Large-scale, Long-term Acoustic Surveys of Marine Mammals

Transcription

1 Large-scale, Long-term Acoustic Surveys of Marine Mammals David K. Mellinger Oregon State University and National Oceanographic and Atmospheric Administration

2 Overview Visual and acoustic marine mammal surveys Techniques for acoustic surveys data collection acoustic analysis density estimation Results and applications mid-atlantic minke whales Pacific blue whales Gulf of Alaska sperm whales Iceland right whales Current and future work research at the OSU/NOAA lab

3 Overview Visual and acoustic marine mammal surveys Techniques for acoustic surveys data collection acoustic analysis density estimation Results and applications mid-atlantic minke whales Pacific blue whales Gulf of Alaska sperm whales Iceland right whales Current and future work research at the OSU/NOAA lab

4 Marine Mammal Surveys Why? basic research management mitigation What is measured? population - absolute, relative, or trend distribution - daily, seasonal, and year-to-year movement Laurinolli & Hay 2004

5 Marine Mammal Surveys Surveys traditionally done visually ship-based line transect surveys aircraft surveys shore counts

6 Visual Surveys Limitations small area -- sighting distance short time span -- while on survey daylight animals must be at surface affected by fog and haze, sun angle, swell, wind waves, observer variability,... labor-intensive -- generally animals are counted one at a time

7 Acoustic surveys Why use sound? good mode of long-range information transmission used by all (known) marine mammals can travel quite long ranges

8 Acoustic surveys "Passive" acoustic surveys passive: no sonar ping rely on sound the animal produces listen for sounds, identify call signatures (sometimes) locate animals estimate population density

9 Acoustic surveys Advantages of acoustic surveys: large areas - potentially, ocean basins long time spans - months to years - can be continuous: 24 hours/day, 365 days/year automatic data processing - can count large numbers of animals - constant bias not subject to fog, sun, waves, observer variability, daylight, surfacing behavior,... in comparisons, acoustic surveys detect times as many individuals

10 Acoustic surveys: Disadvantages Animals have to breathe, but they don't have to make sound but they DO make sound if they want to reproduce, they have to make sound (except perhaps coastal species) even most coastal species make loud sounds underwater Animals make variable amounts of sound again, there are useful exceptions - foraging gsounds - reproductive displays

11 Acoustic surveys: Disadvantages We don't know how many animals make sound Calibrate key data: rate and sound level of vocalizations via acoustic tags on animals via joint visual and acoustic surveys - ==> (more-or-less) independent estimates of number of animals and vocalizations - can deduce number of animals present Or derive relative measures or trends

12 Overview Visual and acoustic marine mammal surveys Techniques for acoustic surveys data collection acoustic analysis density estimation Results and applications mid-atlantic minke whales Pacific blue whales Gulf of Alaska sperm whales Iceland right whales Current and future work research at the OSU/NOAA lab

13 Overview Visual and acoustic marine mammal surveys Techniques for acoustic surveys data collection acoustic analysis density estimation Results and applications mid-atlantic minke whales Pacific blue whales Gulf of Alaska sperm whales Iceland right whales Current and future work research at the OSU/NOAA lab

14 Overview Visual and acoustic marine mammal surveys Techniques for acoustic surveys data collection - autonomous hydrophones - cabled hydrophone arrays - ship-based arrays and sonobuoys acoustic analysis density estimation Results and applications mid-atlantic minke whales Pacific blue whales Gulf of Alaska sperm whales Current and future work research at the OSU/NOAA lab

15 Data collection Autonomous hydrophones typically in deep sound channel, in arrays of 5-6 independent instruments currently deployable for one year with a 2 khz sampling rate constraint: battery life syntactic foam float autonomous hydrophone cable sections acoustic release anchor

16 Data collection and analysis Ocean glider collaboration with UW-APL team mobile buoyancy-driven autonomous platform equipped with hydrophones recording system onboard call detection satellite communication principal constraint is energy 1.8 m Diameter 30 cm Battery pack

17 Data collection and analysis QUEphone Haru Matsumoto buoyancy-driven autonomous platform immobile (more or less) equipped with hydrophones recording system onboard call detection satellite communications principal constraint is energy

18 Data collection Shore-cabled hydrophone arrays U.S. Navy SOSUS IOOS networks

19 Data collection Ship-based trackline recording towed hydrophone arrays and directional sonobuoys P. Folkens

20 Comparisons: Methods for marine mammal detection ti 10 years cabled hydrophone array ime 1 year autonomous hydrophone Dep ployment ti 1 month 1 week towed hydrophone array visual survey, active acoustics satellite tag 1 day acoustic recording tag Distance scale, km

21 Comparisons: Methods for marine mammal detection ti active acoustics 0.01 Positional ac ccuracy, km acoustic recording tag satellite tag visual survey towed hydrophone array 2 cabled hydrophone arrays 1 autonomous hydrophone 3 autonomous hydrophones cabled hydrophone array Number of individuals surveyed 10000

