False killer whale and short-finned pilot whale acoustic identification

False killer whales Pseudorca crassidens and short-finned pilot whales Globicephala macrorhynchus are known to interact with long-line fishing gear in Hawaiian waters, causing economic loss and leading to whale injuries and deaths. The main Hawaiian Islands’ insular population of false killer whales is listed as endangered and the offshore population is considered ‘strategic’ under the Marine Mammal Protection Act due to relatively high bycatch levels. Discriminating between these species acoustically is problematic due to similarity in the spectral content of their echolocation clicks. We used passive acoustic monitoring along with data from satellite tags to distinguish signals from these 2 species. Acoustic encounters recorded with autonomous instruments offshore of the islands of Hawai‘i and Kaua‘i were matched with concurrent and nearby location information obtained from satellite tagged individuals. Two patterns of echolocation clicks were established for the 2 species. The overall spectral click parameters were highly similar (22 kHz peak and 25 kHz center frequency), but false killer whales had shorter duration and broader bandwidth clicks than short-finned pilot whales (225 μs, 8 kHz [−3 dB bandwidth] and 545 μs, 4 kHz, respectively). Also, short-finned pilot whale clicks showed distinct spectral peaks at 12 and 18 kHz. Automated classification techniques using Gaussian mixture models had a 6.5% median error rate. Based on these findings for echolocation clicks and prior published work on whistle classification, acoustic encounters of false killer whales and short-finned pilot whales on auto nomous instruments should be identifiable to species level, leading to better long-term monitoring with the goal of mitigating bycatch.


INTRODUCTION
False killer whales Pseudorca crassidens and shortfinned pilot whales Globicephala macrorhynchus are known to interact with pelagic long-line fisheries in Hawaiian waters, causing economic loss to the fishing industry and at least occasionally resulting in injury or death to the animals (Forney et al. 2011).Since both of these species are known to echolocate, passive acoustic monitoring techniques are being evaluated to assess the suitability of this method for distinguishing species, and potentially for bycatch mitigation.
Both species occur in largely overlapping distributions in tropical to warm temperate waters of all oceans, generally in more offshore and deep waters (Jefferson et al. 2008), where long-line fishing is conducted.The behavior, ecology, and social structure of these species have been studied most extensively around the Hawaiian archipelago (Baird 2009, Baird et al. 2010, 2012, 2013a,b, Bradford et al. 2014, Mahaffy et al. 2015).Behavioral and genetic evidence indicates 3 distinct populations of false killer whales in Hawaiian waters, with an insular main Hawaiian Islands stock, a pelagic, offshore stock, and a Northwestern Hawaiian Islands stock (Chivers et al. 2010, Baird et al. 2012, 2013a, Bradford et al. 2014, Martien et al. 2014).Fishery interactions and other threats have led to the listing of the main Hawaiian Island population of false killer whales as 'endangered' (November 2012 77FR 70915) under the US Endangered Species Act (ESA), and the pelagic stock as 'strategic' under the US Marine Mammal Protection Act (Bradford et al. 2014, Carretta et al. 2014).
A take reduction team was formed by the US National Marine Fisheries Service in 2010.One of its primary research goals is distinguishing false killer whale acoustic signals from those of other toothed whales (November 2012 77 FR 71259).Delphinids produce 2 general types of signals: (1) tonal frequency-modulated signals (whistles), and (2) broadband pulsed signals (clicks) (Herman & Tavolga 1980).Whistles are primarily used for communication, whereas echolocation clicks are used for detection, localization, and target classification in spatial orientation and foraging (e.g.Herman & Tavolga 1980, Au 1993).Field recordings of false killer whale and short-finned pilot whale whistles have been successfully discriminated, and automated classification routines were developed (Rendell et al. 1999, Oswald et al. 2003, 2007).However, misclassification most often occurred with whistles originating from shortfinned pilot whales.The characteristics of false killer whale echolocation clicks have been described (e.g.Au et al. 1995, Madsen et al. 2004, Ibsen et al. 2011) and false killer whale echolocation capabilities have mainly been studied with captive animals (e.g.Thomas & Turl 1990, Brill et al. 1992, Yuen et al. 2005, Kloepper et al. 2012, Madsen et al. 2013).The overall spectral content of their clicks seems to be generally lower than what is known for most other echolocating delphinids, having considerable energy below 20 kHz.While the echolocation strategy of short-finned pilot whales during deep dives is known to consist of foraging sprints accompanied by clicking, buzzing, and communication signals (Aguilar Soto et al. 2008, Jensen et al. 2011), no spectral or temporal information on short-finned pilot whale echolocation clicks is available.Assuming a similar spectral content of echolocation clicks from short-finned pilot whales as described for long-finned pilot whales Globicephala melas (Eskesen et al. 2011), discrimination of shortfinned pilot whale clicks from false killer whale clicks may be problematic based on spectral content.However, the capability to determine species identity from echolocation clicks can enhance passive acoustic monitoring and real-time alert systems.This is particularly true during periods when animals are echolocating without whistling and when combining both vocal production types in the classification decision.
The goal of this study was to describe the differences in echolocation click characteristics between false killer whales and short-finned pilot whales and demonstrate reliable classification results for clicks of these 2 species.This was accomplished by reviewing long-term acoustic recordings taken offshore of the main Hawaiian Islands and comparing these to concurrent and nearby satellite tracking tagged false killer whales and short-finned pilot whales for specieslevel echolocation click identification.These clicks were then used to build classification models which can be applied to acoustic recordings without the necessity of ground-truthing tagged animals.

