Passive acoustic data were collected using High-frequency Acoustic Recording Packages (HARPs) (Wiggins and Hildebrand, 2007). Each deployment included a single moored instrument that was located as close as possible to a target location approximately 5 km off the coast of Kona, Hawai’i (Figure 1), with an average depth of 650 m. Some small variations in the exact drop location (10s of m), combined with the steep slope of the island at this site, resulted in some variation of depths (range 460-720 m). The deployments spanned the 2007–2018 time period, with intermittent gaps between deployments due to servicing schedules and/or limits of battery life and data capacity (Table 1).

Each instrument included a hydrophone suspended 10–20 m above the sea floor with two stages: a low-frequency stage (10–2,000 Hz or 10–25,000 Hz, Table 1) and a high-frequency stage (2–100 kHz or 25–100 kHz) (Wiggins and Hildebrand, 2007). The sensors were connected to custom-built preamplifier boards with bandpass filters. Each hydrophone was calibrated in the laboratory to provide a quantitative analysis of the received sound field. Representative data loggers and hydrophones were also calibrated at the Navy’s Transducer Evaluation Center facility to verify the laboratory calibrations. Additionally, these calibrations were cross-checked with expected amplitudes given local wind conditions (Hildebrand et al., 2021). Instruments recorded at either a 200 or a 320 kHz sampling rate with 16-bit quantization (Table 1). These sampling frequencies were selected with the goal of spanning the hearing and sound production range of all cetacean species.

Different duty cycle schedules were used to maximize recording duration, with each deployment either having a single, consistent duty-cycle or recording continuously. Within duty-cycled deployments, data were recorded for 5 min, followed by an “off” period ranging from 1 to 20 min. Data gaps, both those between deployments and the “off” periods during duty cycles, forced adjustments for recording effort and prevented us from using standard time series analysis methods that are based on continuous measurements across the full duration of the time series.

Acoustic data were processed in multiple ways to meet the requirements of the detection and classification tools that were selected for the various signal types. All signal processing was performed using the MATLAB-based (Mathworks, Natick, MA, United States) custom software program Triton (Wiggins and Hildebrand, 2007¹) and other MATLAB custom routines. The appropriate calibration information for each hydrophone and pre-amplifier was applied during analysis of all data products. A suite of soundscape metrics of various bandwidths was calculated, spanning frequencies below 100 Hz to above 50 kHz.

The entire data set can be viewed as a set of annual long-term spectral averages of each calendar year (Figure 2). This view demonstrates the remarkable extent of the data over time, and also illustrates gaps or changes in monitoring effort (e.g., limited coverage in March throughout all years). The long term and seasonal trends were subtle, with relatively constant soundscape composition across all years and most seasons. The soundscape at lower frequencies (<1,000 Hz) did vary seasonally (Figure 3), with higher levels across all frequencies during winter months (January/February/March). Spring, summer and fall seasons had similar energy across all frequencies in the 10–1,000 Hz range.

Examination of the seasonal spectra (Figure 3) revealed an increase in energy in the winter months (blue spectrum), specifically in the 100-500 Hz band, which corresponds to the presence of humpback whale song. The fall, winter, and spring (October-April) spectra included a peak around 20 Hz, which indicates the presence of fin whale calls (Širović et al., 2013). Additional energy in the low-frequency spectra may be attributable to various fish species, many of which are known to generate a wide variety of sounds. Signal-specific detectors were not implemented for baleen whales other than humpback whales (see below) or for fish sounds, so further analysis of the contribution of these signal types is not included in this study.

The duration of the no-vessel periods were approximately 38 h during Iselle and 33 h during Darby. During each of the 2-week periods around each hurricane there was only a slight correlation between the wind speed and the best-correlated soundscape metric (the 1-Hz wide band at 900 Hz, rho = 0.2, p < 0.0001 for each storm). Wideband spectral plots revealed changes in the overall energy levels during the storms (Figure 10). First, both no-vessel periods had lower sound energy during daytime hours between ∼150 and 800 Hz (Figures 10A,B). These are the frequencies with the majority of vessel energy at this site, as shown in the example spectrum and LTSA (Figures 10E,F), which contain a representative hour with typical vessel passages. Second, the storms passed the HARP during different periods of the day, with Iselle present during the daytime (Figure 10A) and Darby present primarily during the nighttime (Figure 10D). During Iselle the storm produced higher energy in frequencies above ∼200 Hz (Figure 10A, red line in red shaded box), but this energy was still not as high as when vessels were present the week before (Figure 10A, black line). In contrast, the energy was elevated at all frequencies above 30 Hz (Figure 10D, red line in shaded box) during the passage of Darby.

