Analysis for an Acoustic Survey of Beaked Whales

Transcription

1 Analysis for an Acoustic Survey of Beaked Whales Joe Hood, Akoostix Inc. Benjamin Bougher, Akoostix Inc. Hilary Moors, Independent Contractor Olivia Paitich, Akoostix Inc. Prepared by: Akoostix Inc. 10 Akerley Blvd - Suite 12 Dartmouth, NS B3B 1J4 Contractor's Document Number: AI CR Contract Project Manager: Joe Hood, (902) PWGSC Contract Number: W Technical Authority: James Theriault, Defence Scientist The scientific or technical validity of this Contract Report is entirely the responsibility of the Contractor and the contents do not necessarily have the approval or endorsement of the Department of National Defence of Canada. Contract Report DRDC-RDDC-2017-C083 March 2012

2 Principal Author Original signed by Joe Hood Joe Hood President, Akoostix Approved by Original signed by Robert A. Stuart Robert A. Stuart Head, Technology Demonstration Approved for release by Original signed by Calvin V. Hyatt Calvin V. Hyatt Chairman, Document Review Panel Her Majesty the Queen in Right of Canada, as represented by the Minister of National Defence, 2012 Sa Majesté la Reine (en droit du Canada), telle que représentée par le ministre de la Défense nationale, 2012

3 Abstract.. DRDC Atlantic, with support from the Government of Canada s Interdepartmental Recovery Fund (IRF), undertook to apply some of the technology developed for the purpose of mitigating the potential impact of active sonars to enhance acoustic characterization of the Northern bottle nose whale (NBW) (Scotian shelf population) critical habitat. Data were collected during the Canadian Forces Auxiliary Vessel (CFAV) Quest Q338 sea trial (26-30 May 2011) where a mixture of towed array, sonobuoy, and passive acoustic reusable buoy (PARB) systems were used to record acoustic signals in the Gully marine protected area (MPA), east of Nova Scotia. The data were analyzed for presence/absence of species of interest and habitat usage maps were generated using the manual detections and GPS information recording position data for each sensor. During analysis, acoustic samples were extracted to provide sample sounds for the DFO catalogue of marine mammal sounds. In addition to data analysis, an improved band-limited automated detector was developed. The detector was configured for species of interest, and used to generate detection records for each sensor deployed during the trial. The results showed that much of the NBW population resides near the mouth of the Gully. NBW detection rate significantly decreased as the survey moved north to the head of the canyon and into the Gully trough. i

4 Executive summary Analysis for an Acoustic Survey of Beaked Whales: Joe Hood; Benjamin Bougher; Hilary Moors; Olivia Paitich; March Introduction: DRDC Atlantic has a twenty year interest and history in research on the detection, classification, localization, and tracking of marine mammals. From a DRDC prospective, the ultimate goal is to develop technologies that enable the use of active sonar while maintaining a sufficiently low probability of harm to the environment. The specific goals have been to develop acoustic monitoring techniques that characterize the mammal use of the environment and to detect individual animals should they fall within an impact zone. With support from the Government of Canada s Interdepartmental Recovery Fund, DRDC undertook to apply some of the technology developed for the purpose of mitigating the potential impact of active sonars to enhance acoustic characterization of the Northern bottle nose whale (NBW) (Scotian shelf population) critical habitat. Data were collected during the Canadian Forces Auxiliary Vessel Quest sea trial in 2011 where a mixture of towed array, sonobuoy, and passive acoustic reusable buoys were used to record acoustic signals in the Gully marine protected area, east of Nova Scotia. Results: The data were analyzed for presence/absence of species of interest and habitat usage maps were generated using the manual detections and GPS information recording position data for each sensor. During analysis, acoustic samples were extracted to provide sample sounds for the DFO catalogue of marine mammal sounds. In addition to data analysis, an improved bandlimited automated detector was developed. The detector was configured for species of interest, and used to generate detection records for each sensor deployed during the trial. The results showed that much of the NBW population resides near the mouth of the Gully. NBW detection rate significantly decreased as the survey moved north to the head of the canyon and into the Gully trough. Significance: Researchers successfully demonstrated the potential for using acoustic monitoring surveys to assess the distribution and abundance of marine mammals, with a focus on cetacean species found on the Scotian Shelf, particularly in the Gully region. Analysts were able to demonstrate manual and automated detection, showing how these two approaches can work together to ensure both information quality and efficiency in analysis. Additionally, this project allowed the improvement and demonstration of DRDC marine mammal monitoring technologies that in turn can be used to improve the ability to mitigate the potential impact of active sonar operations on marine life. Future plans: Additional surveys and effort to develop improved detection and analysis processes and software are required to broaden the scope of this. Work is also required to improve sensors, reducing the impact of self-noise due to issues like electronic cross-talk, tow noise, and RF interference, as this noise can have a negative impact on the quality of automated detection, while also reducing the rate of manual analysis. ii

