Acoustic monitoring of coastal dolphins and their response to naval mine neutralization exercises

To investigate the potential impacts of naval mine neutralization exercises (MINEX) on odontocete cetaceans, a long-term passive acoustic monitoring study was conducted at a US Navy training range near Virginia Beach, USA. Bottom-moored acoustic recorders were deployed in 2012–2016 near the epicentre of MINEX training activity and were refurbished every 2–4 months. Recordings were analysed for the daily presence/absence of dolphins, and dolphin acoustic activity was quantified in detail for the hours and days before and after 31 MINEX training events. Dolphins occurred in the area year-round, but there was clear seasonal variability, with lower presence during winter months. Dolphins exhibited a behavioural response to underwater detonations. Dolphin acoustic activity near the training location was lower during the hours and days following detonations, suggesting that animals left the area and/or reduced their signalling. Concurrent acoustic monitoring farther away from the training area suggested that the radius of response was between 3 and 6 km. A generalized additive model indicated that the predictors that explained the greatest amount of deviance in the data were the day relative to the training event, the hour of the day and circumstances specific to each training event.


MOL, 0000-0002-7215-4426
To investigate the potential impacts of naval mine neutralization exercises (MINEX) on odontocete cetaceans, a long-term passive acoustic monitoring study was conducted at a US Navy training range near Virginia Beach, USA. Bottommoored acoustic recorders were deployed in 2012-2016 near the epicentre of MINEX training activity and were refurbished every 2-4 months. Recordings were analysed for the daily presence/absence of dolphins, and dolphin acoustic activity was quantified in detail for the hours and days before and after 31 MINEX training events. Dolphins occurred in the area year-round, but there was clear seasonal variability, with lower presence during winter months. Dolphins exhibited a behavioural response to underwater detonations. Dolphin acoustic activity near the training location was lower during the hours and days following detonations, suggesting that animals left the area and/or reduced their signalling. Concurrent acoustic monitoring farther away from the training area suggested that the radius of response was between 3 and 6 km. A generalized additive model indicated that the predictors that explained the greatest amount of deviance in the data were the day relative to the training event, the hour of the day and circumstances specific to each training event.

Introduction
The use of underwater explosives and other high-amplitude impulsive sound in industrial activities, hydrocarbon exploration and military training is becoming increasingly common in many marine habitats around the world, raising concerns about were located 0.5-1 km apart, and their recording periods were offset so that one unit was recording while the other was off. As a result, one of the paired units was always 'on' in order to detect any nearby UNDETs. The two EARs were placed in 14 m water depth approximately 1 km to the north of a site that was identified as the 'epicentre' of MINEX training activity ( figure 1). This is a search field location approximately 7 km offshore from Virginia Beach where the majority (approx. 95%) of MINEX detonations were expected to occur each year. Beginning in September of 2013, two additional EARs were used in a modified deployment configuration. During alternating deployments, four EARs were placed in a 'linear-array' shape, which was aligned to the south, east or north. These EAR units were spaced at approximate distances of 1 km (site B), 3 km (sites E, H or K) 6 km (sites F, I or L), and 12 km (sites G, J or M) from the primary MINEX epicentre (figure 1). The 1 km EAR and 6 km EAR were programmed to record synchronously, and the 3 and 12 km EAR were programmed to record synchronously at an offset duty cycle from the 1 and 3 km EARs in order to ensure the capture of all UNDETS, as in the previous year. Site B was maintained as the 1-km location for all subsequent linear-array deployments to ensure the continuation of the data timeseries from this site. The other deployment configuration in 2013-2016 consisted of site B and three other sites approximately 150 m from site B intended as a localization array; results from the localization effort are not presented here and only data from site B are included in the present analysis.