Motivated by the lack of detailed, wide-band, long-term analysis of marine soundscapes, we set out to assess trends at one nearshore, deep-water site over more than a decade. Here we provide an example of the soundscape in a tropical, deep-water, island-associated habitat, which we hope will be compared to underwater soundscapes from other locations. We looked for cyclical patterns on time scales as short as a day and as long as a year, and documented the sound sources that dominate this soundscape on diel, seasonal and inter-annual time scales at various frequencies, as well as signals that do not dominate the soundscape but contribute to the acoustic environment nonetheless.

The long duration, wide bandwidth, and recording depth of the data in the current soundscape analysis allow us to consider all sound sources within the hearing range of the animals found in this acoustic habitat. The lowest frequency sounds (e.g., 20 Hz) are likely heard by large baleen whales such as blue and fin whales (Erbe, 2002) and may also be sensed by many fish or invertebrates. The highest recorded frequencies (160 kHz) capture signals like the echolocation clicks of Kogia spp. (with energy above 120 kHz) that are above the hearing capabilities of many, but not all, marine mammals (e.g., Szymanski et al., 1999; Li et al., 2011).

In addition to the duration and frequency range of the data set, we were also presented with challenges in how to efficiently analyze the data, considering the total size of the data set and the complexity of the signals present. While a general spectral analysis is useful, and reveals some of the important signals in the soundscape, such as the seasonal presence of humpback whales (Figure 7) or the incessant pinging of echosounders at 50 kHz (Figure 6), much of the detail and nuance is lost in this wide view. To explore some of the smaller-scale details we employed a large suite of analysis tools, each with its own challenges and limitations, and each requiring different levels of human input and expertise. Computer-based methods were essential to analyzing the full data set in a reasonable time frame. A similarly multi-faceted analysis scheme will likely be required for most large-scale, underwater soundscape analysis projects going forward.

The long-term perspective, as shown in the annual long-term spectral averages (Figure 2), allows us to draw some overall conclusions about the soundscape at this location over more than 10 years. Notably absent are strong seasonal or long-term changes, with sound sources and levels being relatively stable and consistent over seasons and years. This is as expected, given the tropical latitude and water depth, which combine to mediate most of the seasonal impacts seen in soundscapes at higher latitudes and/or in shallow water environments (e.g., Haver et al., 2017, 2019). There is a slight seasonal pattern, seen most clearly in the spectra in Figure 3, with increased energy across all frequencies in the winter, likely due to seasonal increases in wind speed and storm-generated noise, as well as localized peaks in the frequencies that correlate with the seasonal presence of humpback and other baleen whale calls. Across all frequencies the wintertime sound energy increase is approximately 2-3 dB. In the long-term view we also notice the consistent presence of energy from vessels (Figure 4), shown as higher energy in lower frequencies (∼150-20 kHz) across all years. With such chronic noise this is clearly not a pristine environment, despite being comparatively remote from continents.

Previous research by Širović et al. (2013) using a subset of these data (∼91 days from fall/winter of 2009, 2010 and 2011, 10-1,000 Hz frequency band) provided some initial findings, upon which we expanded. We found similar seasonal patterns, particularly the winter presence of humpback and fin whales. We also confirmed their initial findings of increased energy during daytime hours. We can attribute that increase to vessel presence based on our vessel and echosounder detections. Širović et al. (2013) identified a daily peak in energy at 500 Hz around 20:00 local time, which is also revealed in our data in the nighttime spectra in Figures 10C,D and as a pulse of energy seen just after sunset in the LTSAs for each hurricane passage (Figure 9). The exact source of this signal remains unknown, but its daily cycle and diffuse energy suggest it may be related to a fish chorus or other daily behavior, such as a diel vertical migration.

