Md and Zc are both medium-sized oceanic whales that live in small mixed composition groups of about 2–4 animals (Baird, et al., 2006; Falcone, et al., 2009; Marques, et al., 2019; Barlow, et al., 2006; MacLeod, et al., 2006) and are usually found in deep offshore waters. Both species also execute regular, breath-hold dives to hunt for mesopelagic and bathypelagic squid and fish (Santos, et al., 2001) using echolocation (Tyack, et al., 2006; Madsen, et al., 2013). Md and Zc frequently dive to depths over 1,000 m (Tyack, et al., 2006) during dives lasting 45 min or more. Zc tagged near SOAR hold the records for both longest (137.5 min) and deepest (2992 m) mammal dives ever recorded (Schorr et al., 2014). During their deep dives Md and Zc emit two distinct types of click waveforms, foraging clicks (Baumann-Pickering, et al., 2013) and buzz clicks (Figure 1). A period of foraging clicks prefaces each buzz and this pattern repeats approximately 30 times per dive, on average (Madsen et al., 2013). Sound production in both Md and Zc is highly directional with an estimated main lobe beam width of approximately 20–30° and significant side lobe suppression (Zimmer, et al., 2005; Zimmer, et al., 2008; Shaffer et al., 2013). Essentially Md and Zc search their environment with an acoustic flashlight where a narrow beam ensonifies finite swaths ahead of the animal along the axis of its melon (Zimmer et al., 2005). From the perspective of an omnidirectional bottom-mounted receiver, the sounds sweep through its hearing radius as the animal searches. The amplitude of the received clicks increase and decrease in response. Over a grid of hydrophones such as AUTEC or SOAR we see the foraging clicks from a diving Zc or Md “move” from one hydrophone to its neighbors as the animal’s main response axis turns away from one toward the other. Frequently the loudest received clicks are not necessarily received on the hydrophone closest to the animal but on the one most directly in line with the main lobe of its biosonar. Therefore, we must consider the ensemble of clicks received by the local group of hydrophones when trying to detect beaked whale dive events.

Output reports from the CS-SVM classifier archived at two Navy Ranges, AUTEC and SOAR, were analyzed for this study. AUTEC, which is located in the Tongue of the Ocean off Andros Island, Bahamas, has 91 broadband hydrophones arranged in a roughly 48 km × 24 km grid. Fourteen of the hydrophones are spaced approximately 2 km apart while the remaining 75 are spaced approximately 4 km apart (Figure 2A). The total instrumented range area is approximately 1,200 km². SOAR is located in the San Nicolas Basin, west of San Clemente Island, California. The range has 153 broadband hydrophones available but the CS-SVM classifier was running on only the newest 89 hydrophones at the time of the study. The spacing of the SOAR hydrophones is depth dependent (from approximately 3.5–5 km) and the total area of coverage is approximately 1700 km² (Figure 2B). The signal conditioning method for acoustic data is the same at both sites.

Analog data from each hydrophone are digitized using a sampling frequency of 96 kHz and passed to a dedicated processing stream. The digitized data for that hydrophone are high pass filtered and a time-domain energy detector is used to detect arriving clicks. If the peak of a click is above the noise-variable threshold (NVT) then a 2.67 ms (256 samples at 96 kHz) long snippet of the time series about the peak is sent to the CS-SVM classifier for feature extraction then the feature vector is input to the CS-SVM decision functions. During the study, the CS-SVM was configured with the following classes: 1) Md foraging clicks, 2) Zc foraging clicks, 3) Sperm whale foraging clicks, 4) generalized dolphin clicks. Since the data input from each hydrophone is a continuous, real-time stream, which can contain clicks from multiple animals of the same or different species, the CS-SVM classifies each click it sees individually. The NVT is typically set to ride 20 dB above the time-varying average noise level calculated using an exponential average (Jarvis et al., 2014). A buzz classifier is launched on the data stream from a given hydrophone only when 80% of clicks detected on that hydrophone over the last 20 s are classified as either Md or Zc. The buzz classifier process then runs for 30 min, the approximate vocal period of an Md dive (Tyack, et al., 2006), before automatically stopping. The NVT for the buzz classifier is set only 6 dB above the average noise level. To further reduce the likelihood of false buzz detections and better match the known amplitude characteristics of buzz clicks, a maximum peak amplitude threshold is also used. The peak value of clicks considered by the buzz classifier must be less than ½ of the average of the local foraging click peaks seen prior to the buzz. That is, the received peak of a candidate buzz click must be a minimum of 6 dB below the local foraging peaks (METEOR Team, 2014). Output reports from the classifier (i.e. detection reports), as well as other M3R data like spectrogram and localization data, from each hydrophone are written to a central archive on a large (∼3 TB) external hard disk drive. When the archive drives at the ranges are full they are replaced and the archive files they contain are transferred to permanent storage on M3R’s petabyte-capable data server in Newport, RI.