Data analysis
Dolphin occurrence at site B was analysed using data from August 2012 to August 2015 (table 1). An experienced acoustic analyst manually examined all available recordings obtained from site B for the presence/absence of dolphin signals during this 3-year period using the Matlab™ program Triton [20]. During deployment #7 (16 Feb-27 April 2014), the disk drive of EAR B malfunctioned so the data from EAR K (3 km from the epicentre) were analysed instead. Each 3 min recording containing dolphin signals Table 1. EAR deployment/recovery information. Superscripts after EAR site indicate analyses conducted at that site: 1, UNDET detection; 2, dolphin presence/absence per recording; 3, dolphin acoustic activity during the days surrounding an UNDET ( (whistles, echolocation clicks or burst pulses) was considered a 'detection' of dolphins in the area, and dolphin occurrence was quantified in terms of the number of detections (i.e. number of recordings with dolphin signals present) per day. Because deployments made after August 2015 were intermittent and/or shorter compared to the first 3 years, they were not examined for dolphin occurrence. Three different analysts were employed sequentially over the course of the project and were trained for consistency using a common training dataset. Analysts also manually searched for UNDETs in recordings. Beginning in 2015, an automated Matlab™ script was used to identify training events containing UNDETs by searching recordings for large deviations in acoustic amplitude from background levels with a near instantaneous rise time. Each UNDET was classified as 'small', 'medium' or 'large' based on the duration of the associated acoustic reverberations (0-2 s, 2-7 s and 7+ s, respectively), i.e. the period from initial onset to when acoustic energy fell below 6 dB above pre-UNDET ambient noise levels.
Dolphin acoustic activity was quantified in detail during timescales of seconds, hours and days surrounding UNDETs. For data from EAR B recorded during the first year of the project (Deployments 1-4), an analyst manually counted whistles in the 30 s before and after detected UNDETs; this analysis was not continued in subsequent years. In addition, all 3 min recordings on EAR B in the four days surrounding each MINEX training event were assigned a scalar acoustic activity index between 1 and   Table 2. Index values used to quantify dolphin acoustic activity in each 3 min recording made the day before, day of and the two days after each detected explosion, based on the abundance of dolphin whistles, burst pulses (BP) and echolocation clicks.   (table 2). Recordings without dolphin signals were assigned a value of 0. The analysis period included the day before the event (to represent baseline activity levels), the day of the event, and the two days after each event. For training events occurring over multiple days, the day of the event was the day when the first UNDET was detected; the day after and the second day after were the two days following the final UNDET of the training event. Data from the three more distant EARs (3, 6 and 12 km) within the linear-array configuration in the latter 3 years of the project were also analysed (in addition to EAR B) for the day before, day of, and day after MINEX training events to investigate the acoustic behaviour of dolphins at greater distances from the training epicentre (table 1). Training events were excluded from analysis if they did not have a baseline period of 3 days prior to the event, in order to avoid confounding the 'before' and 'after' periods with other training events. Acoustic activity indices were used to quantitatively compare the acoustic behaviour of dolphins during the hours (site B) and days (all linear-array EARs) before and after MINEX training events throughout all years of the project. To determine whether the observed differences in the hours surrounding explosions were statistically significant, pairwise comparisons were conducted using a Wilcoxon signed-rank test between the hour before and the hour after, second hour after or third hour after an UNDET. Similarly, for the 'day(s)' time scale, hourly indices for each day before, day of, day after and second day after were averaged for the daytime hours (06.00-17.59) and night-time hours (18.00-05.59). The day before values were then matched with the corresponding day of, day after and second day after values and either a parametric paired t-test (for normally distributed data) [21] or a nonparametric Wilcoxon signed-rank test (for non-normally distributed data) [22] was performed. Seasonal variation in daylight hours was not explicitly accounted for in analyses; the 'daytime' hours do, however, represent the typical working day hours for naval exercises [17].

