ABSTRACTS ARE IN THE ORDER OF PRESENTATION

Transcription

1 Fifth International Workshop on Detection, Classification, Localization, and Density Estimation of Marine Mammals using Passive Acoustics August, 2011 Timberline Lodge, Mount Hood, Oregon, USA ABSTRACTS The abstracts for oral presentations and posters at the 5th International Workshop on Detection, Classification, Localization, and Density Estimation of Marine Mammals using Passive Acoustics, predominantly focus on odontocete sounds and analytical methods for classifying clicks and whistles, as well as density estimation. Research on baleen whale sounds and some non-cetacean marine mammals is also featured, and provides further information about important methodologies. These biennial DCLDE workshops are intended for exchanging information that advances understanding of acoustic methods to detect, classify, locate, track, count, and monitor marine mammals in their natural environment. The goal is to encourage interdisciplinary approaches to solve real-world problems related to the study of marine mammals and the effects of human activities on their behavior. ABSTRACTS ARE IN THE ORDER OF PRESENTATION Alphabetical Index of authors begins on page 86

2 Monday 22-Aug-11 08:40-09:00 Simard Automatic detection and classification of wild beluga whale clicks: performance of a click detecting instrument under various conditions recorded at sea Yvan Simard 1,2, Cédric Gervaise 3, Nathalie Roy 2 (1) Fisheries and Oceans Canada chair in acoustics applied to ecosystem and marine mammals, Marine Sciences Institute (ISMER), University of Quebec at Rimouski, 310 Allée des Ursulines, Rimouski, Québec G5L 3A1, Canada; Yvan.Simard@dfo-mpo.gc.ca (2) Fisheries and Oceans Canada, Maurice Lamontagne Institute, 850 route de la Mer, P.O. Box 1000, Mont-Joli, Québec G5H 3Z4, Canada (3) ENSTA Bretagne, 2 rue François Verny, Brest, France Toothed whales are equipped with efficient biosonars they regularly use for echolocating and sensing objects, preys, predators and navigating in their underwater environment. The recording and analysis of these clicks can reveal the presence and behaviours of the whales in the detection range of PAM (passive acoustic monitoring) devices deployed in their habitats. The detection of the acoustic events of interest can be done at the acquisition, by features embedded in the electronic circuitry of the click detector, and the event characteristics are stored for eventual post-processing classification. This approach was implemented in the C-Pod click detectors which are tested here for their efficiency in detecting and recognising beluga clicks recorded in the St. Lawrence estuary and for their robustness against potential interferences from echosounder pulses and other transient signals. Three co-located C-Pods and a reference hydrophone were deployed from a ship at 6 locations in St. Lawrence beluga habitat in Oct. 2010, in presence and absence of belugas and echosounder clicks, under diverse sea state conditions. Using settings for beluga click detection and post-processing classification, C-Pods false positive and false negative rates were high. Acoustic events detected by C-Pods were much higher than the actual beluga clicks or acoustic transients recorded by the hydrophone, but were often timed with them. The C- Pod post-acquisition filters were not able to reject transients misclassified as clicks. The general performance of the instrument for the tested task is discussed. Improvements of the classification performance of the detected events using new algorithms exploiting their stored attributes are explored. 1 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

3 Monday 22-Aug-11 09:00-09:20 Yack Automated detection and classification of Dall s porpoise in the California Current ecosystem Tina M. Yack 1,2,3, Shannon N. Coates 2, Jay Barlow 1 (1) NMFS/NOAA Southwest Fisheries Science Center, 3333 N. Torrey Pines Ct., La Jolla, CA USA; Tina.Yack@noaa.gov (2) Joint Doctoral Program in Ecology, San Diego State University, San Diego, CA USA (3) Biowaves Inc, 517 Cornish Dr., Encinitas, CA USA Dalls porpoise (Phocoenoides dalli) range throughout the North Pacific and adjacent seas from 32 to 35 N (southern California and southern Japan) in the south to ~63 N (central Bering Sea) in the north. In the California Current their range overlaps with harbor porpoise (Phocoena phocoena). Dall s porpoise can be difficult to sight in high sea states which leads to difficulties in obtaining accurate population estimates. Dall s porpoise produce high-frequency, narrow-band (HFNB) echolocation clicks which are easily distinguished from other odontocetes. Technological advances now allow for continuous recording at frequencies including HFNB signals, providing an opportunity to use passive acoustic monitoring of these species during shipboard line-transect population surveys. During a 2008 Southwest Fisheries Science Center survey of the California Current, Rainbow Click ( software was used for detection and classification of porpoise echolocation signals. The survey resulted in 54 visual-only detections, 31 combined visual/acoustic detections, and at least 90 acoustic-only detections. Preliminary results suggest acoustic detections will significantly increase the sample size of Dall s porpoise detections compared to previous SWFSC cetacean assessment cruises. We describe and characterize Dall s porpoise echolocation clicks from recordings of eight single-species encounters Measurement of 5,560 echolocation clicks showed a bimodal split in the peak-frequency with median peak frequencies at 120 khz and 130 khz. Click types were segregated into two categories based on peak-frequency, with a split at 126 khz. This analysis will allow for development of a more robust click detector and discrimination of clicks of Dalls porpoise from harbor porpoise in areas of sympatry. 2 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