The M3R systems at each Navy range collect archives of detection reports nearly continuously. The archive data analyzed in this study include 36 h of archives recorded 8–10 July 2014 at AUTEC and a set of archives collected over 3.5 years at SOAR (Table 1). To identify click trains that are indicative of foraging dives, archive data are post-process through M3R’s click train processor (CTP) program. CTP generates click trains for each detection algorithm and class (TYP) on a per-hydrophone (HYD) basis. Two beaked whale classes are considered by the CTP. Md foraging clicks are designated as class number 1 and Zc foraging clicks are class number 2. A click-train is started when a click is detected, and clicks of that detector type and class are added to the click-train until at least 180 s (timeout (TMO)) pass without additional detections. At this point, if the click train has at least five clicks (click requirement (RQ)), a click train report is generated; otherwise the click train is discarded. The click train report includes the hydrophone number, detector type (CS-SVM is type 17) and class (Md = 1, Zc = 2), the start and stop times for the click-train, the total number of clicks in the click-train (CNT), and the ICI. An example of the unfiltered CTP output for archives from SOAR is shown in Table 2, with the CS-SVM Zc detections (TYP 17:02) highlighted in blue.

The per-hydrophone click trains identified by CTP are then associated with other click trains on nearby hydrophones to form group dives using our auto-grouper (AG) program. The CTP trains are first filtered for desired type and class (e.g. CS-SVM Md and Zc are types 17:1 and 17:2, respectively), then for a user-selectable ICI range (0.23–0.40 s for Md and 0.35–0.75 s for Zc), followed by duration (click trains must be less than 60 min long and contain a minimum of 300 foraging clicks). Finally, the candidate click trains are sorted by start time. The AG algorithm uses the CTP output to form groups by first identifying trains from the set of filtered click trains with the highest click counts, and then adding to their groups click trains from nearby hydrophones with successively smaller click counts. The set of neighboring hydrophones generated for each hydrophone include all hydrophones within a fixed radius (usually ∼6 km) of the center hydrophone. The radius selected is based on both range to visual Zc observations at SOAR (DiMarzio & Jarvis, 2016; Moretti, et al., 2010) and calculated hearing radius for Md clicks at AUTEC (Ward et al., 2011).

Buzzes are identified using the buzz train processor (BTP) algorithm. The BTP works in much the same way as the CTP. It identifies buzz events by applying a set of heuristic rules to the buzz click detection reports received over time on a given hydrophone. The BTP requires a minimum of five buzz click detections to start and the buzz must be less than 6 seconds long (Tyack, et al., 2006; Madsen et al., 2013). The amplitude of buzz clicks received varies with whale’s motion and received amplitude (Figure 1D) can fall below the NVT during the buzz. Therefore, any buzz clicks received on a hydrophone within a given 6 s window are grouped into one buzz. There also must be a minimum of 10 s without buzz click detections between successive buzzes. This is to guard against buzz classifier false positives caused by distant dolphin clicks. In general, dolphin clicking tends to be continuous over durations much longer than a buzz. Once the times of candidate buzzes are identified, they are compared to the start and stop times for group dives (AG output) on that hydrophone. Only BTP buzz events that occur during an identified group dive are counted.