5 Table of contents Abstract..... i Executive summary... ii Table of contents... iii List of figures... v List of tables... viivii Acknowledgements... viii 1 Introduction Overview of Q338 trial and data Ship and sensor track Sensor characteristics Towed array AN/SSQ-53F sonobuoys PARB Data processing Data formatting Ambient noise processing Automated detection of transients and cetacean vocalizations Manual analysis process Analysis results Manual detection analysis Automated detection Habitat usage maps Acoustic data samples Ambient noise measurements Discussion and recommendations Monitoring and survey plans Sensors enabling advanced processing System/analysis tools and process development Detector testing and improvement False alarm reduction ACDC software application Software sonobuoy GPS decoder Ambient noise monitoring Summary and Conclusions References iii

6 Annex A.. Software Tools A.1 Signal Processing Packages (SPPACS) A.1.1 Background and Design Information A.2 STAR-IDL A.3 Omni-Passive Display (OPD) A.4 Acoustic Cetacean Detection Capability (ACDC) A.5 Acoustic Subsystem (AS) A.6 Passive Acoustic Reusable Buoy (PARB) Annex B... Software development B.1 Detector development B.1.1 SPPACS applications B.1.2 Detector configuration B.2 Geo-referencing detection data Annex C... Configuration management C.1 STAR branch and release information C.1.1 STAR software documentation C.2 Issue summary List of symbols/abbreviations/acronyms/initialisms iv

7 List of figures Figure 1: Map showing the Gully MPA and the Gully trough Figure 2: Trial reconstruction plot for the recording period on 27 May, Positions for sonobuoys were taken from the deployment location Figure 3: Trial reconstruction plot for the recording period on 29 May, Positions for sonobuoys were taken from the deployment location Figure 4: Trial reconstruction plot for the recording period on 30 May, Positions for sonobuoys were taken from the deployment location Figure 5: Block diagram of the processing stream used to generate 5-minute 1.0 Hz ambientnoise spectra and 3 rd -octave spectra Figure 6: Proportion of five-minute recording segments analyzed from each system which had specific types of biological signals present on them from 27 May Figure 7: Proportion of five-minute recording segments analyzed from each system which had specific types of biological signals present on them from 29 May Figure 8: Proportion of five-minute recording segments analyzed from each system which had specific types of biological signals present on them from 30 May Figure 9: Proportion of five-minute recording segments analyzed from each system which had specific types of non-biological signals present on them from 27 May Figure 10: Proportion of five-minute recording segments analyzed from each system which had specific types of non-biological signals present on them from 29 May Figure 11: Proportion of five-minute recording segments analyzed from each system which had specific types of non-biological signals present on them from 30 May Figure 12: Detection-density graphs for each detection stream taken from the first sonobuoy deployed on 27 May 2011, when cetacean vocalization activity was very high. Data for each detection stream are displayed in one row, with triangles marking detection density for each minute that contains at least one detection. Humpback data are not shown, because there were not any detections. Green lines indicate when the sensor was recording, while red lines would mark when the sensor was off. The total detection count for each band is labelled next to the name for the detector configuration Figure 13: Detection-density graphs for each detection stream taken from the second sonobuoy deployed on 30 May 2011, when cetacean vocalization activity was quite low. Data for each detection stream are displayed in one row, with triangles marking detection density for each minute that contains at least one detection. Green lines indicate when the sensor was recording, while red lines mark when the sensor was off. The total detection count for each band is labelled next to the name for the detector configuration Figure 14: NBW habitat usage map for PARB data recorded on 27 May Detection is near continuous for the entire sampling period v

8 Figure 15: NBW habitat usage map for towed array data recorded on 29 May There is a distinct reduction in detection as the survey progresses north Figure 16: Example of a raw spectra ambient-noise graph taken from channel 1 of the sonobuoy on 29 May The highest noise levels may be due to RF interference and not acoustic signals Figure 17: Example of 3 rd -octave averaged ambient-noise graph generated using channel 1 of the sonobuoy data recorded on 29 May Figure B-18: SPPACS block diagram of the detector stream. Boxes outlined with hashed lines are optional processing blocks vi

9 List of tables Table 1: Summary of all data analysis including the number of five-minute recording segments analyzed from each system and which had specific types of biological signals present on them Table 2: Summary of the number of five-minute recording segments analyzed from each system and which had specific types of non-biological signals present on them Table 3: Cetacean sightings from the Q338 trial. No animals were sighted on 30 May Table 4: Metadata for the 20 acoustic samples provided to the DFO marine mammal sound catalogue Table C-5: Issue summary (severity vs. status) for all software on STAR release Table C-6: Known Critical issues for STAR vii

10 Acknowledgements Akoostix would like to acknowledge the efforts of the Defence Research and Development Canada (DRDC) Canadian Forces Auxiliary Vessel (CFAV) Quest Q338 sea trial team, for whom the Chief Scientist was Donald Mosher. The data analyzed under this trial would not have been gathered without their efforts, care, and diligence. We would also like to acknowledge the funding provided to DRDC Atlantic from the government of Canada s Interdepartmental Recovery Fund (IRF) which made this work possible. viii

