We reviewed the three months of data to determine periods of vocalizing marine mammal species and anthropogenic sounds. The raw audio (.wav files) acoustic dataset was processed in PAMGuard (version 2.01.05; Gillespie et al., 2009) using a two-stage process. PAMGuard is a widely used, open access software program that includes automated and semi-automated modules for the detection and localization of marine mammals. PAMGuard is a modular program that allows the user to define what display, detection, localizations and noise monitoring modules they wish to use in an analysis. The first stage of processing occurred in PAMGuard’s standard mode and involved automated detection of calls from several marine mammal species using a combination of the in-built click and whistle and moan detector modules. Settings for the click detector included standard settings and a series of click classifiers to differentiate delphinid and sperm whale echolocation clicks. Two click classifiers were used for the detection of echolocation clicks from sperm whales to differentiate between “regular” and “slow” clicks (Weilgart & Whitehead, 1988; Barlow & Taylor, 2005). We wanted to make sure both click types were detectable in the processing. The “slow” click classifier consisted of a detection/test band of 500 Hz – 8 kHz, and peak frequency range of 500 Hz – 4 kHz. The “regular” click classifier consisted of a detection/test band of 1 – 12 kHz and peak frequency range of 5-12 kHz. All tonal and burst pulse type of calls were detected using the whistle and moan detector. Frequency range was adjusted to detect all potential sounds within the available bandwidth, requiring true positives to be determined in the secondary stage (Gillespie et al., 2009; Gillespie et al., 2013). Default settings for noise and thresholding were retained.

The second stage of processing involved using PAMGuard’s ViewerMode for the post-automation annotation of the automated detections. Annotation was performed by an experienced bioacoustic analyst and involved discrimination of detections into appropriate “acoustic events.” An acoustic event was defined as a bout of calls from a biological source that had no more than 30 minutes gap between calls from the same species for whistles or burst pulse calls, and no more than 60 minutes gap between bouts for calls from sperm whales and baleen whale species. This selection of time interval is mostly just used as a means of organizing data that are “likely” from the same individual or group, given their potential to remain detected in the area for a certain amount of time. We were not however using this information for any density or abundance determination, so the organization is mostly just useful for data management purposes. Calls attributable to marine mammals were assigned to “acoustic events” according to species or species group (e.g., delphinid species, fin whale). Annotations were made by using the Detection Grouper module in PAMGuard, which allows a user to selected detected calls and assign those calls to a group which can then have custom information associated with the group. The assignment of calls to a group relies on an analyst identifying a period of calls within the data, and then boxing calls to assign to the group. Any calls that were ambiguous or unidentifiable simply received an “unidentified sound” or “unidentified cetacean” assignment. Any detections that were attributable to ambient or instrument noise (false positives) were disregarded, so that only annotations of true events were included in further processing. Annotations were intermittently reviewed by a second, expert bioacoustic analyst to ensure species identification were accurate.

Processing data for all potential calls within the same processing run can be computationally intensive. To increase efficiency of the annotation process, three analysis iterations were completed so the analyst could focus on classifying calls from a subset of possible species/species groups. The very low frequency run (10-200 Hz) focused on calls from baleen whales such as fin and blue whales. The low frequency run (200 Hz – 2 kHz) focused on slightly higher frequency calls, such as those from humpback whales. The mid-high frequency run (2 kHz – 32 kHz) was used to annotate calls from odontocetes, such as sperm whales and delphinids. During annotation, several modules are used to expedite the review of automated detections and parse out true detections from noise This specifically included the Long-term Spectral Average module which allows average spectrum measurements over periods of several seconds (10 seconds in this study) and a clip generator that organizes clips of detections for rapid review. Acoustic event annotations were not used to identify individuals or specific groups within the data, as an animal may have moved past the recorder and returned; acoustic events were solely used to assist in the data management of acoustic encounters.

Periods of passing ship often triggered false positives on the click detector. These false positives were reliable for documenting these underwater sounds and were annotated as “Passing ship” acoustic events. Ships were only annotated in this way if they exhibited the same ship event structure observed in the long-term spectral average display by other studies (Ragland et al., 2022). Mid-frequency active sonar was annotated by selecting contours from the whistle and moan detector. No other sources of anthropogenic noise (e.g., echosounders) were annotated in this analysis.