Acoustic recordings
Autonomous high-frequency acoustic recording packages (HARPs) were bottom-moored at 2 locations in the main Hawaiian Islands at water depths of 600 to 700 m, with the hydrophone positioned about 30 m above the seafloor.One recorder location was offshore of the west coast of the island of Hawai'i (19°3 4.9' N, 156°00.9'W) and the other was between the islands of Kaua'i and Ni'ihau (21°57.2'N, 159°53.2'W) (Fig. 1).There were a number of deployments from August 2007 to May 2010 with varying recording schedules (Table 1).The instruments were part of a larger study to monitor cetacean and anthropogenic sounds over long periods near the Hawaiian Islands.
HARPs were set to a sampling frequency of 200 kHz with 16-bit quantization.The recorders were equipped with an omni-directional sensor (ITC-1042, International Transducer Corporation), which had an approximately flat (± 2 dB) hydrophone sensitivity from 10 Hz to 100 kHz of −200 dB re V/μPa.The sensor was connected to a custom-built preamplifier board and bandpass filter.The preamplifiers were designed to flatten the frequency response of the ambient ocean noise, which provided greater gain at higher frequencies where ambient noise levels are lower and sound attenuation is higher (Wiggins & Hildebrand 2007).The calibrated system response was accounted for during analysis.

Satellite tagging and location analysis
Details of the satellite tags and tagging methods have been previously published (Baird et al. 2010(Baird et al. , 2012)), so they are only briefly reviewed here.Satellite tags containing an ARGOS-linked, Wildlife Computers (Redmond) location-only SPOT-5 transmitter were remotely deployed on both species using a pneumatic projector.Tags were attached with two 6.5 cm titanium attachment darts.Location data obtained from Argos were assessed for plausibility with the Douglas Argos-Filter v.7.06 using established parameters (Baird et al. 2010(Baird et al. , 2012)).Straight-line distances between filtered whale locations and the HARP locations were determined for periods when HARP acoustic and whale location data overlapped, using the Posdist macro in Microsoft Excel (www.afsc.noaa.gov/nmml/software/excelgeo.php).Consecutive tagged whale locations that spanned the HARP location (e.g.north and south of the HARP) were used to assess periods when whale movements past the HARP may have occurred and resulted in acoustic recordings of that whale or individuals of its group.While we cannot be certain how many animals were within the detection range of our recorder, the number of individuals recorded during each acoustic encounter is expected to be more than just the single tagged animal.Group sizes of false killer whale and short-finned pilot whales around the main Hawaiian Islands average 15.6 ± 8.8 and 20.0 ± 15.6 individuals, respectively (Baird et al. 2013b).