11 1 Introduction This contractor report documents work performed under contract W for Project Authority (PA), James Theriault. The work was performed between November 2011 and March This work was conducted to support Defence Research and Development Canada (DRDC) Atlantic s efforts to perform enhanced acoustic characterization of the Northern bottlenose whale (NBW) (Scotian Shelf population) critical habitat. Their initiative is intended to address Species at Risk Act (SARA) recovery strategy objectives. In addition, these deliverables will be used to document and provide samples for the Department of Fisheries and Oceans (DFO) Canada catalogue of marine mammal sounds. Data collected during Canadian Forces Auxiliary Vessel (CFAV) Quest cruise Q338 were analyzed during this contract. These data include acoustic recordings from a two-hydrophone towed array, AN/SSQ-53F sonobuoys fitted with GPS, and Passive Acoustic Reusable Buoys (PARB). These data were analyzed by: Executing automated detection using improved Software Tools for Analysis and Research (STAR)-based automated detection software that was integrated it into the Signal Processing Packages (SPPACS) under this call-up. Analysing the Q338 data sets to determine times when species of interest were vocalizing using a combination of manual and automated processes. Creating habitat usage maps for northern bottlenose whales (NBW) and other cetaceans identified during analysis. Processing the Q338 data set to produce ambient-noise level measurements. The data provided an interesting result that can be used to improve confidence in how the NBW population is distributed in the Gully MPA; much of the NBW population resides near the mouth of the Gully, as detection rate significantly decreased as the survey moved north to the head of the canyon and into the Gully trough (Figure 1). This contract was conducted in parallel with other important work. This situation allowed each to leverage knowledge, capability, and effort from the other. Collaborative efforts included: A Department of National Defence s (DND) study [6] intended to: Improve knowledge of acoustic monitoring technologies (sensor systems and analysis software) that may be applied to long- and short-term marine mammal monitoring efforts. Exercise candidate technologies in order to gain a practical understanding of the strengths and weaknesses. Provide recommendations on how to improve knowledge of cetacean distribution and density on spatial and temporal scales so that it might be applied in naval exercise planning and decision making for mitigation of the impact of operations on marine mammals. 1

12 A DRDC study intended to research and prototype practical methods for reducing acousticdata analysis effort for acoustic survey data for density estimation [4]. This report provides a detailed description of the work performed for each of the tasks listed in the Statement of Requirements (SOR). The report is organized as follows: Section 2 provides an overview of the Q338 data and acoustic sensors, Section 3 documents the methods used to process the data, Section 4 explains the analysis process, Section 5 summarizes the results of the analysis, Section 6 provides recommendations on how to improve analysis and for future work, Section 7 summarizes the project results, Annex A provides a brief description of the software used to support this project, Annex B documents the software developed under this contract, and Annex C provides configuration management (CM) information to help users understand which version of the software was used for this work. 2

13 2 Overview of Q338 trial and data CFAV Quest cruise Q338 was conducted between 24 May 2011 and 31 May During this sea trial, acoustic data were collected from the Gully submarine canyon located at the edge of the Scotian Shelf, 200 km southeast of Nova Scotia, including the core region of the Gully Marine Protected Area (MPA) and the Gully trough just north of the MPA (Figure 1). Figure 1: Map showing the Gully MPA and the Gully trough. Data sets were collected over a three-day period using a 2-element towed hydrophone array, sonobuoys, and a passive acoustic reusable buoy (PARB). The related data and information were organized into a directory structure consistent with the STAR analysis process [3] and stored in the DRDC trial repository trials09 under the directory Q338. Detailed sensor information is provided in Section 2.2. A brief summary of each deployment is provided below: On 27 May 2011, two PARB units were deployed over the period of 12:20Z to 18:30Z and 13:20Z to 17:00Z in a core region of the Gully MPA, near the deployment site of a bottommoored Autonomous Multi-channel Acoustic Recorder (AMAR) 1. Two AN/SSQ-53F 1 The AMAR could not be recovered, likely due to a faulty acoustic release, and so no data were available from this sensor. 3

14 sonobuoys operating in calibrated omni-directional (CO) mode were also deployed. This data set is located in the directory Q338/ On 29 May 2011, a line transect was conducted along the north-south axis of the Gully, starting from the mouth of the canyon and running north to the head of the canyon. The towed hydrophone array was deployed around 12:20Z and towed behind Quest at approximately 3 kts. Two AN/SSQ-53F sonobuoys operating in calibrated omni-directional (CO) mode were also deployed. This data set is located in the directory Q338/ On 30 May 2011, a second transect was conducted from the head of the canyon to the northwest corner of the Gully trough. The towed hydrophone array was deployed at approximately 11:22Z and towed behind Quest at approximately 3 knots. Three AN/SSQ- 53F sonobuoys operating in calibrated omni-directional (CO) mode were also deployed, including one at 13:17Z. This data set is located in the directory Q338/ Ship and sensor track Positions for the ship and PARB units used during the Q338 trial were obtained as Quest NADAS (non-acoustic data [NAD] acquisition system) and GPS NMEA logs. Though some sonobuoys were fitted with GPS, these data were not available for this analysis; instead the position of deployment was used. Sonobuoy drift after deployment is unknown. The NAD logs were formatted and cross-referenced to generate STAR-IDL formatted tracks, which are saved in the NAD subdirectory of each data set. They were verified by generating trial reconstruction plots, which are shown in Figure 2, Figure 3, and Figure 4. Figure 2: Trial reconstruction plot for the recording period on 27 May, Positions for sonobuoys were taken from the deployment location. 4

