggests that the study area functions as a foraging ground for odontocetes.The acoustic characteristics of low-and high-frequency clicks resemble those produced by sperm whales and beaked whales, respectively (Stanistreet et al. 2022).These species are renowned for their deep-diving foraging habits.Their foraging activities may enhance the likelihood of sinking food debris, thereby contributing to the downward transport of organic matter and playing a vital source of energy and habitat for scavengers in the deep sea (Roman et al. 2014).Furthermore, the status as a marine mammal hotspot implies a higher likelihood of whale falls, which create complex localized ecosystems supporting deep-sea organisms, including chemosynthetic communities that sustain primary production in the absence of sunlight (Smith et al. 2015).Therefore, the presence of marine mammals acts as a significant driver of deep-sea ecosystems, and their sounds may contribute to the spatiotemporal heterogeneity of soundscapes, potentially guiding organisms with auditory capabilities toward habitats rich in food resources.Nevertheless, it is crucial to recognize that due to the absence of a pressure-sensitive gas chamber, many deep-sea fishes and invertebrates predominantly respond to particle motion over sound pressure (Popper and Hawkins 2018).Future research should examine how deep-sea organisms detect the nuances found in soundscapes and vibroscapes.
While this study offers a unique investigation into the year-round variation of the abyssal plain soundscape, there are several limitations that should be considered.First, the use of intermittent recordings may limit the completeness of the soundscape phenology (Mooney et al. 2020).Future studies necessitate a higher duty cycle and broader geographical coverage will enhance our understanding of soundscape dynamics in diverse deep-sea habitats.Second, the spectrograms used for source separation may not effectively capture all transient sounds, such as the low-frequency calls of baleen whales.Moreover, the separation of high-intensity transient sounds could inadvertently overshadow sounds emitted by deep-sea animals, which generally produce low-amplitude sounds (Rountree et al. 2012).Developing methodologies adept at discerning these low-amplitude sounds will be crucial for a comprehensive assessment of deep-sea soundscape.Lastly, while this study focuses on soundscapes, it is important to acknowledge that deep-sea organisms may rely on other rhythmic environmental cues, such as tidal cycles, currents, and prey migration patterns (Wagner et al. 2007;Modica et al. 2014;Aguzzi et al. 2015;Mat et al. 2020).Incorporating behavioral and physiological experiments would provide a more holistic understanding of how deep-sea organisms respond to auditory and other environmental cues.Such research would contribute to the emerging field of deep-sea chronobiology (Häfker et al. 2023) and inform management strategies by predicting consequences of changes in deep-sea soundscapes.
Despite their potential significance, deep-sea soundscapes face a drastic alteration due to impending mining activities (Miller et al. 2018;Christiansen et al. 2020).Projected noise levels from such operations are expected to exceed existing ambient noise levels, which could extensively alter soundscapes across a vast region and diverse habitats (Chen et al. 2021;Williams et al. 2022).There is also a high risk that noise pollution from mining operations and shipping activities will mask biological sounds, making these auditory cues imperceptible to deep-sea organisms (Lin et al. 2019).This imminent threat underscores the urgent need for more extensive research in this field, particularly in habitats expected to be most affected by deep-sea mining.Advancing our understanding of the soundscape and the potential impact of anthropogenic activities will be critical in safeguarding these largely unexplored and unique deep-sea ecosystems.

Fig. 1 .
Fig. 1.Map showing the recording site located south of Minamitorishima Island and the topography of the western North Pacific.The study site is situated in the world's largest abyssal plains.

Fig. 2 .
Fig. 2. Visualization of the abyssal plain soundscape off Minamitorishima Island.Two specific periods of recordings are depicted by using the medianbased LTS: (a) from 21 May 2020 to June 2, 2020, and (b) from 24 September 2020 to 06 October 2020, which encompasses the period when Typhoon Kujira was in close proximity to Minamitorishima Island (27-28 September 2020).Example short-time spectrograms for (c) the Type I chorus and (e) the Type II chorus, alongside their diel and seasonal variations (d, f).Each point displayed corresponds to the SNR obtained from each 1-min fragment.The lines and shaded areas indicate the smoothed time series data and the corresponding 95% confidence intervals, respectively.

Fig. 3 .
Fig. 3. Identification and categorization of biological transient sounds via semi-supervised source separation.(a) The spectral diversity of transient sounds, with each bin representing the average signal strength in the frequency domain (y-axis) for all signals falling within a specific range of UMAP coordinates (x-axis).(b) The histogram of UMAP coordinates showing the divergent distribution of whistles (W), low-frequency clicks (LF), mid-frequency clicks (MF), and high-frequency clicks (HF).Colors delineate the quantity of signals detected during daytime hours (from 08:00 to 16:00) compared to nighttime hours (from 20:00 to 04:00).(c-f) Example short-time spectrograms of the four groups of transient sounds.

Fig. 4 .
Fig. 4. Temporal dynamics of the diversity in biological transient sounds.(a) The seasonal patterns of whistles (W), low-frequency clicks (LF), midfrequency clicks (MF), and high-frequency (HF) clicks.The lines and shaded areas indicate the smoothed time series data and the corresponding 95% confidence intervals, respectively.(b) The seasonal variations of all transient sounds across UMAP coordinates (y-axis), with colors signifying the detection probability within each 3-d interval.