22 Overview Visual and acoustic marine mammal surveys Techniques for acoustic surveys data collection acoustic analysis density estimation Results and applications mid-atlantic minke whales Pacific blue whales Gulf of Alaska sperm whales Iceland right whales Current and future work research at the OSU/NOAA lab

23 Overview Visual and acoustic marine mammal surveys Techniques for acoustic surveys data collection acoustic analysis - manual call counting - automatic call recognition density estimation Results and applications mid-atlantic minke whales Pacific blue whales Gulf of Alaska sperm whales Current and future work research at the OSU/NOAA lab

24 Acoustic analysis Large data sets, in both duration and data volume Examples: Alaska 1-year whale survey: 12 hydrophone-years years of sound Monterey Bay 1-year harbor seal project: 2 years of sound Eastern Tropical Pacific autonomous hydrophone recordings (6 years): 38 hydrophone-years of sound Need automated processing

25 Automatic ti call recognition A type of acoustic pattern recognition Usually based on a spectrogram representation S ( t, f ) k 1 n 0 x ( t n ) w ( n )exp( i 2 fn / k ) fr requency, Hz x10 time, s P. Folkens

26 Automatic ti call recognition Methods matched filter - template matching in the sound signal - well-known digital signal processing technique spectrogram correlation - template matching in a spectrogram representation neural network statistical pattern discrimination recognition of timing patterns

27 Automatic ti call recognition Methods matched filter - template matching in the sound signal - well-known digital signal processing technique spectrogram correlation - template matching in a spectrogram representation neural network statistical pattern discrimination recognition of timing patterns

28 Bowhead whale call detection ti Problem: Detect bowhead whale (Balaena mysticetus) sounds sounds recorded during a combined acoustic & visual census bowhead sounds had been analyzed manually needed to speed up the analysis Decided to detect end-notes, which... occur in a repeating sequence at end of each song are somewhat similar from year to year have frequency contours; detection method should be useful for other species call types Data set: bowhead songs from two early censuses, 1986 and 1988, conducted near Barrow, AK

29 Bowhead whale call detection ti End-note examples P. Folkens Hz 1x 2x s Hz Hz s s

30 End-note detection ti Challenges: interfering sounds are common in this dataset - from other whale species, ice cracking, bearded seals, cable strumming,... notes vary from one instance to the next... and from one whale to the next wanted to develop methods that did not require a large training set - less time commitment - methods useful for other species, many of them little-recorded

31 End-note detection ti Approach: detect frequency sweeps Detection method based (loosely) on responses of mammalian auditory system neurons frequency sweeps in a certain frequency band... of a certain sweep rate Method: synthetic-kernel spectrogram correlation synthetic kernel to improve response to noise

32 Spectrogram correlation in a nutshell Hz S ( t, f ) k( t, f ) Hz s s = s (t)

33 Spectrogram correlation Cross-correlation is done only in time, not in frequency result is detection function (t) Can determine detection events by setting a threshold The key is in kernel design: characteristics of sweeps are measured from a small number of example calls each sweep in the kernel is synthesized, s then all are combined kernel has + and flanking regions to give good performance when interference is present width of + region affects response to call variability and noise Hz

34 Spectrogram correlation Comparison with other detection methods: matched filter back-prop neural net (80 input units from spectrogram, 15 hidden units, trained on half the input data) test data: bowhead whale sounds noise and interference sounds

35 Handling variability Variability in frequency can be handled by 2-D cross-correlation, or limited 2-D for speed take maximum over all frequencies correlated Variability in sweep rate can be handled to some extent by changing width of kernel Variability in timing i of call segments can be a problem; timing between parts of the call changed frequently

36 Detecting ti timing i patterns: Minke whale ACR Problem: Detect minke whale (Balaenoptera acutorostrata) sounds recordings made in Feb by autonomous hydrophones near the Mid-Atlantic Ridge study of minke seasonal distribution Minke pulse train sound: 50 P. Folkens frequ uency, Hz x :41:10 20:41:15 20:41:20 20:41:25 20:41:30 20:41:35 20:41:40 20:41:45 20:41:50 20:41:55 time C:\Dave\demos\MAR99 CE disk1 datafile /3x/0.125/Hamming

37 Minke whale automatic ti call recognition Method: rely on repetitive nature of the pulses precondition spectrogram with Weiner filter to equalize background noise, remove tonal sounds such as ships at each time t, sum the energy e(t) in the Hz band use summed autocorrelation to recognize repetitive pulses: ( t) max i N / i k 1 t t t e( t) e( t ki) dt look kfor peaks in the detection function (t( )

38 Minke whale automatic ti call recognition frequency detection function (t) time

39 Overview Visual and acoustic marine mammal surveys Techniques for acoustic surveys data collection acoustic analysis density estimation Results and applications mid-atlantic minke whales Pacific blue whales Gulf of Alaska sperm whales Iceland right whales Current and future work research at the OSU/NOAA lab