15 Figure 3: Trial reconstruction plot for the recording period on 29 May, Positions for sonobuoys were taken from the deployment location. Figure 4: Trial reconstruction plot for the recording period on 30 May, Positions for sonobuoys were taken from the deployment location. 5

16 2.2 Sensor characteristics Three types of sensors were used during Q338: an array towed from CFAV Quest, six AN/SSQ- 53F sonobuoys, and two PARB units. These relevant characteristics are provided in the following subsections Towed array The towed array used during the trial by was borrowed from the Whitehead Lab at Dalhousie University, and was built by researchers at St. Andrew's University in the UK. The entire array was 100 m in length, with a pair of Benthos AQ4 hydrophones (frequency response ± 1 db re 1 V/μPa from 5-30 khz) spaced 1.8 m apart at the end of the cable in a fluid filled tube. The hydrophones were connected to custom-built preamplifiers (which added 27 db re 1V/μPa to the incoming signal), a custom-built high-pass filter and a Kemo Pocketmaster 1600 low-pass filter. The analogue array was connected to DRDC s Environmental Acoustic Data Acquisition System (EADAQ) for recording on channels 7 and 8. EADAQ was set to record 16-bit signed data, covering the range ±10 Volts, sampled at Hz. Data are stored in Defence Research Establishment Atlantic (DREA) DAT formatted files AN/SSQ-53F sonobuoys AN/SSQ-53F are standard sonobuoys that operate in three modes: DIFAR, constant shallow omni (CSO), and calibrated omni-directional (CO). Buoys were operated in CO mode for this trial because CO mode provides the greatest bandwidth, covering 10 Hz to 40 khz, though calibration is only for 10 Hz to 20 khz. These sensors transmit acoustic time-series as analogue data on a very high frequency (VHF) frequency modulated (FM) radio link that provides approximately 40 db of dynamic range. Data were received using a standard AN/ARR-502B sonobuoy receiver and recorded using the same EADAQ system as for the towed array. Sonobuoy data were recorded on channel 4 and 5. Some of the sonobuoys were fitted with GPS, but GPS logs were not provided with the data. It may be possible to post-process data files to extract GPS data, as described in Section PARB Two different PARB units, named LP1 and LP2, were used during the trial. PARBs are digital floating buoys with sensors deployed on a cable, similar to sonobuoys. PARBs contain a computer and storage, providing onboard processing and digital recording. An omni-directional hydrophone was suspended on a cable of unknown length, while a suspension system was used to decouple the hydrophone from surface motion. The ambient recording function was used for this trial. In this mode, the system was set to record 30-minutes of data per file. There is a 4-minute gap between the end of one file and the start of 6

17 the next, as the system closes off recording and prepares to start recording the next file. 2 Data are stored in single-channel 16-bit WAV files, sampled at Hz. More detail on PARBs is provided in Annex A.5. 2 If required, this gap has been reduced to 7 seconds for newer data sets and could be further reduced. 7

18 3 Data processing This section provides detail on how trial data were processed by Akoostix. This processing was performed using scripts that were executed via semi-automated processes. Raw acoustic data from three data sets were provided to Akoostix as WAV files, and were processed for further analysis as follows: The data were converted into little-endian DREA-DAT-formatted files that could be processed using SPPACS applications. This included addition of time-stamp information. This process is described in Section 3.1. The acoustic data were processed to compute approximately 1.0 Hz spectra and 3 rd -octaveaveraged spectra for ambient-noise analysis. This process is described in Section 3.2. The acoustic data were processed to automatically detect transients and cetacean vocalizations. These results were stored in data files and logs compatible with STAR detection-analysis requirements, so that the results could be analyzed using STAR-IDL, Omni-passive Display (OPD) and the Acoustic Cetacean Detection Capability (ACDC). This process is described in Section Data formatting Raw acoustic data from each sensor system was converted into little-endian DREA-DATformatted files, time stamps were applied, and the new files were stored in the raw_data directory associated to each data set. A subdirectory was created for each sensor system, named to simplify identification of the source of the data and to allow for cross-referencing sensor position information. Acoustic data from each sensor type required different data formatting: Sonobuoy and towed-array data were recorded using EADAQ and was received by Akoostix as WAV files with file time-stamps contained in the filenames. The WAV files were converted to DREA DAT format using sp_wav2dat, sonobuoy and towed-array data were separated using sp_extract, then the output DREA DAT file headers were timestamped using sp_ph. The processing streams are executed using sono_wav2dat.py and towedarray_wav2dat.py, which are contained in the main_scripts directory of the trials09/q338 trial. PARB data were received in WAV files with time information stored in an accompanying header file. Data were converted to DREA DAT format and time-stamped using the SPPACS application sp_wav2dat, which is executed using the PARB_wav2dat in the scripts directory of the data set. 3.2 Ambient noise processing Acoustic data were processed to form 5-minute averaged 1.0 Hz spectra and then 3 rd -octaveaveraged spectra. These data were not calibrated, though these corrections could be applied at a 8