Acoustic data analysis and satellite tag location match
Signal processing was performed using the MAT-LAB (Mathworks) based custom software program Triton (Wiggins & Hildebrand 2007) and other MAT-LAB custom routines.Long-term spectral averages (LTSAs) were calculated for visual analysis of the long-term recordings.LTSAs are long-term spectrograms with each time segment consisting of an average of 500 spectra, which were created using the Welch algorithm (Welch 1967).The averages were formed from the power spectral densities of nonoverlapped 10 ms Hann-windowed frames.The resulting long-term spectrograms have a resolution of 100 Hz in frequency and 5 s in time.This averaging technique provides visual detectability of echo location activity in a 5 s bin containing as few as 1 or 2 clicks.Trained analysts (A.E.S., S.B.P.) manually screened 1 h windows of LTSAs and identified echo location acoustic encounters in the HARP data when echolocation click energy fell below 20 kHz.This criterion was chosen based on towed hydrophone recordings in the presence of short-finned pilot whales in which echolocation click energy was below 20 kHz (authors' unpubl.data).Images of these LTSAs were saved, and start and end times of acoustic encounters were noted.
For each acoustic encounter, it was noted if satellite tag locations were reported within 5 nautical miles and 3 h prior or subsequent to the encounter.Based on the time difference between the tag location and the start or end of the acoustic recording, these encounters were labeled as potential encounters of the tagged species with low (1 to 3 h) or high (0 to 1 h) confidence (and will be shown to be similar  1. Deployment periods of high-frequency acoustic recording packages (HARPs) at sites offshore of Hawai'i (19°34.9'N, 156°00.9'W) and Kaua'i (21°57.2'N, 159°53.2'W) with overlapping periods of satellite tag recording.
Recording start and end dates and times given as mm/dd/yy, GMT to one another via classification experiment).Only echolocation clicks from high confidence acoustic encounters were used to compute echolocation click parameters.
Individual echolocation signals within these selected encounters were automatically detected using a 2 step approach computer algorithm (Soldevilla et al. 2008, Roch et al. 2011).The individual click detections were digitally filtered with a 10 pole Butterworth band-pass filter with a pass-band between 5 and 95 kHz.Filtering was done on 800 sample points centered on the echolocation signal.Spectra of each detected signal were calculated using 2.56 ms (512 samples) of Hann-windowed data centered on the signal.The frequency-related signal parameters peak frequency, center frequency, and bandwidth (−3 and −10 dB) were processed using methods from Au (1993).Click duration was derived from the de tector output.All detected echolocation signals, independent of distance and orientation of the recorded animal with respect to the recorder, were included in the analysis.Mean spectra were computed for all clicks across all acoustic encounters for each species.Mean noise spectra were also estimated from 2.56 ms of the 5 to 95 kHz bandpass-filtered acoustic signal (512 samples) ending 1.3 ms before the signal, similar to the echolocation click.

Classification
The classification methodology, which uses Gaussian mixture models (GMMs), is similar to that reported in Roch et al. (2011), but was modified to increase the variability of training data to better measure the susceptibility of the classifier to overtraining.Features consisted of duration and 14 cepstral coefficients that provide a compact representation of the spectra (Roch et al. 2011) extracted from the spectra of each echolocation click between 10 and 92 kHz.
Subsets of data from the high confidence acoustic encounters were used to train 16 mixture GMMs.For each species, the acoustic encounters were randomly permuted and distributed as evenly as possible into 3 groups for a 3-fold cross-validation experiment.For each of the 3 trials, the 2 folds used for training were further reduced by a random draw with replacement (a click could be selected more than once) of 85% of the echolocation clicks.The encounters from the remaining fold were then classified using the shortfinned pilot and false killer whale models.For each encounter in the test fold, joint probabilities were computed for each of the 2 models for groups of 100 echolocation clicks, making the simplifying assumption that echolocation clicks are independent.Click groups were assigned to species based on the Bayes decision rule (the model that produced the maximum probability).This process was repeated 100 times, resulting in 300 individual experiments.The low confidence acoustic encounters were classified using the 300 models trained from subsets of the high confidence models.