4 Monday 22-Aug-11 09:20-09:40 Oswald Automatic detection and classification of bowhead and beluga whale calls in the Eastern Chukchi Sea Julie N. Oswald 1, Xavier Mouy 1, David K. Mellinger 2, Del Leary 3, Bruce Martin 3, David Hannay 1 (1) JASCO Applied Sciences, Suite Markham Street, Victoria, British Columbia V8Z 7X8 Canada; oswald.jn@gmail.com. (2) Cooperative Institute for Marine Resources Studies, Oregon State University, 2030 SE Marine Science Dr., Newport, OR USA (3) JASCO Applied Sciences, Suite Troop Avenue, Halifax, Nova Scotia B3B 1Z1 Canada. A large scale passive acoustic monitoring program was conducted in the Eastern Chukchi Sea from 2007 to A goal of this program was to monitor migration distributions of bowhead and beluga whales. Automated systems for detecting calls in the acoustic recordings had to be developed due to the large datasets collected. The bowhead acoustic repertoire included low frequency moans (< 1000 Hz) produced during the summer season and higher frequency, more complex songs produced in the fall and early winter. Beluga sounds recorded included whistles in the 1-8 khz frequency band. This paper investigates an automated method for detecting and classifying bowhead moans, bowhead songs and beluga whistles in the Eastern Chukchi Sea. The Chukchi Sea is a challenging acoustic environment; ice noise occurs in the winter, anthropogenic noise is present during the summer, and other marine mammal (e.g. bearded seal, walrus) calls occur in the same frequency bands as bowhead and beluga calls. These factors present challenges for automated detection and classification systems. The method investigated here was as follows. Normalized spectrograms were calculated using a split-window normalizer. Vocalization events were detected in the normalized spectrograms and their time-frequency contours were extracted. Each contour was represented by 48 features and presented to two-class random forest classifiers (e.g. bowhead vs. other, beluga vs. other ). Detection and classification of bowhead and beluga sounds were performed separately. The algorithms were applied to a large set of manually annotated vocalizations. The average Precision and Recall values were 73.9 (sd=0.5) and 74.4 (sd=2.8) for bowhead song, 77.2 (sd=2.9) and 77.6 (sd=3.5) for bowhead moans, and 84.1 (sd=1.4) and 85.7 ( sd=2.8) for beluga whistles. Description of results for different signal to noise ratios and applicability for a real-world monitoring scenario will be discussed. 3 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

5 Monday 22-Aug-11 09:40-10:00 Birleanu Recurrence plot analysis (RPA): a new tool for click detection and clustering clicks into trains Florin-Marian Birleanu 1, Cornel Ioana 1, Cédric Gervaise 2, Yvan Simard 3,4 (1) GIPSA-lab, Department Image-Signal, 961 rue de la Houille Blanche, Saint Martin d Heres, France (2) ENSTA Bretagne, 2 rue François Verny, Brest, France; cgervaise50@gmail.com (3) Fisheries and Oceans Canada chair in acoustics applied to ecosystem and marine mammals, Marine Sciences Institute (ISMER), University of Quebec at Rimouski, 310 Allée des Ursulines, Rimouski, Québec G5L 3A1, Canada (4) Fisheries and Oceans Canada, Maurice Lamontagne Institute, 850 route de la Mer, P.O. Box 1000, Mont-Joli, Québec G5H 3Z4, Canada A new algorithm to detect and classify short transients such as clicks is studied. In cases o f multiple simultaneous emissions, such preliminary step of recognizing the clicks and the train they belong to is a prerequisite step to source localization and census. In a recording time segment, m(t), containing several click trains, our algorithm is aimed to detect the time interval corresponding to the click and associate similar clicks to a same cluster or train. Our classifier relies on RPA, a n ew way of representing transients. First, m(t) is translated in a multi-dimensional space by forming the trajectory of vector v(t)={m(t),m(t+τ), m(t+(n-1) τ) where n a nd τ are some RPA parameters. Second, recurrences in the trajectory v(t) are looked for with the distribution of distance (D(i,j)) between v(t i ) and v(t j ) when t i and t j span the initial recording time segment. The map D(i,j) is the Recurrent Plot that can be used for Analyses to detect repetitive patterns associated with a click train. Our RPA algorithm is then applied on synthetic and real data implying several simultaneously clicking bottlenose dolphins in Cotentin, France, and beluga whales in Saguenay fjord mouth, Québec, Canada. 4 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

6 Monday 22-Aug-11 10:00-10:20 Azzara Assessment of large vessel noise impacts on sperm whales in the Gulf of Mexico Alyson J. Azzara 1, Wyndyln M. von Zharen 2 (1) Department of Marine Biology, Texas A&M University at Galveston, P.O. Box 1657 Galveston, TX USA; aly.azzara@gmail.com (2) Department of Marine Sciences, Texas A&M University at Galveston, P.O. Box 1657 Galveston, TX USA There is currently a p aucity of regulations for vessel noise, particularly for large shipping vessels, although it is the number one contributor of noise to the marine environment. It has already been shown that vessel noise impacts the vocalizations of both humpback and killer whales. The potential exists for it to also impact sperm whales. To determine this potential impact, this research uses archival, bottom moored, acoustic buoy data to assess the effects of shipping noise on sperm whales in the Gulf of Mexico to improve understanding of modifications in marine-mammals behavior. The project use data collected by the Naval Research Lab s (NRL) EARS buoys through the Office of Naval Research (ONR) project, Littoral Acoustic Demonstration Center (LADC) from 2001 a nd T hree buoys recorded a single channel continuously and autonomously at a sampling rate of 11.7 khz. Buoys were deployed 50 meters from the bottom at depths from 645m to 1034m. These buoys recorded for 36 and 55 consecutive days during the summer months of 2001 a nd Presentation includes preliminary results on calculated decibel shift during ship passage as well as both the received levels at peak noise and the sound exposure levels at the buoys for the duration of the ship passage. The resident sperm whale population in the Gulf of Mexico is primarily found along the Louisiana- Mississippi shelf directly within the path of the New Orleans shipping lanes and the location of the buoys. By comparing the number of clicks before, during and after shipping events, it may be possible to determine whether vessel passage has an impact on the click behavior of sperm whales. If there is a significant difference, it could indicate changes in behavior likely linked with anthropogenic impacts. A ssessing potential impacts is imperative for future stock management and noise regulation. 5 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