Predictive modelling
In order to test which predictors best explained the variation in dolphin acoustic activity in the hours and days surrounding MINEX training events, a generalized additive model (GAM) was constructed in the mixed GAM computational vehicle (mgcv) package using R software, version 3.  from the training epicentre. The response variable was dolphin acoustic activity, labelled 'mean acoustic activity' (MAA) in the dataset, which was quantified as the mean acoustic activity index for each hour of data recorded the day before, day of, day after or second day after an explosion event, normalized to be between 0 and 1. MAA was tested as a function of nine predictors: season, hour of day, day/night, hours since explosion, size of explosion, number of previous explosions, day (relative to training event), deployment and cumulative hours (table 3). The first three predictor variables were chosen to account for the observed seasonal and diurnal variation in dolphin activity. The next four predictors were chosen because any changes in dolphin activity relative to MINEX training events were expected to be influenced by how close in time they were to explosions and the magnitude of the explosion or event. Deployment (number) was included to account for inter-deployment variation in conditions, and cumulative hours (elapsed since beginning of the dataset) was included to explain variation due to temporal autocorrelation.
A GAM with a β distribution and logit link was chosen for the starting model because the response variable (MAA) was bounded between zero and one. The residuals from the resulting model appeared to have fairly constant variance and showed only slight departures from the expected normal distribution. GAMs with β distributions were also fit to test the probit, cloglog and cauchit links. The different link functions did not have much influence on the residuals or explained deviance. Therefore, the logit link was used in further analysis.
The optimal combination of the nine predictors was determined with dredge in the MuMIn package [26]. All possible combinations of predictors generated a set of models ranked by corrected Akaike information criterion (AICc). The best candidate models (within two AICc of the lowest AICc) all included hours since explosion, hour of day, deployment, season, cumulative hours and day as predictors (table 4). The model with the lowest AICc was selected as the 'best-specified' model as it was the simplest of the three.
To address autocorrelation, the predictor variable cumulative hours was included as described previously. Remaining dependence (autocorrelation) among residuals was tested using acf and the Durbin-Watson test in the lmtest library [27]. The acf function plots a correlogram of the residuals, displaying the approximate correlation at each time lag. The Durbin-Watson test checks for serial correlation between the residuals of the model. The acf plot suggested some autocorrelation exists; however the Durbin-Watson test indicated that autocorrelation was not a significant problem (p = 0.764).

EAR deployments
Fifteen EAR deployment cycles were conducted during the study period (table 1). Two instruments were lost, but were replaced, and one instrument malfunctioned. In total, 609 584 3 min recordings were made among all EARs totalling 30 479 h of data. Of these, 213 176 recordings totalling 10 659 h of data were made at site B, representing coverage of approximately 888 days of the 1423 day study period.

Dolphin occurrence near 'epicentre' area of W-50
Dolphin presence/absence at site B was analysed for the period from 15 August 2012 to 30 August 2015 (deployments 1-12), totalling 799 days of recordings. Dolphins were present almost daily from August 2012 to August 2015 in or near the MINEX range, with detections made on 97% of recording days. During the 3 year period analysed, a clear seasonal trend was observed in the mean number of daily detections each month. Dolphin detection rates (detections/day) were greatest between the months of April and October, decreased substantially between November and March, and were reduced by 90% during the month of February compared to the peak in October (figure 2). However, although the mean number of daily detections decreased during winter months, dolphins were still detected in the area nearly every day throughout the year. For the more comprehensive analyses at the hours and day scales, 31 of the 38 total training events coincided with data successfully obtained from site B, whereas seven training events took place during two deployments when the EAR at site B was either lost or malfunctioned. Of the training events recorded at site B, 31 included baseline data obtained from the day before the event, 30 from the day after the event, and 29 from the second day after the event. The disparity in the number of days recorded is due to the timing of EAR recovery and recording duration relative to two of the training events.

Dolphin acoustic response at site B
Dolphin acoustic activity at site B decreased over the 3 h following an UNDET compared to the hour before ( figure 4). This decrease was statistically significant for the second hour after and third hour after compared to the hour before (Wilcoxon signed-rank test, N = 31, Z = 2.29, p = 0.02, effect size r = 0.41, and N = 31, Z = 2.77 p = 0.006, r = 0.50, respectively).       Dolphin acoustic activity in the days surrounding a MINEX training event varied depending on the time of day. During the day before an event, dolphins were most active during midday (11.00-12.00 local time) and night-time hours (20.00-06.00 local time) (figure 5). On the day of MINEX training and the following day (day after), the midday activity peak was reduced or absent. The reduction in daytime dolphin acoustic activity on the day of and day after was significant compared to the day before an event (Wilcoxon signed-rank test, N = 31, Z = 3.46, p < 0.001, effect size r = 0.62 and N = 30, Z = 2.15, p = 0.032, r = 0.39, respectively). However, on the second day after a training event, daytime dolphin acoustic activity was significantly higher than the day before an event (Wilcoxon signed-rank test, N = 28, Z = 2.07, p = 0.038, effect size r = 0.39). By contrast, during night-time hours, no significant differences were found between dolphin acoustic activity levels from the day before compared to the day of (