19 later date. The data may be used to examine the time-dependence of ambient noise in the area and differences in acoustic sensitivity for each sensor system. The ambient-noise processing stream is depicted in Figure 5. Minor differences in spectral resolutions due to sensor sampling rate are noted below: Spectral processing FFT window sizes were rounded down to the closest power of two, making spectral resolutions. Specific processing parameters are documented in the ambient_3rd_octave.py Python script located in the main_scripts directory of the trials09/q338 trial. Figure 5: Block diagram of the processing stream used to generate 5-minute 1.0 Hz ambientnoise spectra and 3 rd -octave spectra. 3.3 Automated detection of transients and cetacean vocalizations Detection data were generated for the Q338 data sets using the automated detector described in Annex B.1. The detector was configured for the cetacean species and call types that were expected for the Gully, using the settings generated during Glider 2012 trial analysis [6]. Detector configurations were generated for the following signal types: General transients, Sperm whales, Fin whales, Minke whales, Sei whales, Humpback whales, General squeals, and High-frequency clicks (still named delphinids in automated detection graphs; see below). For the high-frequency click detector, the following logic was used to adjust the detector s name from that used in [6]. In [8], researchers reported a mean peak frequency of NBW foraging (echolocation) clicks of 24 khz, although the frequency range of these clicks extend to at least 40 khz (the upper frequency limit of the sensing system used). More recent analysis of recordings of 9

20 NBWs which span a higher frequency range indicate that the energy of their clicks extends well above 40 khz, and most of energy roughly occurs between khz. The high-frequency click detector was originally called the delphinid click detector in [6]. This detector was configured to detect high-frequency clicks, which were most likely produced by dolphins and pilot whales in the area from which the February 2012 Glider trial recordings were obtained. While this detector triggers on dolphin and pilot whale clicks in the Gully area, it also detects the high frequency clicks of northern bottlenose whales and therefore was renamed the high-frequency click detector. Each detector configuration was applied to the Q338 data and detection logs were saved to the NAD/transient_detect directory related to each data set. This processing was executed using the Python script cetacean_script.py, stored in the main_scripts directory of the trials09/q338 trial. The detection processing used the target files contained in the main_target_files directory of the same trial. The resulting detections logs can be used for further analysis using STAR-IDL, OPD, and ACDC, which are described in Annex A. 10

21 4 Manual analysis process Analysis was performed using OPD for visual spectral analysis and Audacity for aural classification. Data were divided into five-minute segments 3 and then each segment was analyzed for presence/absence of the following pre-defined signals: Northern bottlenose whale clicks: high frequency clicks with most energy usually concentrated at frequencies > 15 khz and the spacing between successive clicks, known as the inter-click interval (ICI), of approximately 0.4 sec. Sperm whale clicks: low frequency clicks with most energy usually concentrated at frequencies <10 khz and ICI usually > 0.5 sec. Other clicks: clicks produced by other species (such as dolphins or pilot whales), usually high-frequency clicks with most energy concentrated at frequencies > 10 khz and ICI usually between sec. High-frequency whistles: narrowband whistles and squeals with most energy usually occurring at frequencies > 10 khz (and most often within the khz range) with few harmonics present, likely produced by dolphin species such as common dolphins or Atlantic white-sided dolphins, although they may also be produced by pilot whales or other species of the area. Low frequency whistles: narrowband whistles and squeals with most energy usually occurring at frequencies <10 khz, with multiple harmonics usually present, likely produced by pilot whales. Buzzes: high frequency clicks with very short ICI (<0.05 sec), usually occur when other clicks are present (a "zzzzzzzz" sound). Hydrophone/system noise: noise attributed to the recording sensor or system itself such as the acoustic signal cutting in and out, knocking or clanking unlikely to be biological sounds and static or electrical interference (crackling). Depth sounders: 12 khz depth sounder (onboard CFAV Quest). Engine noise: noise attributed to Quest's engine or other vessel engine noise such as noise produced by the RHIB boat during deployment and retrieval of the PARB systems. Other: any other unique sounds heard on the recordings. Notes for each data segment pertaining to general data quality, specific times of interest, and other analysis notes were provided. These notes were used to extract and generate the acoustic data samples, described in Section 5.4. The results of the presence/absence analysis are summarized in Section 5.1 and presented as habitat usage maps in Section While most of the segments were five-minutes long, some, particularly segments at the very beginning or very end of a recording, were less than five minutes long. 11