7 Monday 22-Aug-11 10:40-11:00 Moretti The detection of Cuvier s beaked whales (Ziphius cavirostris) across ocean basins David Moretti 1, Mark Johnson 2, Susan Jarvis 1, Ronald Morrissey 1, Jessica Ward 1, Charles White 1 (1) Naval Undersea Warfare Center (NUWC), Newport, RI 02841; morettidj@hotmail.com (2) Woods Hole Oceanographic Institution, 86 Water St, Woods Hole, MA Cuvier s beaked whales (Ziphius cavirostris) have been recorded in multiple locations around the globe. The signal structure of Ziphius clicks recorded in the Ligurian Sea, the Canary Islands, the Tongue of the Ocean (TOTO) in the Bahamas, and off San Clemente Island in the Pacific are compared. These clicks are then processed through FFT and linear matched filter based detectors and through a simple classifier which uses frequency segmentation and Inter-Click Interval (ICI) and a Support Vector Machine (SVM) classifier, to investigate the performance of multiple algorithms for the same species across widely diverse environments. These multiple environments present both a potential difference in signal structure along with markedly different non-gaussian noise backgrounds that include vocalizations from competing species and anthropogenic generated noise. Both the detector (probability of detection, and false alarm rates) and classifier performance across each data set are measured. 6 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

8 Monday 22-Aug-11 11:00-11:20 Gérard Analysis of various beaked whale recordings in eastern Atlantic Odile Gérard 1 (1) DGA Naval Systems, Avenue de la Tour Royale, BP 40915, Toulon cedex, F rance; odileg@free.fr Beaked whales are a group of more than 20 genetically confirmed species; they are very elusive and were among the least known species until few years ago. The strandings of beaked whales have been reported to occur in conjunction with sonar exercises, although the exact mechanism is not clear. These standings leaded to an increased research effort dedicated to these very sensitive species and in particular to their signals. Despite this effort, the signals of many beaked whale species remain unknown. Signals of Blainville s (Mesoplodon densirostris), Cuvier s (Ziphius cavirostris), and Gervais (Mesoplodon europaeus) have been studied (Johnson et al. 2004; Zimmer et al. 2005; Gillepsie et al. 2009). The signals of two unknown species have been reported (Mcdonald et al. 2009; Baumann-Pickering et al. 2010), it is thought that these signals were produced by two different species of beaked whale other than Blainville s, Cuvier s or Gervais. All five species produce upsweep frequency modulated signals which seem species specific. Previous recordings of other beaked whale species have been reported, but restrictions in sampling frequency do not allow determining if the signals have a frequency sweep. In 2010, N ATO Undersea Research Centre (NURC) conducted a sea-trial in Eastern Atlantic Ocean, Southwest of Portugal. Various beaked whale signals were recorded; the characteristics of these signals will be presented and compared to the known signals of beaked whales [1-5]. 7 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

9 Monday 22-Aug-11 11:20-11:40 Ward Mesoplodon densirostris click characteristics for passive acoustic monitoring Jessica Ward 1, Nancy DiMarzio 1, David Moretti 1, Ronald Morrissey 1, Mark Johnson 2, Peter Tyack 2 (1) Naval Undersea Warfare Center Division, 1176 Howell St., Newport, Rhode Island 02841; jessica.ward@navy.mil (2) Woods Hole Oceanographic Institution, Woods Hole, MA Beaked whales (Cetacea: Odontoceti: Ziphiidae) as a group are broadly distributed, but due to their elusive behavior and deepwater habitat, only recently have the characteristics of their sound production been documented. Analysis of sound and orientation data from Digital Acoustic Recording Tags (DTag) has contributed significantly to the understanding of beaked whale behavior and acoustic characteristics. Beaked whales have some ideal characteristics for passive acoustic monitoring including long dive times with consistent clicking, unique click time/frequency characteristics, and a small repertoire of click types. During the 2007 Behavioral Response Study in the Tongue of the Ocean, Bahamas, DTag s were placed on Mesoplodon densirostris and synchronous passive acoustic recordings were made from an 82 hydrophone, broad-band deep water array. These combined data were used to estimate M. densirostris acoustic transmission characteristics including beam pattern. The results from three tagged M. densirostris will be compared and contrasted with discussion of the effects on PAM. 8 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