Dolphin acoustic response along linear array
Seven linear-array deployments were conducted between September 2013 and March 2016. Overall, the baseline (day before) mean acoustic activity indices at the 3 km and 6 km sites were approximately 15% lower than EAR B, and the mean index at sites 12 km away was approximately 40% lower than EAR B (figure 6). However, hourly patterns in baseline dolphin acoustic activity at the linear-array sites were similar to site B, with elevated activity at night and a peak in activity during the daytime, although the magnitude of the daytime peak decreased with increasing distance from EAR B (figure 6). To statistically infer whether training events influenced dolphin acoustic activity at distances of 3, 6 and 12 km from the epicentre, the hourly indices for each day before, day of and day after were averaged for the daytime and night-time. Because more than 90% of recorded UNDETs occurred between 10.00 and 17.59, the hypothesis was tested that daytime effects would be observed between these hours at distances further away from the epicentre on the day of the training event. The night-time hours examined were those between 18.00 and 05.59, while the day after daytime hours considered were those between 06.00 and 17.59. The values corresponding to the equivalent day before periods were then matched with the  corresponding day of and day after values and either a parametric paired t-test or a non-parametric Wilcoxon signed-rank test was performed. Dolphin acoustic activity on the day before, day of, and day after training events varied depending on the distance of the monitoring site from the epicentre. For the pooled 3 km data (N = 15 MINEX training events recorded), the mean daytime dolphin acoustic activity on the day of the training event was significantly reduced compared to the day before (paired t-test, t = 2.26, d.f. = 14, p = 0.040, Cohen's d = 0.58). No significant differences were found between the daytime hours of the day before and the day after, or between the night-time hours of the day before and either the day of or day after. At 6 km from the epicentre (N = 10 MINEX training events recorded), no significant differences in dolphin acoustic activity were found between the day before and the day of or the day after for either the daytime or night-time periods. Lastly, for the pooled 12 km data (N = 14 MINEX training events recorded), the mean daytime dolphin acoustic activity on the day of the training event was significantly lower than the day before (Wilcoxon signed-rank test, N = 14, Z = 2.41, p = 0.016, effect size r = 0.64). However, at the 12 km sites, no significant differences were found between the daytime hours of the day before and the day after, or between the night-time hours of the day before and either the day of or day after.
To further explore the significant reduction in daytime acoustic activity at the 12 km sites during the day of the training event, data were grouped according to the linear array orientation (north, south or east). At the northern site, the baseline (day before) mean dolphin acoustic activity indices were more than three times greater than the southern site and more than 18 times the activity levels observed at the eastern site ( figure 7). At both the northern and southern sites, mean daytime dolphin acoustic activity decreased by nearly 50% during the day of a training event. At the eastern site, activity levels remained low and unchanged.

Predictive model
The best-fitting model was determined to be the MAA as a function of day, deployment, cumulative hours, hour of day, hours since explosion and season. The explained deviance of this model was 39.07%.   The deviance explained by each individual predictor was found by refitting the model without that predictor, then comparing the explained deviance in the new model to the explained deviance of the original model. Deployment, hour of day, and day (relative to a training event) were the predictors that explained the most deviance (table 5). Figure 8 shows the smoothing splines from the best-fit model for each continuous variable. The standard error for the cumulative hours variable increases as the number of cumulative hours increases beyond approximately 20 000 h, probably due to the small number of data points late in the dataset   Dolphin mean acoustic activity is plotted against categorical variables in figure 9. MAA was greatest in the autumn and lowest in the spring (figure 9a). MAA decreased on the day of and day after a MINEX training event, whereas on the second day after, MAA increased relative to baseline levels of the day before (figure 9b). There was also variation in acoustic activity as a function of deployment (number) (figure 9c), indicating that differences in deployment-specific conditions influenced MAA.