Acoustic event-based information including the species identification, the count of calls (number of calls per audio file, per species), and select additional parameters were extracted using the R-based package ‘PAMpal’ (v0.13.0; Sakai et al., 2020). PAMpal extracts information from the binary files and database generated by PAMGuard, and allows users to calculate parameters from detected calls and summarize this information in user selected time intervals. A call or detection count was summed for each five-minute wav file, and multiple acoustic events occurring within the same file were summarized independently. Additionally, select parameters calculated from the “contours” (which are the detections made by the whistle and moan detector) included the mean call amplitude (dB), mean call bandwidth (Hz), and mean frequency (Hz) determined for each 5-minute wav file. For echolocation clicks, we extracted the total number of clicks as well as average measures of mean click amplitude (dBPP - a relative measure of intensity/amplitude calculated as 20 * log10 of the difference between the maximum and minimum value of the waveform), mean 10 dB bandwidth (kHz), and mean peak frequency (kHz) for each five-minute wav file associated with sperm whale acoustic events. We did not differentiate between "slow" and "regular" clicks for parameter measurement extraction, so data from both click categories were combined for analyses. Additionally, we extracted occurrence information (e.g., labeled a five-minute wav file as containing a type of sound) for anthropogenic sounds such as passing ships and mid-frequency active sonar.

Output from R calculations of acoustic indices were merged with summaries of information in five-minute files from marine mammal and anthropogenic acoustic events. We assigned details of call types to indicate if each five-minute wav files contained contours from a single event, from overlapping (different species) events, contours and clicks from single or overlapping events, or periods where biophony and anthrophony were not detected. When correlating call characteristics to indices, we excluded all five-minute wav files with multiple species. For all analysis and graphs comparing index measurements with a specific sound type, only data containing calls from that specific species/species group or anthropogenic noise source were included. We did not include periods where calls or sounds overlapped. We are confident that all possible sounds from marine mammals and anthropogenic noise were annotated in each dataset, meaning the “Vocalizations absent” category can only have included ambient and instrument noise or potentially noise from geophony. We did not exclude periods with rain or storms from this “Vocalizations absent” category as we wanted the natural variability of ocean ambient noise to be reflected in those periods without marine mammal or anthropogenic sounds. Finally, prior to performing statistical tests and descriptive graphics, we determined the average call count from the data specific to each recorder for all marine mammal species and excluded all five-minute wav files that contained fewer calls than the mean call count value across all acoustic events. This was done to ensure that data used in the statistical tests and comparative figures included a sufficient number of calls to relate to quiet periods without vocalizations.

The statistical analyses included in our study were identified due to the presumed potential for pseudoreplication, or observations that could include temporal or spatial dependence. As posed by Alcocer et al. (2022), the data had the large potential for temporal autocorrelation given successive five-minute periods were identified as individual data points. To account for this issue we incorporated statistical measures that are robust to the non-independence, as demonstrated by Moreno-Gómez et al. (2019) in a study of acoustic indices and bird and anuran richness.

Evaluating the Relevance of Species to Index Measurement Variation: Determining the effect of species or anthropogenic sound source on acoustic indices was an important element to this study. Boxplots grouped by recorder and reported by species/species group or sound type were initially created. We then assessed the contribution of species to the acoustic index measurements by fitting a per-index generalized linear mixed-effects model using the ‘lme4’ R package (Bates et al., 2015). Given the size of this dataset and the variability in detected sounds on each recorder, as well as the difference in month for one of the recorders, each index was fitted with a single model that examined species as a fixed effect and used recorder as a random effect. As we did not produce separate models to compare in this analysis, we do not report AIC values or assess the best fit model.

Correlating Index Measurements to Call Characteristics: We examined the relationship between each of several acoustic indices and characteristics from species where a notable difference from periods without vocalizations were observed in the descriptive figures. This was performed using a Spearman correlation test to determine a correlation coefficient (“ρ”) to measure the strength and direction of the variables' relationship using the spearmanRho function of the ‘rcompanion’ package in R (Mangiafico, 2016). The nonparametric Spearman correlation was used due to all index measurements exhibiting a non-normal distribution. To mitigate issues that could potentially arise due to data dependence for each recorder, we performed a bootstrap routine with 1,000 iterations. Bootstrapped correlation coefficient measurements and associated confidence intervals were calculated for all extracted call measurements.