10 Monday 22-Aug-11 11:40-12:00 DiMarzio Temporal and spatial distribution of Blainville s beaked whales using passive acoustic detection data Nancy A. DiMarzio 1, Len Thomas 2, Ian L. Boyd 3, Ronald P. Morrissey 1, David J. Moretti 1, Jessica A. Ward 1, Ashley N. Dilley 1, Elena M. McCarthy 1, Susan M. Jarvis 1 (1) Naval Undersea Warfare Center Division, 1176 H owell St., Newport, Rhode Island 02841, USA; nancy.dimarzio@navy.mil (2) Center for Research into Ecological and Environmental Modelling, University of St. Andrews, St. Andrews, KY16 9LZ, Scotland (3) Sea Mammal Research Unit, University of St. Andrews, St. Andrews, KY16 9LZ, Scotland CONTACT FIRST AUTHOR FOR ABSTRACT 9 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

11 Monday 22-Aug-11 12:00-12:20 Abeille Robust bi-density automatic determination of the inter-pulse interval of the Physeter macrocephalus click Régis Abeille 1, Pascale Giraudet 1, Hervé Glotin 1 (1) DYNI-LSIS, Université du Sud Toulon Var, La Garde 83130, France; glotin@univ-tln.fr State of the art provides efficient click extraction system, and correct localization methods. However an important challenge remains : the individual identification for each click and for each specimen which is recorded. Many studies (Gordon, 1991)consider that the Inter-Pulse Interval (IPI) is the key to identify each animal. State of the art (Antunes, 2010) is based on auto-correlation, cepstrum and manual measurement to extract IPI, but recent author conclusion was that manual measurement is the most robust and is still required (Antunes, 2010). This issue is related to the click structure which depends on the orientation between sperm whales and the hydrophone. In most cases, the right pulse identification becomes a very hard task and finally the IPI cannot be extracted. To tackle this matter, we have developed a new algorithm which allows to determine the sperm whale size from a whole set of detected clicks, without any preliminary click filtering. It is based on bihistogram density estimations. Our algorithm labels the pulses present in each click. The first labeled pulse is called P0, the second P1 and the third P2. Then we compute the delay between each pulse. The bi or three density estimation of all these delay are showing particular properties that we use to estimate reliable IPI. This method is robust to echo, pulse miss-detection and different off-axis. Our evaluation are conducted on test data from 2005 DCL workshop, from the BOUSSOLE experiment (see submitted abstract Monnin et al.), and from LSIS recordings made in front of Toulon (see submitted abstract Glotin et al.). Some manual validations have been made which show the efficiency of our algorithm. Our method allows sperm whale identification and open ways for large base statistical studies on whale populations. 10 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

12 Monday 22-Aug-11 13:40-14:00 Zimmer Detection and classification of delphinids using clicks and whistles Walter M.X. Zimmer 1 (1) NATO Undersea Research Centre, Vle San Bartolomeo 400, I La Spezia, Italy; wzimmer@whoi.edu The focus of the 2011 DCL workshop is on detection and classification of odontocete sounds, including both whistles and clicks. While previous workshops concentrated on d etection, classification and localization of short, transient type cetacean sounds, the explicit extension to tonal sounds constitutes not only a novum, but also a challenge. The work to be presented is based on the dataset provided by the organizers for this workshop and picks up the implicit challenge of the workshop. The methods that are used for detection and classification of clicks and whistles follow closely and extend the algorithms in the recently published textbook on Passive Acoustic Monitoring of Cetaceans. Emphasis is hereby placed on an integrated classification scheme combining features that stem from both click and whistle detectors, which are implemented in a straight forward way by means of standard algorithms. The classification scheme follows the basic multidimensional cluster analysis and is therefore easy to implement. As with all classification schemes the proper choice of the feature vector is key to a successful solution. The presented work discusses the basic concept of the detection and classification algorithms and analyses the influence of the selected feature vectors on the performance of the combined click and whistle classification scheme. Testing the algorithms on the dataset provided by the workshop organizer will facilitate discussion and will further allow a direct comparison with approaches and solutions presented by other workshop participants. All algorithms are implemented in Matlab and the scripts will be made public available after the workshop. 11 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

13 Monday 22-Aug-11 14:00-14:20 Frasier Odontocete species identification by analysis of whistle component shape and sequence Kaitlin E. Frasier 1, E. Elizabeth Henderson 1, Marie A. Roch 2, Hannah R. Bassett 1, John A. Hildebrand 1 (1) Scripps Institution of Oceanography, University of California, San Diego, 9500 Gilman Drive, La Jolla, California ; kefrasie@ucsd.edu (2) San Diego State University, Dept. of Computer Science, 5500 Campanile Dr, San Diego, California, Most automatic classification systems for odontocete whistles are based on statistics describing the mean characteristics of calls. In this work we model whistles using Hidden Markov models (HMMs), drawing on methods used in speech recognition. Whistles are modeled as a series of components defined by easily detected phenomena such as inflection points. Each component is represented as a sequence of states, using state dependent probability distributions. A bigram model is created to predict the probability of seeing one type of component having just seen another. Models are created for each species using training data. Sequences that did not occur in the training data are given non-zero probabilities using backoff strategies (e.g. Katz backoff). In this way, component models can be sequenced to recognize whistles that may be characteristic of a species, but that did not occur in the training data. For testing, the probability that an unknown whistle was produced by one of a set of known species is computed using a search strategy that takes into account both the acoustic information (the component HMMs) and the component sequencing information (bigram model). The system is evaluated on the DCL conference data set, using about an hour of calls from five different species: bottlenose dolphins (Tursiops truncatus), common dolphins (Delphinus sp.), melon-headed whales (Peponocephala electra), and spinner dolphins (Stenella longirostris longirostris) collected in the Southern California Bight and at Palmyra Atoll. R esults are reported both using analyst generated contours and automatically detected contours (Silbido contour extractor). 12 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