Discussion
Dolphins were detected nearly daily in the MINEX W-50 training area from August 2012-August 2015. Seasonally, there was a period of low occurrence or reduced acoustic activity during the winter months with a minimum in February. This finding is consistent with reported seasonal trends in bottlenose dolphin abundance offshore of Virginia Beach [13,15]; this is by far the most commonly sighted delphinid in this region [13] and the majority of detections in this study are presumed to be of this species. Other delphinid species such as common dolphins are sighted infrequently, but may also have been detected occasionally. The near-continuous presence of dolphins in the MINEX training area indicates that dolphins are exposed to noise from UNDETs, although it is not clear at what range most exposures occur.
There is strong evidence that dolphins respond behaviourally to MINEX training events. The data from site B, comprising 31 monitored training events in 4 years of data collection, paint a clear picture: after a short-term increase in whistling rates in the seconds following an UNDET . Dolphin acoustic activity returned to baseline levels in the evening and night-time hours of that day (figure 5). However, during the daytime hours of the following day, acoustic activity was again reduced compared to baseline levels, but activity normalized again in the evening and at night. Interestingly, during daytime hours of the second day following a training event, dolphin acoustic activity at site B was generally higher than the baseline period.
It cannot be confirmed in the present study whether the observed changes represent individuals moving away from the area, a shift in acoustic signalling behaviour, or both. If the former is true, then the observed patterns suggest that dolphins temporarily move away from the epicentre during the day of the training event as well as the day afterward, but return during night-time hours (although it cannot be determined from this study whether the same individuals were recorded). Perhaps because training events often occur over multiple days, dolphins may anticipate additional UNDETs beyond the final day of the training event, which could explain the reduced acoustic activity observed during the day after the final UNDET. In other words, dolphins may hedge against potential future exposure to an UNDET by avoiding the area. The increase in dolphin acoustic activity on the second day following the training event may indicate that dolphins occupied the area in greater numbers, but this cannot be confirmed at present.
Alternatively, acoustic activity levels may also be a function of individual dolphin sound production rates. In a study of wild Western Atlantic coastal dolphin populations, vocalization rates varied geographically and as a function of behavioural state, and decreased with increasing group size at most locations [31]. The higher acoustic activity levels determined two days after a training event could indicate more frequent signalling by dolphins, perhaps reflecting altered behaviour and/or group sizes compared to the baseline. In captive animals, stressful events can lead to periods of reduced or no acoustic activity lasting hours or even days [32,33]. It is not known whether free-ranging animals respond similarly, but lower acoustic activity levels the day of and day following a training exercise may also signify a stress response.
The data obtained from the EARs 3, 6 and 12 km away from the epicentre further help inform our understanding of the response by dolphins to MINEX training events. There is evidence that dolphin acoustic activity is reduced 3 km away from the epicentre during the day of an UNDET, but not 6 km away, which suggests that the dolphins potentially respond to UNDETs at ranges up to 3-6 km (figure 6). These results are consistent with studies of captive bottlenose dolphins and a beluga whale (Delphinapteras leucas), which demonstrated disruption to trained behaviours in response to simulated explosions at ranges of 1.5 to 9.3 km depending on the size of the charge [4]. Of note, however, is that a significant reduction in acoustic activity on the day of an UNDET was also observed at the two 12 km sites towards the north and south of the epicentre ( figure 7). This suggests that animals from these areas responded to relatively distant UNDETs or other aspects of training exercises. It is unclear whether this response indicates animal movement, behavioural changes or both. One possibility is that the animals may be moving away from more distant areas toward the epicentre to exploit prey fauna harmed or disoriented by the UNDET, during the night-time hours in between or after training days when acoustic activity was at normal levels. Another possibility is that habituation to UNDETs exists among animals typically occurring approximately 6 km away, but not among those substantially further away, which may be exposed to training events less frequently. More detailed studies are needed to better understand the response by these more distant animals to the north and south of the epicentre.