The nearshore waters of Hawai’i are much-used for many human activities, most of which generate underwater noise. The diurnal presence of vessels and echosounder pings (Figures 4, 6) was expected, given the known diurnal patterns of humans in the area. Such diel trends have been documented in this soundscape at lower frequencies (∼500 Hz) by Širović et al. (2013). However, the extent and persistence of the acoustic presence was remarkable. The chronic daytime vessel noise is likely coming from both smaller sport-fishing as well as commercial fishing vessels. The typical schedule of these vessels is to depart a nearby harbor each morning around sunrise, travel to various established fishing locations, including a fish aggregating device that is located in the same area as the HARP, and then return to shore in the evening.

Despite being over deep water, sounds from these vessels, particularly propeller cavitation and engine noise, introduce acoustic energy into the local environment as revealed by the constant, lower frequency (< 5,000 Hz) energy in our data (Figures 2, 4). Although the duration of each vessel passage is relatively short (∼15-30 min), the number of encounters per day (∼1-5 per hour during daylight hours, for an average of 14 min of vessel presence per hour during daytime hours, reveals the considerable impact this sound source may have on the area across a wide frequency band (10 Hz-50 kHz), similar to what was found by Putland et al. (2017). Also, given the broadband nature of these signals, they are likely to cause significant masking for many marine animals, including protected marine mammal species, whose acoustic signals and hearing ranges fall well within this frequency band (e.g., humpback, pilot, killer and sperm whales, etc.).

We found that the frequency spectra produced by most vessels in our data was different from what is typically described for underwater commercial vessels. Unlike large ships that have considerable energy below 100 Hz (e.g., McKenna et al., 2012b and McKenna et al., 2013), the energy in the vast majority of vessel passages near the Hawai’i HARP peaked between 200 and 400 Hz, and dropped off below 150 Hz (Figure 10E,F). This may be a result of propagation, or due to the majority of vessels being smaller in size, or a combination thereof. This difference in sound signature may help explain a lack of correlation found by Širović et al. (2013) who found no relationship between the energy in a 40 Hz band at this site and the presence of vessels detected using the Automated Identification System (AIS). If most of the vessels passing this site are smaller it is likely that they do not transmit AIS signals. Therefore they would not be included in any AIS-based monitoring, as was found by Magnier and Gervaise (2020), who determined that less than 4% of vessels near their site in the Pelagos marine sanctuary in the Mediterranean were using AIS systems. Similarly, Cholewiak et al. (2018) determined that non-AIS vessels contributed significantly to the noise in the Stellwagen National Marine Sanctuary, reducing the communication space of baleen whales by approximately 30%. Similar conditions may be impacting the cetaceans off Hawai’i, and should be taken into consideration when assessing the environmental impacts of any noise-generating human activities.

In contrast to the broadband signals from vessel engines and propeller cavitation, echosounder pings are very narrow band, with each ping generally being only a few Hertz in bandwidth. However, despite the limited frequency band, the signals have considerable energy, with a median energy per ping being 130 dB re 1 μPa at this site. Echosounder pings have not been reported in other soundscape studies, perhaps because they tend to be above the frequency range of many passive acoustic monitoring systems. Yet, the frequencies of these pings and the side-lobes of higher frequency pings are within the hearing range of many marine mammals, particularly odontocetes (Deng et al., 2014). The biological impact of the near-constant daytime pinging at this location should not be underestimated and may be impacting marine mammals in the area. It is fortunate that there appears to be some temporal acoustic niche separation, with anthropogenic noise not overlapping with the majority of echolocation clicks from delphinids, because of their nocturnal behavior (Figures 4, 5, 8). However, the possibility of significant physical or acoustic interactions during dawn and dusk should be explored further, and anthropogenic noise may still impact their daytime resting even when they are less acoustically active. Additionally, several species of odontocetes are active and detected at this site during all hours of day and night (e.g., Baumann-Pickering et al., 2014; Merkens et al., 2019), increasing the chances of acoustic overlap and masking for those species.