14 Monday 22-Aug-11 14:20-14:40 Johansson High-resolution whistle detection and frequency tracking using adaptive notch filters A. Torbjorn Johansson 1 (1) FOI (Swedish Defence Research Institute), Gullfossgatan 6, SE Stockholm, Sweden; torbjorn.johansson@foi.se. An adaptive filter-based method for detection and frequency estimation of whistles and other tonal calls is proposed. Such calls are used by e.g. birds and marine mammals and are typically analyzed in the timefrequency domain using a sp ectrogram. The approach taken here is instead based on adaptive notch filtering applied to the recording waveform. The choice of this approach is motivated by a desire to obtain a more accurate characterization of these whistles, hoping that this in turn will permit improved call-based classification of species and species groups. In spectrogram-based processing, data analysis is performed in short temporal windows, typically 256 or 512 samples long. The proposed method instead moves a temporal window of effective width samples along the recording waveform. The use of a shorter window permits a h igher temporal resolution, effectively making it possible to track faster frequency changes and extract a more detailed time-frequency contour of each whistle. Adaptive notch filtering is an established technique for tracking of non-stationary tonals, e.g. in telecommunications. This paper describes its adaptation to the analysis of tonal animal calls and the development of a method for whistle detection. It is important that a w histle detector can cope with whistles that occur at the same time as well as interfering clicks and other short-duration transients. We describe developments that aim at correct whistle detection and estimation even in such conditions. The performance of the proposed whistle detector is illustrated using data from the 5 th DCL workshop dataset. The results show that the proposed method is able to extract multiple simultaneous whistles, can cope with frequency crossings and interfering clicks, and appears to reports correct whistle start and end times. It is also able to extract long whistles into a single detection. 13 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

15 Monday 22-Aug-11 14:40-15:00 Hager Marine mammal monitoring during a Navy explosives training event off the coast of Virginia Beach, Virginia Carl A. Hager 1, Mandy Shoemaker 2, Anu Kumar 2 (1) Oceanography Department, United States Naval Academy, Chauvenet Hall, 572C Holloway Road, Annapolis, MD 21402, USA; hager@usna.edu (2) Environmental Conservation - Marine Resources, Naval Facilities Engineering Command (NAVFAC) Atlantic, 6506 Hampton Blvd, Norfolk, VA 23508, USA Navy explosives training events pose a risk to marine mammals and sea turtles. In order to minimize the potential impacts and in support of the monitoring plan for the Virginia Capes (VACAPES) Letter of Authorization under the MMPA, visual and acoustic surveys were conducted during explosive training events conducted in August of 2009, 2010 and Each year's monitoring plan consists of visual and acoustic surveys onboard the U.S. Naval Academy's 109 foot research vessel one day prior to, during and following the explosive training events. In 2009, visual surveys were conducted and a single hydrophone was deployed as a feasibility study and approximately 20 minutes of acoustic data was collected. I n 2010, visual and acoustic surveys were conducted which included the use of five SSQ53F sonobuoys modified for autonomous use. Approximately 20 hours of acoustic data was collected over three days. The August 2011 monitoring cruise will include visual surveys and will utilize six autonomous hydrophones able to record data locally with a sampling rate of 96 khz. Methods, preliminary results, and lessons learned will be presented. 14 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

16 Monday 22-Aug-11 15:00-15:20 Freitag Reducing mutual interference between marine mammals and acoustic modems Lee Freitag 1, James Partan 1, Sandipa Singh 1, Eric Gallimore 1 (1) Woods Hole Oceanographic Institution, 266 Woods Hole Road, MS 18, Woods Hole, MA, USA; lfreitag@whoi.edu An introduction to a means for reducing or avoiding interference between man-made acoustic devices such as acoustic modems and marine mammals is discussed. The ultimate goal is to create an acoustic communications device that has characteristics of what is termed cognitive radio, an expression used to describe radio-frequency systems that exploit un-used or under-used sections of spectrum. A cognitive modem, much like a cognitive radio, would be able to employ policies that determine allowable behaviors, along with detection capabilities that measure usage of particular bands, and to set operating modes that include carrier frequency, modulation and power. There are a number of enabling technologies that will make up a cognitive acoustic communications system able to reduce mutual interference to the benefit of both marine mammals and man-made underwater networks. One, methods for detection of marine mammals are necessary, and must be implemented in real-time and in parallel to normal receivers. Two, a means for making decisions on how to act in the presence of mammals, is required. Finally, while simple responses to detection of mammals include ceasing transmission or lowering power, the acoustic modem may modify frequencies of use, requiring very broad-band transducers. The goal of this presentation is provide a introduction to typical signaling characteristics of a coustic modem systems, their methods of use, their potential to become cognitive of the environment and what practical problems must be solved. As part of this presentation the latest version of the WHOI Micro- Modem and its capabilities will be described, along with specific thoughts on how such a system could be implemented by taking advantage of advances in passive marine mammal detection and classification algorithms. Specific examples of interference between marine mammals and acoustic communications systems in a number of environments will also be presented. 15 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