A summary of sample sizes were reported for all call types and all recorders. All statistical analysis was performed using the native stats package in R (v4.2.2; R Core Team, 2022) and figures were generated using the ‘ggplot2’ package in R.

Marine mammals, passing ships, and sonar activity were identified in the dataset and the number of five-minute audio files containing each sound category varied for each sound category and recorder ( Table 1 ). Differences in recording coverage during each month varied, resulting in differences in the total sample size of five-minute files for each site (shelf = 6,410 files, slope = 4,725 files, base of the slope = 7,101 files). Periods of vocalizing marine mammals occurred less frequently at the base of the slope (HYDBBA 103) as compared to the slope (HYDBBA 105) and the shelf (HYDBBA 106). Calls from delphinid species (not identifiable to species level), humpback whales (Megaptera novaeangliae) and fin whales (Balaenoptera physalus) were detected at all three sites. Delphinid calls predominantly consisted of burst pulses and could not be classified to species as they could have been produced from one of several species and visual validation was unavailable. Echolocation clicks from sperm whales (Physeter macrocephalus) were detected on the slope and at the base of the slope. Blue whale (Balaenoptera musculus) A and B calls and mid-frequency active sonar were detected on the recorder located on the shelf. Possible fish sounds were only detected at the base of the slope. Passing ships were detected at the slope and shelf sites. Within each set of recordings, all five-minute files where vocalizations were absent were included in the comparison (4188 – 6943 files). The combined dataset including information on marine mammal occurrence, call measurement and acoustic indices can be found in Supplementary Table 1.

The acoustic index measurements varied in association with calls produced by several marine mammal species and anthropogenic noise sources at each of the OOI recorder sites ( Figures 2 - 5 ; Supplementary Figures 1, 2 ). ACI measurements from five-minute wav files notably increased during periods of vocalizing delphinid species and sperm whales ( Figure 2 ). AD and AE resulted in complementary, opposite measurements and are thus summarized collectively. These indices responded inconsistently to different marine mammal calls, and were more obscure in terms of the measurement response ( Table 2 ; Supplementary Figures 1, 2 ). BI measurements decreased in associationwith to sperm whale echolocation clicks for both the recorders ( Figure 3 ). Humpback whales had a smaller significant effect on BI at all three sites exhibiting a slightly higher distribution of values than delphinids, fin whales and blue whales. The NDSI varied for all marine mammal groups across the three recorders ( Figure 4 ). Humpback whale calls resulted in a reduction ofNDSI measurement on the slope and at the base of the slope, but were insignificant on the shelf. There was a notable reduction in the measurements associated with the possible fish category at the base of the slope. The NP and H measurements also exhibited variable responses to calls from marine mammals ( Supplementary Figure 3 and Figure 5 , respectively). Although differences in NP measurements were noted for several species, the increase or decrease in measurements as compared to periods without vocalizations was inconsistent at each recorder and displayed no clear trend in the data. The H measurements increased in association with blue whale or delphinid calls and decreased in connection to the possible fish sounds ( Figure 5 ).

The results of the generalized linear mixed-effects models provide additional evidence of the contribution of each species or species groups to the variability of the dependent variable (each index; Table 2 ). For ACI, all categories but “Possible Fish” and “Sonar” indicate they significantly contribute as a predictor variable in the model. Delphinid species and sperm whales demonstrate a higher likelihood of increasing this metric. For AD and AE, although blue whales, delphinid species, fin whales, and sperm whales indicate a statistically significant effect on the model, the coefficient estimate is relatively low for these indices. Only humpback whales and sperm whales significantly contribute to the variation for BI, with the sperm whales resulting in a large negative coefficient. The model for H model does indicate significant differences between many species, but they have a lower likelihood of increasing or decreasing this measurement. All but humpback whales have the potential to predict the NDSI variable, although blue whales are more likely to increase this value while possible fish are more likely to decrease it. Finally, the NP measurements are significantly predicted by all species but fin whales, although the likelihood of this impact is nominal.