22 5 Analysis results The following sections describe the results of the manual and automated analyses. Results from the manual detection analysis, described in Section 5.1, were used to create the habitat usage maps discussed in Section 5.3 and to select appropriate data samples for the DFO database discussed in Section 5.4. Automated detection processing was also performed, as discussed in Section 5.2, while the ambient-noise analysis is documented in Section Manual detection analysis In total, 815 five-minute recording segments were analyzed for the presence of various types of signals including both biological and non-biological sounds. The full presence/absence data set is stored with the data as an MS Excel spreadsheet, while Table 1 and Table 2 provide a summary of the number of recording segments analyzed and which segment had a specific type of signal present on them for each system of each day. A fair amount of hydrophone/system noise occurred on some of the recordings (Table 2), especially those taken from the PARBs and towed array (see Figure 9 thru Figure 11). A relatively high proportion of the recordings collected on 27 May 2012 also had engine noise present (Figure 9). Despite the presence of noise on the recordings, cetacean signals could easily be manually detected on almost all of the analyzed recordings. Noisier recordings, however, are expected to affect automated detector performance (the false alarm rate is expected to increase on noisier recordings). The recordings made on 27 May and 29 May were obtained from the core region of the Gully MPA and within the NBW critical habitat. As expected, NBW clicks were present on most of the recordings obtained from these two days. Similarly, sperm whale, pilot whale, and dolphin vocalizations were also present on most of the recordings obtained on these days (Figure 6 thru Figure 8). In most cases, sperm whale and other high-frequency clicks occurred on a higher proportion of recordings than NBW clicks. Low-frequency whistles occurred more frequently on the recordings than high-frequency whistles. Pilot whales and NBW were the only species sighted on 27 May and 29 May (Table 3). Very few of the recordings obtained from 30 May had cetacean vocalizations present on them (Table 1). High-frequency clicks not created by NBW and some low-frequency whistles were detected on a very low percentage of the recordings (Figure 8). The trough of the Gully is not considered to be critical habitat for NBW, therefore it makes sense that no NBW clicks were detected on the recordings obtained from this area. Sperm whales, dolphins, and pilot whales have occasionally been sighted in this area, but it is not an area especially known to attract a high abundance of cetaceans. No cetaceans were sighted on 30 May. 12

23 Table 1: Summary of all data analysis including the number of five-minute recording segments analyzed from each system and which had specific types of biological signals present on them. Date of recording System Channel Number of 5-min recording segments analyzed Bottlenose whale click Sperm whale click Other click High freq. whistle Low freq. whistle Buzz 05/27/11 PARB 1 (LP1) /27/11 PARB 1 (LP2) /27/11 Sonobuoy /27/11 Sonobuoy /29/11 Sonobuoy /29/11 Sonobuoy /29/11 Towed array 7, /30/11 Sonobuoy /30/11 Sonobuoy /30/11 Towed array 7, Figure 6: Proportion of five-minute recording segments analyzed from each system which had specific types of biological signals present on them from 27 May

24 Figure 7: Proportion of five-minute recording segments analyzed from each system which had specific types of biological signals present on them from 29 May Figure 8: Proportion of five-minute recording segments analyzed from each system which had specific types of biological signals present on them from 30 May

25 Table 2: Summary of the number of five-minute recording segments analyzed from each system and which had specific types of non-biological signals present on them. Date of recording System Channel Number of 5-min segments analyzed Hydrophone/ system noise Depth sounder Engine noise Other 05/27/11 PARB 1 (LP1) /27/11 PARB 1 (LP2) /27/11 Sonobuoy /27/11 Sonobuoy /29/11 Sonobuoy /29/11 Sonobuoy /29/11 Towed array 7, /30/11 Sonobuoy /30/11 Sonobuoy /30/11 Towed array 7, Figure 9: Proportion of five-minute recording segments analyzed from each system which had specific types of non-biological signals present on them from 27 May

26 Figure 10: Proportion of five-minute recording segments analyzed from each system which had specific types of non-biological signals present on them from 29 May Figure 11: Proportion of five-minute recording segments analyzed from each system which had specific types of non-biological signals present on them from 30 May

27 Table 3: Cetacean sightings from the Q338 trial. No animals were sighted on 30 May Date Local Time Species Number of individuals Distance from boat [m] Bearing to the animal [degrees starboard] 05/27/2011 8:05-8:14 Bottlenose whale /27/2011 8:44-9:26 Bottlenose whale /27/ :07-14:14 Pilot whale /27/ :25-15:39 Pilot whale /29/2011 7:14-7:16 Pilot whale /29/2011 8:14-8:15 Bottlenose whale /29/2011 8:52-9:12 Bottlenose whale /29/2011 8:56-9:07 Pilot whale /29/2011 9:44-9:48 Bottlenose whale /29/2011 9:46-9:53 Pilot whale /29/ :22-10:23 Bottlenose whale /29/ :43 Bottlenose whale /29/ :18-15:23 Bottlenose whale Automated detection The detector described in Annex B.1 was configured with the processing streams described in [6] and Section 3.3. Plots of detection density versus time were generated for each detection stream, recorder, and deployment. Sample density graphs are provided in Figure 12 (27 May 2012) and Figure 13 (30 May 2012), while all graphs are available in the analysis_results/detection_analysis subdirectory of each data set. It is clear from Figure 13, where there were few valid detections that the false-alarm rate is high for some species yet quite low for others. Generally false-alarm rate is higher for click detection, while it is very low for narrowband signals (e.g. humpback). This is because it is harder for the detector to discriminate a single cetacean click from static, broadband transients, etc., while narrowband transients are more distinct. One option for improvement is to consider a sequence of detections as a click train and improve discrimination by examining this set, as considered in [4][5]. Here time delay between subsequent clicks on individual sensors and across sensors is considered, as ICI and source location provide discrimination features. There are a number of other options for improving detector performance, which are further discussed in Section 6. 17

28 18 Figure 12: Detection-density graphs for each detection stream taken from the first sonobuoy deployed on 27 May 2011, when cetacean vocalization activity was very high. Data for each detection stream are displayed in one row, with triangles marking detection density for each minute that contains at least one detection. Humpback data are not shown, because there were not any detections. Green lines indicate when the sensor was recording, while red lines would mark when the sensor was off. The total detection count for each band is labelled next to the name for the detector configuration.