17 Monday 22-Aug-11 15:40-16:00 Thode Automated detection and localization of bowhead whale sounds in the presence of seismic airgun surveys, Aaron M. Thode 1, Katherine H. Kim 2, Susanna B. Blackwell 2, Charles R. Greene, Jr. 2, Christopher S. Nations 3, Trent M. McDonald 3, A. Michael Macrander 4 (1) Marine Physical Laboratory, Scripps Institution of Oceanography, University of California San Diego, La Jolla CA U.S.A.; athode@ucsd.edu (2) Greeneridge Sciences, Inc., 6160-C Wallace Becknell Road, Santa Barbara CA 93117, U.S.A. (3) Western EcoSystems Technology Inc., 2003 Central Ave., Cheyenne, WY 82001, U.S.A. (4) Shell Exploration and Production Co., 3601 C St. Suite 100, Anchorage AK 99503, U.S.A. The development of automated techniques for processing acoustic data collected by oil companies in the Arctic has been spurred by both cost concerns and by rapid turnaround requirements for regulatory reporting. O ne such procedure has been developed for detecting and localizing frequency-modulated bowhead whale sounds in the presence of seismic airgun surveys. The procedure has been applied to four years of data, collected from over 30 directional autonomous recording packages deployed over a 280 km span of continental shelf in the Alaskan Beaufort Sea. The procedure has six sequential stages that begin by extracting 25-element feature vectors from spectrograms of potential call candidates. Two cascaded neural networks then classify the feature vectors as bowhead calls or reject them. Detections classified as calls are subsequently matched between recorders to triangulate locations. To train the networks, manual analysts reviewed twelve non-consecutive days of data from 2008 and 2009, flagging 219,471 bowhead call examples. The manual analyses were also used to identify 1.17 million transient signals that were not whale calls. During training each network output threshold was set such that 10% of true whale calls in the training data were rejected (i.e. both networks rejected 20% of whale calls in the training data). The trained algorithm was then validated on data collected in 2007 and During those years automated analyses missed 30-40% of manually detected calls. In addition, 20-40% of the sounds flagged as calls by the algorithm are not present in the manual analyses. These extra detections incorporate some legitimate whale calls overlooked by human analysts. The automated procedure tends to reject high signal-to-noise ratio calls that generate incoherent reverberation, and the manual analyses tend to miss short, weak calls; however, both methods produce similar spatial and temporal call distributions. (Work supported by Greeneridge Sciences and the Shell Exploration and Production Company). 16 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

18 Monday 22-Aug-11 16:00-16:20 Mouy Automatic detection of underwater walrus sounds in the eastern Chukchi Sea Xavier Mouy 1, Julie N. Oswald 1, Del Leary 2, David Hannay 1, Bruce Martin 2, David K. Mellinger 3 (1) JASCO Applied Sciences, Suite Markham Street, Victoria, British Columbia V8Z 7X8 Canada; Xavier.Mouy@jasco.com (2) JASCO Applied Sciences, Suite Troop Avenue, Halifax, Nova Scotia B3B 1Z1 Canada. (3) Cooperative Institute for Marine Resources Studies, Oregon State University, 2030 SE Marine Science Dr., Newport, OR USA Pacific walruses (Odobenus rosmarus divergens) are loquacious animals able to produce a large variety of airborne and underwater sounds. Their underwater repertoire includes grunting sounds, knocks and belllike sounds. Grunts are the most common sounds recorded in the eastern Chukchi Sea and therefore the most relevant targets for passive acoustic monitoring studies. They consist of short ~0.2-second sounds below 1 khz, usually repeated in pairs. Principal difficulties in developing detection algorithms for such sounds are the presence of bowhead whale calls in the same frequency band, natural and anthropogenic low frequency noise, and distortions of the calls due to acoustic propagation effects in this area. Here we investigate a robust automatic algorithm for detecting walrus grunts. The algorithm first calculates the spectrogram and normalizes it for each frequency band using a split window normalizer. The spectrogram is then analyzed in consecutive 0.7-second frames overlapped by 50 percent. For each frame, 50 features representing salient characteristics of the spectrogram are extracted in the frequency band Hz. These features include (but are not limited to) spectral entropy, harmonicity, frequency distribution, and frequency and amplitude modulation indices. Extracted features are then presented to a two-class random forest classifier composed of 600 decision trees to determine the class of the sound in the analyzed frame (i.e., walrus grunt or other ). The algorithm was trained and optimized using acoustic data collected in the eastern Chukchi Sea during summer The optimized algorithm was then applied to similar data collected in the Chukchi during summer 2010 and its performance was quantified using the precision and recall metrics for various signal-to-noise ratios. Applicability of this algorithm to a real-world monitoring scenario is discussed. 17 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