We observed a seasonal trend in echosounder pings, with an approximate doubling of detections during summer months (June–September) (Figure 6). This likely reflects the seasonality of tourism as well as chances of good weather, both of which are higher during the summer. The corresponding lack of a seasonal pattern for vessel detections may be the result of echosounders pinging while vessels are moving slowly or drifting and therefore not generating detectable levels of sound energy. This disconnect may also be due to the detector function, where the threshold for “presence” shifts dynamically based on the overall energy in each individual 2-h window. Windows with almost no vessel noise and a single higher-energy passage (e.g., during a month with less overall vessel activity) might result in a similar detection rate as windows with a high baseline of vessel noise (e.g., during peak summer months) and a single higher-energy passage. Despite this potential limitation in detector performance there are currently no other options available for automated detection of vessels, and overall performance was found to be satisfactory by validating a subset of manual vessel detections. Further analysis to assess the impacts of this dynamic detection method and improvement of this detector are beyond the scope of the current work.

We considered whether soundscape metrics could be used as a proxy for anthropogenic signals, and found that this depends greatly on the signal type and the detection method. The automated vessel detections (Figure 4), may work well at a location where the vessel passages are distinct from each other and when analyzing a continuous recording, which could result in a high correlation with the soundscape metrics (e.g., energy in the 4 kHz octave). However, when the data are duty cycled, or there is a constant, high level of vessel energy, the correlation suffers. This varying performance reveals that the soundscape metrics could possibly be used as a proxy for vessel passages, but more importantly, the impact of recording parameters and ambient noise conditions on detector performance should be considered. In contrast, detection of the echosounders was much more consistent across time and with low variability in ambient noise conditions at higher frequencies, resulting in a strong correlation between the soundscape metric (50-50.1 kHz band) and the automatic detections. In this case, the sound energy measurements can be used reliably for monitoring the presence of echosounder pings over time.

While anthropogenic sounds were shown to be prevalent throughout the time series, biological sounds are also significant contributors to the soundscape at this location. The dominant driver of seasonal change in frequencies below 1,000 Hz was the wintertime songs of humpback whales (Figures 3, 7). These results confirm the earlier findings by Širović et al. (2013), who found humpback whales and other baleen whale species were an important component of the soundscape at multiple sites in the North Pacific.

Our analysis revealed only a small correlation between automated humpback whale detections and soundscape metrics, which have been used by other researchers to monitor and quantify humpback whale presence at other locations in the Hawaiian Islands (e.g., Au et al., 2000; Kügler et al., 2020). There were likely other sound sources in our data set contributing to the 125 Hz octave levels, adding to the signals of the whales in this band. There was possibly less energy from humpbacks at this location as chorusing generally occurs in shallower water. Perhaps most importantly, the humpback whale detector determined presence or absence of humpback calls in a 75-s bin and did not indicate the total number of calls in that bin. Call counts may correlate more closely with sound energy in that frequency band.

Another question to be explored is why humpback whale detections correlated best with the energy in the 125 Hz octave, when those calls are known to primarily span 200-1,000 Hz. One possible explanation is that the energy of vessels dominates the daytime soundscape between 150 and 1,000 Hz at this site, as described above. Given this potential masking, the frequencies just below the vessel energy may contain enough energy from humpback whales to produce a correlation. Examination of the correlation at a site without intense vessel noise may reveal a stronger relationship between humpback whale calls and the energy in a higher frequency band (e.g., the 250 or 500 Hz octaves).

Two additional details related to humpback whale detection patterns are revealed through examination of Figure 7. There is a subtle mismatch between the seasonal peak in the 125 Hz octave levels and the peak of humpback whale presence, and there are a few unexpectedly low soundscape metric levels in March and April. The overall mismatches are likely a result of the poor correlation between these two parameters, described above, particularly due to the methods of the humpback whale call detector, which does not discriminate between one whale call and many calls per 75-s window. The peak in the number of calls per hour may be more aligned with the peak in the 125 Hz band, but those measurements are not possible at this time. The low levels in March and April are likely an artifact of non-uniform sampling over time. As revealed in Figure 2, there is limited coverage of March and early April across the 11 years of monitoring, resulting in greater variability of all measurements during those months. Ongoing monitoring at this site during the spring months will hopefully reduce variability and reveal a clearer trend in both parameters.