29 Figure 13: Detection-density graphs for each detection stream taken from the second sonobuoy deployed on 30 May 2011, when cetacean vocalization activity was quite low. Data for each detection stream are displayed in one row, with triangles marking detection density for each minute that contains at least one detection. Green lines indicate when the sensor was recording, while red lines mark when the sensor was off. The total detection count for each band is labelled next to the name for the detector configuration. 5.3 Habitat usage maps Akoostix was tasked to overlay detection data on a map, showing regions containing active NBW and other species of interest. These maps are created using STAR-IDL as described in Annex B.2. The maps were created as segments with the same display area for each sensor, enabling the end user to overlay different species on a single map if required. Another advantage of these maps is that they show the sample area. Sample habitat usage maps are provided in Figure 14 (27 May 2012) and Figure 15 (29 May 2012), while all plots are available in the main_analysis_result subdirectory of the trials09/q338 STAR trial repository. 19

30 Figure 14: NBW habitat usage map for PARB data recorded on 27 May Detection is near continuous for the entire sampling period. Figure 15: NBW habitat usage map for towed array data recorded on 29 May There is a distinct reduction in detection as the survey progresses north. 20

31 5.4 Acoustic data samples Twenty sample acoustic files were submitted to Jack Lawson of DFO (via DRDC), including recordings of northern bottlenose whales, sperm whales, pilot whales and dolphins. STAR annotations that contain all metadata required to define the data source for each sample were generated, and are saved in main_analysis_result/dfo_library_annotations.ann. Audio extracts were generated from the annotations and are saved as WAV format in the main_analysis_result/dfo_wav_library directory. Human readable metadata containing for each sample is contained in main_analysis_result/dfo_library_annotations.csv as well as presented in Table 4. Table 4: Metadata for the 20 acoustic samples provided to the DFO marine mammal sound catalogue FILENAME sample_20.wav sample_1.wav sample_2.wav sample_3.wav sample_4.wav sample_5.wav sample_6.wav sample_7.wav sample_8.wav sample_9.wav sample_10.wav sample_11.wav START TIME 13:29: :36: :11: :13: :13: :16: :16: :48: :50: :58: :01: :13: END TIME RECORDER LATITUDE LONGITUDE DESCRIPTION 13:30: LP :37:17 clicks and whistles (unsure of species) 2011 LP Sperm whale clicks 12:12: sonobuoy :13: sonobuoy :15: sonobuoy :19: sonobuoy :27: sonobuoy Sperm whale clicks (plus some high freq. clicks) Bottlenose whale clicks (plus some sperm whale clicks) Busy recording with sperm whales and bottlenose whales, maybe delphis as well Busy recording with sperm whales and bottlenose whales, maybe delphis as well Bottlenose whale and sperm whale clicks 12:49: sonobuoy Sperm whale clicks 12:51: sonobuoy Sperm whale clicks 12:59: sonobuoy Sperm whale clicks 15:02: sonobuoy :17: sonobuoy High freq. clicks, whistles and buzzes (likely delphiniums, unsure of species) High freq. clicks, whistles and buzzes (likely delphiniums, unsure of species) 21

32 FILENAME sample_12.wav sample_13.wav sample_14.wav sample_15.wav sample_16.wav sample_17.wav sample_18.wav sample_19.wav START TIME 15:22: :26: :49: :08: :38: :41: :48: :33: END TIME RECORDER LATITUDE LONGITUDE DESCRIPTION 15:23: sonobuoy :27: sonobuoy :52: sonobuoy :17:02 High freq. clicks, whistles and buzzes (likely delphiniums, unsure of species) Clicks and buzzes (likely delphiniums, unsure of species) High freq. clicks and low freq. whistles (unknown species but likely pilot whales) High freq. clicks and interesting low freq. whistles (unknown species but likely pilot whales) 2011 sonobuoy :39: towed_array Bottlenose whale clicks :48: towed_array Bottlenose whale clicks and low freq. whistles :49: towed_array Bottlenose whale clicks :34: towed_array Sperm whale slow clicks 5.5 Ambient noise measurements The data was processed for ambient-noise measurements, as described in Section 3.2. Plots of mean, median, maximum, and minimum spectra were generated for raw spectra and 3 rd octave averaged spectra using one-hour sample sizes. Plots were generated for each sensor and deployment using the process_an_data2.pro script located in the main_idlprog subdirectory. All graphs and corresponding csv files are located in the analysis_results/ambient_results/ subdirectory of each data set. Examples of raw and 3 rd -octave 5-minute-averaged ambient-noise graphs are shown in Figure 16 and Figure 17 respectively. 22

33 Figure 16: Example of a raw spectra ambient-noise graph taken from channel 1 of the sonobuoy on 29 May The highest noise levels may be due to RF interference and not acoustic signals. Figure 17: Example of 3 rd -octave averaged ambient-noise graph generated using channel 1 of the sonobuoy data recorded on 29 May