19 Monday 22-Aug-11 16:20-16:40 Denes A comparison of marine mammal detections from co-located sub-sampling passive acoustic monitors Samuel L. Denes 1,2, Jennifer L. Miksis-Olds 1,2, Jeffrey Nystuen 3, David K. Mellinger 4 (1) Graduate Program in Acoustics, The Pennsylvania State University, 201 Applied Science Building, University Park, PA USA; sld980@psu.edu (2) Applied Research Laboratory, The Pennsylvania State University, PO Box 30, State College, PA USA (3) Applied Physics Lab, University of Washington, 1013 NE 40th Street, Box , Seattle, WA USA (4) Cooperative Institute for Marine Resources Studies, Oregon State University and NOAA Pacific Marine Environmental Laboratory, 2030 SE Marine Science Dr., Newport, OR USA Acoustic detection of marine mammal vocalizations from two passive acoustic monitors, AURAL and PAL (Passive Aquatic Listener), co-located in the Bering Sea were compared. Power and data storage limitations require sub-sampling during long-term passive acoustic monitoring. The AURAL is a continuous recorder that was set to a 30% duty-cycle and samples at 8 khz. The PAL is an adaptive subsampling recorder with a 100 khz sampling rate that collects 4.5 sec time series, or soundbites, at varying intervals based on feedback from onboard processing algorithms with a typical, but variable duty cycle of 1.5%. The number of soundbites was limited to a daily quota, but the PAL also records a quasi continuous spectral time history. We compared detections of marine mammal vocalizations for four days each month from a yearlong dataset of AURAL recordings and a combination of PAL spectra and soundbites. Identifications of marine mammal species with stereotyped or easily defined vocalizations, in particular, bowhead whale, bearded seal, ribbon seal, and Pacific walrus, from PAL spectral time history data were made by comparing spectral time history samples to vocalization templates from the PAL soundbites. Inherent strengths and weaknesses of each device were readily apparent. Detections from the higher duty-cycle but lower sample rate AURAL were limited to species and vocalizations with energy below 4 khz. This configuration precluded detection of echolocation signals, but was more likely to include detections from species with lower vocalization rates. PAL soundbite data were often limited to a portion of each day by the soundbite quota during periods of high levels of acoustic activity. Vocalizations detected from the spectral time history data increased the amount of time from which vocalizations could be detected in PAL data. Performance metrics comparing detections from the AURAL and PAL will be presented. 18 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

20 Monday 22-Aug-11 16:40-17:00 Renken An approach to generalized marine mammal vocalization detection based on spectrogram representations Martin C. Renken 1, David Gerber Jr 1., Shaari Unger 1 (1) Naval Undersea Warfare Center Division Keyport, 610 Dowell St., Keyport WA, 98345, USA; martin.renken@navy.mil Typical marine mammal detection techniques focus on a particular species of interest to the investigator. While many of these techniques show great promise, ultimately they rely on information specific to the vocalization of that particular species. However, for navy range operations the important question is whether or not any marine mammal species might be present within specific areas of interest. Consequently, effort has been focused on de veloping methods for detecting any marine mammal vocalization without relying on species specific features. The approach described in this presentation was developed using transient orca calls recorded on the Pacific Northwest range, and follows a three step process. The first step reduces the general noise of the spectrogram, the second step uses an edge detection algorithm to highlight certain canonical types of marine mammal sounds, and the final step uses a combination of a threshold and classification technique to identify the potential marine mammal vocalizations. The initial results from this developed process will also be presented. 19 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

21 Monday 22-Aug-11 17:00-17:20 Bougher Improved band-limited marine mammal detector using signal excess Benjamin B. Bougher 1, Joey D. Hood 1, James A. Theriault 2 (1) Akoostix Inc, 10 Akerley Blvd., Suite 12, Dartmouth, NS B3B 1J4, Canada; jhood@akoostix.com (2) Defence Research and Development Canada Atlantic, 9 Grove St., Dartmouth, NS B2Y 3Z7, Canada Akoostix continues to experiment with flexible, low-processing-load marine mammal detection options suitable for implementation in both workstations and low-power embedded systems. Building on previous work, additional processing stages have been added to normalize and de-noise spectrograms using a wide variety of user-configurable options. These pre-processing options can be optimized for the signals of interest, which enhances targets while significantly reducing the impact of structured noise that cause false alarms with other methods. Band-limited signal excess is then computed using one of several weighting functions, after which detection is performed on the signal excess time series. Out-of-band tests are also performed to ensure band-limited signals. The detector is tested on the workshop Minke whale localization dataset with configurations of varying complexity and performance (detection and processing load) are compared to other detection algorithms. 20 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

22 Tuesday 23-Aug-11 08:40-09:00 Roch Recent improvements to the Silbido graph-based whistle detector Marie A. Roch 1 (1) San Diego State University, Dept. of Computer Science, 5500 Campanile Dr, San Diego, California, , marie.roch@sdsu.edu The Silbido whistle detector is designed to detect tonal calls in complex auditory scenes. It does so by constructing a graph representation of the ridges in time series of spectra (spectrograms). Whistle starts, ends, and crossings are all represented as nodes that are interconnected by edges from one node to another. Once a graph has been generated, its interior nodes represent ambiguous crossings which must be resolved. Most whistle detection algorithms make decisions as each frame of data is processed. Silbido delays decisions about whistle crossings until a set of interconnected whistles have been completely processed. By delaying the disambiguation until additional information is available, Silbido permits the algorithm to use information from both sides of whistle crossings to decipher the correct associations. We review the functioning of Silbido, and discuss recent improvements to the algorithm. New additions to the signal processing chain reduce the impact of echolocation clicks and provide additional noise compensation. Part of Silbido s ability to analyze complex acoustic scenes is its capability to bridge gaps in time-frequency ridges of a spectrogram. U nfortunately, while this can also promote the linking of spurious peaks into edges in the graph that produce false positives and occasional excursions from the underlying whistle. We introduce a peak density metric that can be used to remove such edges. The performance of Silbido will be reported on the conference dataset. 21 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