For the purposes of this study, it was not possible or feasible to empirically determine detection distances of dolphin signals or UNDETs, as the locations of sound sources were not available (or able to be localized). Janik demonstrated the range of detectability could be up to 20-25 km for dolphin whistles at maximum source level in Moray Firth, Scotland [34]. However, this was a relatively deep water habitat and these distances were estimated in low noise and Beaufort 0 sea state. Previous studies of bottlenose dolphins in more comparable high-noise, shallow, mud/sand bottom habitats in the eastern US and western Australia reported lower source levels of whistles and estimated the active space of dolphin whistles to range from less than 1 km to 2-6 km [35][36][37]. In 2014, work was conducted to measure received levels of UNDETs and derive a semi-empirical model for the propagation of underwater detonations in shallow water in/near the VACAPES W-50 range [38]. Sound exposure levels and peak pressure are calculated for 1, 3, 6 and 12 km for typical MINEX UNDET charge weights in table 6. Using the equations in [38], the sound exposure level for a receiver at 12 km away from a 20 lb charge would be 173 dB re 1 µPa 2 s (table 6). Because of this and the offset duty cycles between neighbouring EARs, it is likely that EARs in the linear array configuration recorded independent groups of dolphins, and also that dolphins 12 km from an UNDET were still exposed to the sound.   The modelling effort corroborated the empirically derived results and suggested some additional factors to consider. The explanatory power of the variables day and hour of day supports the finding that dolphin acoustic activity varied significantly on the days during and after an event compared to the day before, as well as the consistent diel pattern observed in dolphin acoustic activity. The inclusion of the deployment variable suggests that MAA was influenced by deployment-or event-specific conditions, such as the number of dolphins present prior to and during the training event, ambient noise conditions and the precise UNDET locations relative to the 1 km EAR site. Other predictors included in the best-fit model were season, hours since explosion and cumulative hours. The variation in MAA in response to season was consistent with the observed peak dolphin activity in autumn and lowest activity in spring. The increase in MAA with increasing hours since explosion for time periods of 60 h to two months may indicate that dolphins gradually returned to natural activity levels after a training event. The decrease in MAA with more than two months since the previous explosion does not lend itself to meaningful interpretation because conditions could have changed or training events may have taken place between deployments (typically approximately 2 months in duration). The variation in MAA observed over the course of the study (as a function of cumulative hours) may reflect the fact that three different analysts manually scanned different periods of recordings, which roughly corresponded to the different shape of the MAA function for 0-5 months, the middle 2 years and the final approximately 7-8 months. Although analysts were trained for consistency, some subjective variability in assigning acoustic indices may have persisted. However, because the response to training events was statistically analysed in a pairwise manner, any inter-analyst variability would not have affected the results.
The biological implications, if there are any, of the documented responses cannot be resolved by the present study. However, among the deleterious effects at the organismal level that might be expected are: chronically elevated stress hormone levels, acute or cumulative hearing damage/loss, reduced access to foraging habitat, and/or increased energy expenditures to avoid training areas [1,[39][40][41]. For example, Rolland et al. [39] provide evidence that exposure to ship noise was responsible for increased levels of stress-related hormones in North Atlantic right whales (Eubalaena glacialis). Such effects may result in reduced fitness and possible declines in reproductive rates at the population level. Other possible impacts could include shifts in distribution and local reduction in abundance; such effects were demonstrated in bottlenose dolphins (Tursiops sp.) in Australia as a result of chronic disturbance from tourism vessels [40]. More detailed data on received sound characteristics are needed to interpret responses and link sound exposure to potential population-level effects [41].
Follow-on work would be useful to investigate the conservation implications and to help the Navy improve mitigation strategies. For example, future studies could investigate sound exposure levels as well as the types (physiological, behavioural, etc.) and magnitude of individual and group responses, potentially with the use of short-term tags, focal follow studies, unmanned aerial vehicles, hydrophone arrays and other means. In addition, analyses of the existing data could target calls from other marine mammals, including baleen whales and pinnipeds that occur in the area, to determine whether they also exhibit an acoustic response to MINEX training. Finally, long-term monitoring using passive acoustics, visual, and/or other means should be considered in order to detect any changes over time in dolphin occurrence and behaviour, and other environmental and anthropogenic factors should be integrated into cumulative impacts studies.