In contrast to the humpback whales, our odontocete detections showed no seasonal trend, but were highly nocturnal. Although we did not identify the exact species present, many of the odontocetes in this region that produce clicks in the frequencies targeted by our automated detector (∼5–90 kHz) are primarily delphinids, some of which are known to have strong diel behavioral and acoustic cycles (e.g., Norris and Dohl, 1980; Lammers, 2004; Benoit-Bird and Au, 2009; Soldevilla et al., 2010; Wiggins et al., 2013). Most of these species are primarily nocturnal, with increased foraging at night and resting during the day. Because they rely heavily on echolocation for foraging, this behavioral pattern leads to increased acoustic activity at night (e.g., Henderson et al., 2011), which was very clear in our data set (Figure 8). The moderate correlation between the detection of odontocetes and the sound energy in an overlapping frequency band (31.5 kHz octave) indicates that the presence of odontocetes could be monitored using soundscape metrics as a proxy. Although the delphinids’ nocturnal behavior provides some temporal niche separation from the majority of anthropogenic activities and their associated sounds, there is overlap between these two signals during dawn and dusk, which may present opportunity for physical and/or acoustic interactions during these times. It is important to acknowledge that we did not include an analysis of delphind whistles in this study, and although the presence of whistles is very closely linked to echolocation activity for monitoring dolphin presence, we have not explored the contribution of this signal type to the overall soundscape.

Two unique opportunities to explore the soundscape in the absence of anthropogenic noise (no close vessels or echosounder pings) arose from the passages of two hurricanes during the monitoring period (Figures 9, 10). The eyes of hurricanes Iselle (2014) and Darby (2016) passed close to the HARP (within 20–30 km) as the storms each moved past the southern end of the island of Hawai’i.

Iselle moved by the HARP during daytime hours, so the sound energy from wind and rain was present during the daytime (Figure 10A), and was nearly absent from the nighttime soundscape (Figure 10C). This revealed that the storm was characterized by broadband energy at frequencies of approximately 150 Hz and higher (Figure 10A, red), but that the noise from this storm was not higher energy than the typical noise from vessels in those same frequencies (Figure 10A, black).

In contrast, Darby was present near the HARP primarily at night, so we saw lower levels across all frequencies in the daytime spectra due to the absence of vessels (Figure 10B, red), and very elevated levels across all frequencies above 40 Hz in the nighttime spectra due to the storm (Figure 10D, red). This aligns generally with the results of Wiggins et al. (2016) who found a marked increase in sound energy above 300 Hz during the passage of a hurricane close to a HARP in the Gulf of Mexico. Examination of the LTSAs in Figure 9 show that the sound from the storms is slightly different from the daily sounds of vessels, with energy distributed more diffusely across a wider frequency range.

There was a small correlation between wind speeds and sound energy levels at 900 Hz, which was the frequency band with the highest correlation of any of the bands that were tested. This soundscape metric appears to be a combination of energy from both wind and vessels, as shown in Figure 9, such that it reflects vessel presence when vessels are present (e.g., daytime) and reflects wind energy when vessels are absent (e.g., nighttime or during a no-vessel event). The overall low correlation between wind speed and the soundscape metric is likely due to masking of the dominant wind frequencies by vessel noise and to local variations in wind speed due to the topography of the island of Hawai’i, as well as the distance between the HARP site and the nearest wind speed monitoring station, which is onshore at the Kona International Airport.

The passages of both hurricanes also revealed some nuanced human behavior. On one hand, people were eager to maximize time at sea, with vessels present as soon as sunrise on the day after the storms passed. In contrast, the uncertainty and risks associated with the passage of a hurricane are highlighted by examining the time series of wind speed during Hurricane Darby (Figure 9B). Wind speeds were highly variable during the storm, but were not notably higher than a week prior, yet vessels did not leave the dock during the storm. This discrepancy may also be impacted by the issues associated with the wind speed monitoring station described above. Whatever the cause, it seems that wind speeds are not a direct correlate to the presence or absence of vessels.

It is clear that during both Hawaiian hurricanes the absence of vessels had a large impact on the soundscape, even if the addition of sound energy from the storms themselves prevented an assessment of daytime ambient noise under lower wind conditions.