RESULTS
Satellite-tagged false killer whales and shortfinned pilot whales utilized the areas within the vicinity of the HARPs on a number of occasions and were matched with acoustic encounters containing click energy below 20 kHz (Tables 2 & 3, Figs. 2 & 3).Satellite-derived positions from 10 individual false killer whales and 15 individual short-finned pilot whales were matched with acoustic encounters.High and low confidence matches were achieved for 7 and 6 acoustic encounters with false killer whale locations (Table 2) and 15 and 3 acoustic encounters with short-finned pilot whale locations (Table 3), respectively.
The spectral echolocation click parameters peak and center frequency were highly similar when comparing clicks extracted from all high confidence acoustic encounters of both species (Table 4).False killer whale echolocation clicks had a much broader bandwidth but shorter duration clicks, resulting in a similar time−bandwidth product (a metric of signal processing gain).The LTSAs and mean spectra of acoustic encounters that were matched with shortfinned pilot whale showed spectral banding with peaks at 12 and 18 kHz (Fig. 4A,B).These spectral peaks were absent in encounters matched with false killer whales.When examining individual echolocation clicks of both species, presumable off-axis echolocation clicks of short-finned pilot whales appeared with more cycles in the waveform and with a slight frequency modulation (Fig. 4D).GMM classification of echolocation clicks from high confidence acoustic encounters resulted in a 6.5% median (mean 7.9 ± 8.0%) error rate across the 300 randomized trials.Performance on individual species in an accept/reject task had a median classification error of 2.8% (mean 7.5 ± 14.7%) for false killer whales, which was slightly lower than the error for short-finned pilot whales (median 3.9%, mean 7.6 ± 10.2%) (Fig. 5A).Classification of low confidence acoustic encounters using the models trained on high Index Before encounter

DISCUSSION
Acoustic recordings from bottom-moored auto nomous instruments in combination with locations of satellite tagged individuals from 2 odontocete species in the vicinity of the recorder have proven to be a useful method for gathering species-specific recordings for descriptions of their acoustic signals.Uncertainty remains, however, whether the recorded signals truly belong to the tagged individual(s) and their conspecifics or rather to other species concurrently in the detection range of the recorder.The 2 species rarely associate in Hawaiian waters; in 50 sightings of false killer whales around the main Hawaiian Islands, short-finned pilot whales were only present during 2 sightings (4%), and spatial overlap between the 2 species was brief in both sightings (R. W. Baird unpubl.data).Regardless, the consistency of the acoustic results in regards to spectral and temporal patterns as well as stable classification results support the method and results presented here.While ship-based recordings in the presence of animals are often used to describe the echolocation clicks of free-ranging animals, there are drawbacks to this method.Recordings from shallow depth hydrophones collected during ship-based acoustic surveys often have considerable noise originating from the sea surface and from the survey vessels, including acoustic sources such as breaking waves, vessel propulsion, machinery sounds, and electronic noises originating from onboard power generation or other sources.With deep water auto nomous recordings, these noise sources are minimized.Also, introduced noise in recordings can be highly confounding when initially defining these species-specific click characteristics.Classifiable differences in echolocation clicks are, at times, only small differences in spectral peaks or small shifts in overall spectral energy distribution, so low noise recordings are needed for reliable classifications (Soldevilla et al. 2008, Baumann-Pickering et al. 2010, 2013a).The data presented here (all clicks collected during acoustic encounters) are likely not comparable to other acoustic studies that focus on the presentation of features from on-axis clicks.Spectral and temporal parameters of echolocation clicks vary largely with the orientation of the animals' head relative to the hydrophone (Au et al. 1995(Au et al. , 2012a,b),b), their distance to it (Urick 1983), their individual hearing range (Kloepper et al. 2010), the behavioral state of the animals (Kloepper et al. 2012), as well as potentially other aspects such as the number of animals being acoustically active or the acoustic environment (Baumann-Pickering et al. in press).Given the median and percentile values in Table 4 and in comparison to published on-axis click values (Au et al. 1995, Kloepper et al. 2010, 2012), one can assume that the majority of clicks described in this article were recorded off-axis.For the purpose of passive acoustic monitoring using a single hydro phone, these intricacies of angle, distance, and be havior cannot easily be distinguished.Hence, stable features ob served across recorded signals and a description of the observed variability are of interest from free-ranging species in their natural habitat.
Click classification models should be able to incorporate variability in click characters without detailed knowledge of its cause and still be able to classify with adequate precision.Species-specific peaks in the mean spectra and LTSAs of short-finned pilot whale acoustic encounters are similar to species-specific peaks reported for Pacific white-sided Lagenorhynchus obliquidens and Risso's Grampus griseus dolphins off the coast of California (Soldevilla et al. 2008), as well as for several beaked whale species (Baumann-Pickering et al. 2013a,b).These click characteristics make acoustic encounters reliably identifiable to species level.Soldevilla et al. (2008) also point out that echolocation clicks with longer duration, presumably recorded off the central axis, show stronger spectral peaks for Pacific white-sided and Risso's dolphins.Hence, these spectral peaks may occur due to anatomical structures that shape off-axis echolocation signals.When spectral peaks are not an available parameter, as is the case for false killer whales, then species identification becomes more challenging (e.g.Baumann-Pickering et al. 2010, Roch et al. 2011).Another species with overlapping geographic distribution to false killer whales that may have similar broadband clicks without distinct spectral peaks could be roughtoothed dolphins Steno bredanensis, based on recent work by Rankin et al. (2015).Future work should investigate classification success of encounters including rough-toothed dolphins.GMM classification with low error rates for false killer whale and short-finned pilot whale echoloca-tion clicks confirms the possibility of distinguishing between the signals of these species.Error rates were even lower when testing on the low confidence acoustic encounters.This further strengthens our result that the 2 observed patterns of echolocation clicks are indeed classifiable with high accuracy.The next steps for optimizing and further understanding these classification results would be to more closely investigate those test tokens (groups of 100 echolocation clicks) that had consistent errors (Fig. 5II) to establish exactly what made these clicks different from those in the correctly classified tokens.Adding to the classification accuracy would be details on how the click signals attenuate with frequency over distance and how the signal's beam pattern, along with animal heading, affects the recorded signal.Using literature values collected from a small number of individuals in captivity can be useful; however, it may be possible to collect such data in the field by deploying sensors in an array that can track animals acoustically (e.g.Wiggins et al. 2012).Furthermore, such an experiment would ideally include acoustic archival-tagged animals swimming through the array, providing fine-scale spatial and temporal swimming and acoustic behavior.
Based on our results on the classification of echolocation clicks and published work on whistle discrimination, passive acoustic monitoring of false killer whales and short-finned pilot whales has the potential to be reliable with high accuracy.This opens up new methods for management and bycatch mitigation by investigating long-term trends in behavior and ecology, in addition to including passive acoustic recorders on fishing gear or hydro phones attached to long-line fishing vessels to identify behavioral patterns, confirm species of bycaught animals or implement real-time alarm systems.
Acknowledgements.The authors thank 2 anonymous reviewers for their insightful comments.We also thank John A. Hildebrand, Erin O'Neill, Brent Hurley, Chris Garsha, and Rachel Gottlieb for instrument preparation, data collection, and analysis assistance.HARP deployments and recovery off the island of Hawai'i were undertaken with the assistance of Greg Schorr and Daniel Webster.
Date (mm/dd/yy) and time (GMT) of acoustic encounter and matching satellite tag locations for false killer whales Pseudorca crassidens (Pc) within 3 h before and after the encounter.Indices are separated into high (H) and low (L) confidence based on distance and time delay of satellite tag detections in relation to highfrequency acoustic recording package (HARP) acoustic encounter, or possible false classification due to interfering short-finned pilot whale