23 Tuesday 23-Aug-11 09:00-09:20 Eshafanian Classification of dolphin whistle types using an eigen-whistle algorithm Mahdi Eshafanian 1, Hanqi Zhuang 1, Nurgun Erdol 1, Ryan Thew 1 (1) Department of Computer and Electrical Engineering and Computer Science, Florida Atlantic University, Boca Raton, FL 33431; erdol@fau.edu We propose a classifier of types of dolphin whistles by adapting the method of eigenfaces used to identify human faces. The association is based on the observation that time-frequency representations such as spectrograms are images that are informative displays of marine mammal vocalizations. The narrow-band dolphin whistles and short duration clicks display identifying features that are localized contours in the image. Their contour-like nature suggests that the class of signals may be efficiently represented by a small number of eigenvectors derived from the principle component analysis (PCA). The Eigen-Whistle Classifier is developed by applying PCA on a set of pre-classified training whistles to construct an eigenwhistle space. The classification is performed by projecting the unclassified testing whistles onto the eigen-whistle space. In this research, we attempt to classify types of dolphin whistles only. Pre-classification processing to remove underwater noise and clicks and multiple harmonics uses spectral subtraction and various image enhancement techniques. Initial results on a small data set are encouraging. Most misclassifications are among whistles whose contours have dominant similarities. The performance of the algorithm can be improved by refining the eigen-whistle space using a larger training data set and with effective preprocessing. This project is partially supported by a grant from the Southeast National Marine Renewable Energy Center. 22 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

24 Tuesday 23-Aug-11 09:20-09:40 Binder Passive classification of marine mammal vocalizations using an automatic aural classifier Carolyn M. Binder 1, Paul C. Hines 1, Joe Hood 2 (1) Defence R&D Canada - Atlantic, P.O. Box 1012, Dartmouth, NS, Canada, B2Y 3Z7; Carolyn.Binder@drdc-rddc.gc.ca (2) Akoostix Inc., 10 Akerley Blvd, Suite 12, Dartmouth, NS, Canada, B3B 1J4 Passive acoustic methods have become increasingly widespread for detection of marine mammals. However, passive sonar systems that are used to localize and track marine mammals by their vocalizations are often triggered by other transient sources, producing a large number of false alarms. Even in the case of successful marine mammal detection, classification of the genus and species is often required, which typically requires expertise in marine mammal vocalization. The shortage of such expert listeners makes this classification task both difficult and costly; automated classification of marine mammal vocalizations would help to overcome these issues. To address this, a prototype classifier has been developed at DRDC that uses perceptual signal features similar to those employed in the human auditory system. A previous effort has shown that the classifier reduced false alarm rates and successfully discriminated marine mammal vocalizations from four species: the northern right whale, sperm whale, humpback whale, and bowhead whale. In the current paper, the dataset has been extended to include low SNR vocalizations from the four species mentioned previously as well as minke whale vocalizations obtained from the Fifth International DCL workshop dataset. 23 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

25 Tuesday 23-Aug-11 09:40-10:00 Caillat A Bayesian approach to study the uncertainty generated by miss-classified acoustic detections Marjolaine Caillat 1,2, Len Thomas 2, Douglas Gillespie 1 (1) Sea Mammal Research Unit, St Andrews University, St Andrews KY16 8LD UK; mc326@st-andrews.ac.uk (2) Centre for Research into Ecological and Environmental Modeling, The Observatory, Buchanan Gardens, University of St Andrews, St Andrews KY16 9LZ, UK In order to estimate abundance or to apply mitigation procedures to specific cetacean species using acoustic data it is necessary to correctly identify the species acoustically detected. Many different classifiers have been developed and tested in the last decade to try to classify cetacean clicks, whistles or baleen whales calls. Whistles can vary within a species from population to population or conversely can be very similar between two different species. For this reason developing automatic classifier with 100% correct classification rate for each species is probably unrealistic. To use acoustic data, it is necessary to develop a method which will re-estimate the true number of detections by species from the classification output data containing miss-identified detections. A Bayesian statistical model has been developed to solve this problem. Each classifier need to be trained with identified data. From this data, the quality of a classifier can be represented through a confusion matrix which will show the correct and the miss-classification rates between species. This confusion matrix can be used as a prior in the Bayesian model to estimate the true number of detections for each species. When this prior is very informative (similar classification rates when the classifier is trained with different datasets) the Bayesian model allows the re-estimation of the true number of detections. However if the confusion matrix is more variable between training datasets, our Bayesian model has difficulty reestimating the true number of detections. 24 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR

26 Tuesday 23-Aug-11 10:00-10:20 Yang Noise reduction for better detection of beaked whale clicks Yang Lu 1, Holger Klinck 1, David K. Mellinger 1 (1) Cooperative Institute for Marine Resources Studies, Oregon State University and NOAA Pacific Marine Environmental Laboratory, 2030 SE Marine Science Dr., Newport, OR USA ; Lu.Yang@noaa.gov We seek to improve the signal-to-noise (SNR) ratio of clicks recorded from Blainville s beaked whales (Mesoplodon densirostris). One method is an optimal Weiner filter, but it can have high computational cost. To reduce the computational cost, a sub-optimal Weiner filter has been designed. To test the filter, a detector using the filter output has been designed for the detection of beaked whale calls. Through simulations, the proposed detector is shown to be capable of detecting most of the desired clicks, but is not able to differentiate other co-existing species such as Risso s dolphins and pilot whales that is, it efficiently detects all clicks. By combining the proposed detector with the Energy Ratio Mapping Algorithm (ERMA; Klinck and Mellinger in press), which measures energy differences between different species, higher detection accuracy for beaked whale clicks can be achieved. The filter can also be used to improve the SNR of other marine mammal acoustic signals. 25 Abstracts, DCLDE Workshop 2011, Mt. Hood, OR