In assessing the difference in ACI, BI and NDSI measurements for periods with sperm whale clicks and periods without vocalizations, we observed an incremental differentiation when qualitatively parsing data into categories of loud, moderate, and quiet echolocation clicks ( Figure 6 ). The BI measurements were dramatically reduced as compared to periods without vocalizations for loud and moderate clicks. To a lesser extent, the same trend was observed for the ACI and NDSI values particularly for those echolocation clicks that were qualitatively classified as “loud” by the analyst.

We explored the correlation of call characteristics of sperm whale for both ACI and BI ( Figure 7 ). For delphinid calls we only evaluated data that contained “contour” detections of burst pulses, and did not include measurements of echolocation clicks in the correlation comparison. Annotated echolocation clicks from delphinids resulted in too low of a sample size to evaluate independently, and burst pulses make up the majority of the annotations for the delphinids in this region. Although additional significant relationships were found for other species and indices, these two indices provided consistent trends in the data distribution from multiple recorders. Figure 7 indicates the Spearman correlation (ρ) and bootstrapped confidence intervals for each sperm whale call characteristic and ACI and BI. The 10 dB bandwidth (ρ₁₀₃ = 0.41, ρ₁₀₅ = 0.37), click amplitude (ρ₁₀₃ = 0.69, ρ₁₀₅ = 0.51), peak frequency (ρ₁₀₃ = 0.63, ρ₁₀₅ = 0.30), and Total Clicks (number of clicks per 5-min file) (ρ₁₀₃ = 0.47, ρ₁₀₅ = 0.61), for ACI values of sperm whale clicks had a moderate to strong, positive correlation with ACI at both sites. Confidence intervals were greater for the recorder at the base of the slope due to smaller sample size, but were consistent with the trends of the slope recorder (105) ( Figure 7 ). For the BI measurements, the 10 dB bandwidth (ρ₁₀₃ = -0.33, ρ₁₀₅ = -0.43), click amplitude (ρ₁₀₃ = -0. 39, ρ₁₀₅ = -0.65), peak frequency (ρ₁₀₃ = -0.37, ρ₁₀₅ = -0.43), and Total Clicks (number of clicks per 5-min file) (ρ₁₀₃ = -0.34, ρ₁₀₅ = -0.55) all indicate a moderate to strong negative correlation with sperm whale clicks. Of these metrics, the click amplitude and total number of clicks per 5-min file suggest the strongest influence.

For delphinids, inconsistency across in the correlation results, closer to zero correlation values for some call parameters, and wide confidence intervals do not suggest any specific influence on ACI ( Figure 8 ). These results were unexpected given the increase in ACI observed from the descriptive figures. Supplementary Tables 2 and 3 include the correlation coefficient values and confidence intervals calculated in this analysis.

Annotated anthrophony consisted of ships on the shelf and the slope and mid-frequency active sonar pings at the shelf. Five-minute files containing noise from passing ships resulted in an increase of ACI as compared to periods without vocalizations ( Figure 2 ). ACI did not change significantly during periods with sonar, although sample size was low ( Table 1 ). Sonar was noticeably lower for AD and higher for AE Supplementary Figures 1, 2 ). Similar to sperm whale observations, sonar was associated with reduced measurement of BI in comparison to periods where vocalizations were absent Figure 3 ). Ships had a negligible significant difference from quiet periods for BI, but again were noticeably different from quiet periods on the slope.

The NDSI measurements were greatly reduced for passing ships as compared to periods lacking vocalizations at the shelf and slope recorders ( Figure 4 ). For sonar, large increased difference in NDSI was observed as compared to periods where vocalizations were absent A similar trend was observed for the NP measurements for both passing ships and sonar ( Supplementary Figure 3 ). Sonar periods are associated with increased measurements of H as compared to periods without vocalizations, however passing ships were dramatically reduced from quiet periods on the shelf as well as along the slope ( Figure 5 ).

The results of the generalized linear mixed-effects models for passing ships and sonar indicate interesting influences on several index variables ( Table 2 ). Passing ships have a small likelihood of increasing ACI, but a larger likelihood of decreasing BI and decreasing BI and NDSI. The sonar sound category had a large positive effect on AE and NDSI, and had a large negative coefficient relating to BI. The likelihood of sonar and passive ships increasing or decreasing NDSI respectively is far greater than the intercept (Vocalizations Absent).