Fig. 5 .
Fig. 5. Classification results from (A) training and testing on high confidence acoustic encounters and (B) testing on low confidence acoustic encounters using models constructed from high confidence data.(I) Histograms show combined (left) classification results (3 folds, 100 models), and species-specific results: Pc: false killer whale Pseudorca crassidens (middle), Gm: short finned pilot whale Globicephala macrorhynchus (right).(II) Performance of Pc (top) and Gm (bottom) test token groups of 100 echolocation clicks grouped by (A) high and (B) low confidence acoustic encounters.Grey shading in the scale bar represents the per token error rate over the 100 different models that were trained with different training data Satellite tags were deployed by Greg Schorr, Daniel Webster and Brad Hanson.Tagging was undertaken under NMFS Scientific Research Permit No. 731-1774 issued to R.W.B. Funding for this project was provided by the NOAA Pacific Islands Fisheries Science Center and Northwest Fisheries Science Center, the US Navy (through the Living Marine Resources Program, Naval Postgraduate School, National Oceanographic Partnership Program, and Office of Naval Research), the Bureau of Ocean Energy Management, and the John F. Long Foundation.Editorial responsibility: Brendan Godley, University of Exeter, Cornwall Campus, UK Submitted: October 10, 2014; Accepted: June 16, 2015 Proofs received from author(s): July 7, 2015