CS-SVM classifier detection reports, including CS-SVM buzz classifier reports, were extracted from approximately 36 h of M3R AUTEC archives from 8–10 July 2014. These archives were collected just after the buzz classifier was deployed at AUTEC. Md foraging click detections were grouped into click trains and groups using the rules of the CTP and AG processes discussed above. The detections from the groups were plotted and the dives manually reviewed and validated. Figure 3 shows an example of output of the CS-SVM classifier showing an Md foraging click train as received on hydrophone 64 and neighbouring hydrophones. Blue dots indicate the time of Md foraging click detections on a given hydrophone (listed on vertical axis). Magenta squares indicate the first and last clicks in a click train according to the rules of the CTP. Initially we also reviewed Zc foraging click detections (both Md and Zc have been visually verified at AUTEC) but the number of dives detected during the study period was low and several of these were detected only on edge hydrophones. Such edge phone detections are not typically included in dive counting analysis because the animals are likely off the range proper with their heads pointing in the direction of the receiving edge hydrophone.

Figure 4 shows the CS-SVM classifier detections versus time from a single hydrophone. Here, again, blue dots are Md foraging clicks and blue x’s are Md buzz click detections. The black dots represent generalized dolphin detections. In Figure 4A these are most likely false positives and only Md are present. However, Figure 4B shows a case where there likely was dolphin activity and buzz detections persist after the foraging click detections stop. These buzz detections are viewed as false positives and fail the rules of the BTP. Figure 4C shows detection of a long buzz on hydrophone 64. The three clumps of buzz detections are grouped as one buzz because they occur within the same 6-s window. The 5 s gap between foraging clicks suggests that only a fraction of the clicks in the buzz were detectable on the bottom-mounted hydrophone. Still, the buzz event, itself, was detected.

A total of 78 Md group dives containing 258 buzzes were identified within the 36 h manual analysis window. Buzzes were detected in 61 of the group dives. The BTP rules identified an average of 3.321 buzzes per foraging dive over all Md dives (high 19, low 0) and 4.246 buzzes per dive in the dives with buzzes (Table 3). This is a small fraction of the expected number (∼29) of buzzes produced per dive as reported from acoustic tag data, but buzzes were detected on 45 of 91 hydrophones across the entire range area during the study window. The most buzzes per dive were detected during group dives that occurred over hydrophones 1 to 7. Adjacent hydrophones within this hexagonal array, dubbed Whiskey-1, are only 2 km apart vice 4 km apart elsewhere on the range. Of the 14 Md group dives that had 5 + detected buzzes, six occurred over the Whiskey-1 array. Note that the CS-SVM classifier was not running on hydrophones 8–14 during the study period.

M3R SOAR archives collected over approximately 3.5 years from July 2014 through December 2017 (Table 1) were batch processed through the CTP and AG processes and Zc group dives identified as part of a separate study on the long-term spatio-temporal distribution and habitat use of Ziphius at SOAR (DiMarzio & Jarvis, 2016). These archives where subsequently processed through the BTP to isolate potential buzzes according to the following rules: minimum of 5 clicks per buzz, buzz duration less than 6 s long, 10 s timeout between buzzes. The times of candidate buzzes were then compared to the start and stop times of group dives output by AG. Only buzzes that occurred within a group dive were retained (Figure 5). ICI, as calculated from the CS-SVM report times for foraging clicks and buzz clicks can provide an interesting visualization of vocal part of a foraging dive cycle (Figure 6).

During the study period, 46817 Zc group dives and 97384 buzzes were detected, an average of 2.080 buzzes per dive. However, only 43.7% of the group dives contained buzz detections (as compared with 78.2% of dives at AUTEC). Within those dives, 4.758 buzzes were detected per dive (Table 3). The start times of the Zc group dives are fairly uniformly distributed across time. This agrees with the available tag data that undisturbed Zc forage in regular cycles both night and day (Schorr, et al., 2014; Falcone, et al., 2017). A slight diel pattern was (qualitatively) noted in the average number of group dives detected across the study period with somewhat fewer dives recorded during local night of 6 p.m. to 6 a.m. (0200-1400Z). The number of buzzes per dive followed a similar slightly diel pattern (Figure 7). This is not surprising as buzz detection is constrained to occur only during identified group dives. Only a small fraction of the expected number of buzzes (∼30 on average) produced per Zc foraging dive were detected, but buzz events were detected across the entire SOAR range area within the study period (Figure 8).