Index measurements are recorder dependent making them incomparable in terms of exact values, but this study results in several trends in the response of some of these measurements to calls from marine mammals and anthropogenic noise sources. We therefore focus our interpretation of these results on sound categories that exhibited a consistent, strong effect on the measurements of a select set of indices. Given the difference in instrument and ambient noise on the shelf, slope, and base of the slope recorders, these robust trends are anticipated to persist on different instruments located in similar acoustic environments:

We did observe additional trends in the data that could be used as a further means of rapidly determining periods of biophony and anthrophony in a dataset, however the data were either limited to one site (e.g., blue whales) or require a complex machine learning approach to discriminate which combination of indices can provide meaningful information. Based on the efforts of this study, we suggest that sperm whales and delphinid species could serve as representative acoustic indicators in coastal ecosystem monitoring efforts that incorporate acoustic indices.

Anthropogenic sources of noise in this dataset provided information beyond the well-established relationship between intensity indices such as sound pressure level and ship traffic. Additionally, we found sonar to have a significant influence on several indices, however sample size of five-minute files containing sonar were low (n=22) and only occurred at the shelf recorder. A larger dataset may elucidate trends in this anthropogenic noise source. Although noise filtering was suggested by several studies as a means of focusing on biodiversity of calling species, we suggest this might mask the effects of louder signals on index measurement response.

Additionally, we recommend evaluating the effects of call type and rate of a variety of fish species to better understand their relationship with acoustic indices. Armed with a more robust understanding of key influential sources of biophony, marine ecologists can more effectively determine those combinations of acoustic indices that contribute to understanding biodiversity in ocean environments.

To our knowledge, this study is the first to compare acoustic indices across large bandwidth (10 Hz to 32 kHz) data for assessing multiple types of underwater sounds. However, the integration across such a large bandwidth may be marginalizing the low frequency calls from baleen whales within these acoustic indices results. It is likely that incorporating a series of calculations exclusively on a lower bandwidth could result a larger response from acoustic indices, as was evident in Parks et al. (2014). We suggest that bandwidth possibly impacts acoustic index associations as it pertains to both low frequency calling animals and cases where bandwidths are much greater than the 32 kHz of this study. We could not evaluate calls from species that produce vocalizations above the 32 kHz bandwidth, such as beaked whales and harbour porpoise. Given the influence on measurements for several indices in response to sperm whale echolocation clicks, evaluating other echolocating species would be beneficial. Another caveat is that some species did not produce vocalizations at each recorder site, limiting our ability to compare between different acoustic recorder environments.

The results are further limited as they were collected in the same region of the northwest Pacific Ocean and used data from the same type of instrumentation. However, the variability in the recorders in terms of instrument noise from the hydrophone or co-located sensors and acoustic environment provided more confidence in those persistent trends mentioned above. Additional analysis of data from a different region and using different instrumentation would assist in further validation of these results.

The concept of utilizing acoustic indices to characterize the biophony diversity found in acoustic environments is not novel, nor is the acknowledgement that further research is required to improve the reliability of these measurements as indicators (Sueur et al., 2014). Roca and Van Opzeeland (2020) have already tested incorporating a suite of acoustic index measurements in machine learning methods for monitoring acoustic environments using large acoustic datasets. We anticipate this work will contribute to current efforts seeking to incorporate acoustic indices in autonomous, machine learning methods for continuous environmental monitoring (Sethi et al., 2019; Williams et al., 2022).

Implementing meaningful large-scale ecosystem health monitoring has additional challenges. Recommendations from the acoustic community suggest that fluctuations in oceanographic measurements associated with indicator species may elucidate periods of ecological disturbances (Pieretti and Danovaro, 2020; Minello et al., 2021). To truly assess this would require the co-location of additional physical and chemical oceanographic data as is the case with the OOI cabled array sensors. Mooney et al. (2020), indicated the need to record over sufficient timescales to observe deviations from typical seasonal and interannual variability in a marine habitat. Therefore, combining acoustic indices as indirect, biological indicators in concert with physical and chemical monitoring over long time periods, such as is possible with the OOI Coastal Endurance and Continental Margin arrays would be an ideal scenario for accurately monitoring ecosystem health. In this context, acoustic indices could provide a solution to the issue of underutilized passive acoustic data and contribute to biodiversity conservation within marine ecosystems.