34 6 Discussion and recommendations This project provided an excellent opportunity to use current acoustic monitoring and data processing technologies and processes to generate information from acoustic data. This section provides a discussion and summary of what was learned, along with recommendations for future monitoring, surveys, and technology improvement. 6.1 Monitoring and survey plans The idea of generating habitat usages maps from acoustic survey data appears useful, but much more data is required to generate a proper map. Additional survey should be conducted with the towed-array recording continuously as the ship conducts line transects of an area of interest. Ideally, deployable sensors such as sonobuoys should be used to help assess changes in the local area over time, as surveys simultaneously sample time and space with a relatively short dwell time in any area. Along the same lines, surveys should ensure that multiple passes are made in the same area at different times of the day and during different seasons. Another method for improving temporal sampling would be to use a field of long-life bottom-moored recorders as part of a long-term monitoring plan, as suggested in [6]. 6.2 Sensors enabling advanced processing If towed sensors had a larger aperture (e.g. two sensor pairs spaced 100 m or more apart) more sophisticated processing methods might be used such as the time-difference-of-arrival (TDOA) approach used in [4] and [5]. This advanced processing may add the benefit of reduced falsealarm rate and improved knowledge of population density. 6.3 System/analysis tools and process development Methods and tools for analyzing acoustic data must be efficient, reliable, and robust in order for monitoring plans to be effective. Vast quantities of acoustic data must be transformed into useful information during processing. This section provides some recommendations on investment that would help ensure that the tools and processes are ready to support the overall objective Detector testing and improvement A proper assessment of the accuracy of the detectors configured for the various cetacean species requires running the detectors through multiple recordings known to have vocalizations for target species. The next stage in development of automated detectors is thus testing them out on additional biologically-rich data sets. This analysis will no doubt lead to a better understanding of the current detector and further ideas for improvement. 24

35 6.3.2 False alarm reduction Detailed analysis of the detections gave a better understanding of causes of false detections on the three employed systems. Analysts hypothesize that post-processing of the detections would eliminate many of the false detections, and in particular, writing software to do the following would substantially help speed up analysis: Omit detections within the first few minutes of system/sensor deployment, as acoustic data are often corrupted by deployment noise (e.g. sonobuoy hydrophone run-out). Omit detections that occurred within the time period that the vessel was approaching the system for retrieval, as nearby vessels cause significant transient noise and therefore false alarms. For a similar reason to the one above, run a vessel detector through the data to determine areas of relatively loud ship noise and omit detections occurring within that time period. Omit detections within the last few minutes of a sonobuoy recording, when the radio is noisy and the buoy is starting to self-scuttle. Omit any fin whale, sei whale, minke whale, sperm whale, or delphinid detections that occur without any other detections within a one minute period (e.g. if there is a single sperm whale detection then no other sperm whale detections within a minute before or after that, eliminate the detection). These species are expected to have a substantially higher vocalization rate than one pulse/upsweep/click per minute. As an alternative, more accurate but also more complex algorithm than the one presented in the previous bullet; for fin whale, sei whale, minke whale, sperm whale, and the delphinid detectors, only count detections if a minimum number of detections occur within a specific time period based on the expected vocalization rate of the target species (e.g. for sperm whales a minimum of five clicks occurring within a ten second period, half the expected detection rate for a single animal, in order for a detection to be counted). In other words, only detect trains of clicks and pulses, rather than individual pulses and clicks. Develop other species-specific detection filters to help remove false alarms based on expected vocalization patterns, such as the filter proposed above. For species that produce sets of vocalization, consider performing a more complex analysis of this set to determine if characteristics such as ICI are consistent with the targeted species. For this case, generation of sequential time-difference-of-arrival (TDOA) estimates will help determine if detections are consistent with a single animal, group of animals, or more random noise. This method would require a sensor array as described in [4] and [5]. Additionally, testing the detectors on other acoustic datasets that have a mix of cetacean vocalizations and structured noise would allow detector settings to be adjusted to maximize the number of present cetacean signals detected while minimizing the false alarm rate ACDC software application ACDC is becoming a very useful tool for live and post-analysis of acoustic data for detection and classification of cetacean vocalizations. However, there are various improvements that could be made to increase analysis efficiency. These include: 25

36 Have the sonogram window function more like OPD; Provide the ability to zoom in and out on the time axis; Provide the ability to zoom in and out on the frequency axis, or choose a specific frequency band to examine; Have the cursor move across time (i.e. track the time) when playing back audio for a file; and Provide the ability to playback audio for a specific or highlighted region of data. ACDC users would also benefit from a comprehensive user manual, similar to that for OPD Software sonobuoy GPS decoder Some sonobuoys are available with GPS and include NMEA-data modulated on the same carrier as for acoustic data. User of these data requires expensive hardware decoders that will only monitor on channel at a time. The modulation scheme is relatively simple (binary frequency shift keying [BFSK] and a software decoder that could run on recorded data for all channels simultaneously could be developed. Once completed, it could be provided to a number of units including the Acoustic Data Analysis Centres (ADAC) and the Canadian Forces Maritime Experimental and Test Ranges (CFMETR) Ambient noise monitoring This analysis reinforced the dynamic nature of acoustic sound levels. More work should be performed to assist with visualizing noise levels versus time. Different formats including compressed spectrograms or energy time indicator (ETI) graphs might be used. ETI might be generated for sub-bands. These formats may also reveal methods for compressing data to efficiently determine regions of frequent vocalization, as was performed in STAR for compression of data containing clicks. 26