40 Density estimation Estimate population size or population density of a species within a defined darea Current standard methods: Visual line transect relies on seeing the animals Mark-recapture relies on marking (tagging) animals, or their having natural markings Many species produce frequent loud vocalizations (e.g., communication calls; echolocation clicks): hard to see but easy to hear Use passive acoustics to estimate density

41 Density estimation: Strip transectst animal population density D n a number of detected animals area of surveyed region (line length x strip width)

42 Density estimation: Line transectst Line transects: a type of distance sampling survey D n ap proportion of animals in surveyed region detected: average probability of detection

43 Density estimation: Estimating p if you saw everything at all distances, on average the histogram bars should be here count these animals are estimated to have been missed perpendicular distance from line, x pˆ area under curve area under rectangle

44 Density estimation Some complications: Some proportion of detections are false positives (estimate from manual analysis of a sample of data) D n(1 c) apg 0 Only a proportion of animals are vocalizing as the ship passes within range (estimate from, e.g., a tagging study or focal follow)

45 Density estimation: Array of fixed hydrophones Requirements: each sensor operates independently can get distance to detected objects (or direction if can assume a depth or depth distribution) If you impose a set of T snapshot moments then you can use point transect methods another type of distance sampling D n(1 c) Tapg 0 area surveyed circle around each sensor probability of detection within surveyed circles can use distances to estimate

46 Density estimation: Cue counting number of cues (e.g., calls or echolocation clicks) detected during time period T false positive proportion for detected cues nc ( 1 c) D Tapr time period cue production rate (e.g., number of echolocation clicks per unit time) probability of detecting a cue

47 Density estimation example Example: Blainville s beaked whales in Bahamas (Marques et al. 2009, J. Acoust. Soc. Am.) Image: Diane Claridge vertical off-axis angle distance horizontal off-axis angle izontal off-axis angle hor distance ve ertical off-axis angle distance Northing (km m) Easting (km) Result: 25.3 animals per 1000 km 2

48 Overview Visual and acoustic marine mammal surveys Techniques for acoustic surveys data collection acoustic analysis density estimation Results and applications mid-atlantic minke whales Pacific blue whales Gulf of Alaska sperm whales Iceland right whales Current and future work research at the OSU/NOAA lab

49 Results: Mid-Atlantic minke whales 100 NW 100 NE CW 100 CE SW 100 SE day in 1999 day in 1999

50 NOA AA/PMEL A A B C B May Jul Sep Nov Jan Mar May NE Pacific blue whale calls D C May Jul Sep Nov Jan Mar May Chilean blue whale calls Antarctic blue whale calls May Jul Sep Nov Jan Mar May Fin whale calls Stafford et al. 1999

51 Results: Sperm whales N 147 W 53 N 157 W O N D J F M A M J J A S O N D J F M A M O N D J F M A M J J A S O N D J F M A M N 140 W O N D J F M A M J J A S O N D J F M A M N 153 W O N D J F M A M J J A S O N D J F M A M N 149 W O N D J F M A M J J A S O N D J F M A M N 135 W N 145 W O N D J F M A M J J A S O N D J F M A M O N D J F M A M J J A S O N D J F M A M km Mellinger et al. 2004

52 Results: Right whales One-year passive acoustic survey near Iceland right whales existed historically i in this area have been seen only twice here in last 50 years Mellinger et al. 2011

53 Mid-Atlantic Ridge airguns w ith airg uns N024W 0 airguns N025W Jan-99 % days with airguns Jan-00 Jan-01 Jan-02 Jan-03 Jan-04 Jan-05 Jan-06 Jan-07 Jan-08 % days with airguns Jan-99 Jan-00 Jan-01 Jan-02 Jan-03 Jan-04 Jan-05 Jan-06 Jan-07 Jan-08 Jan-99 Jan-00 Jan-01 Jan-02 Jan-03 Jan-04 % d a y s w % days with Jan-05 Jan-06 Jan-07 Jan-08 0 Jan-99 Jan-00 Jan-01 Jan-02 Jan-03 Jan-04 Jan-05 Jan-06 Jan-07 Jan-08 39N034W 37N029W Jan-99 Jan-00 Jan-01 Jan-02 Jan-03 Jan-04 Jan-05 Jan-06 Jan-07 Jan-08 Jan-99 Jan-00 Jan-01 Jan-02 Jan-03 Jan-04 % days with airguns % days with airguns Jan-05 Jan-06 Jan-07 Jan-08 32N035W Jan-99 Jan-00 Jan-01 Jan-02 Jan-03 Jan-04 Jan-05 Jan-06 Jan-07 Jan-08 Jan-99 Jan-00 Jan-01 Jan-02 Jan-03 Jan-04 % days with airguns Jan-05 Jan-06 Jan-07 Jan-08 32N043W % days with airguns 26N050W 26N040W Jan-99 Jan-00 Jan-01 Jan-02 Jan-03 Jan-04 Jan-05 Jan-06 Jan-07 Jan-08 Jan-99 Jan-00 Jan-01 Jan-02 Jan-03 Jan-04 % days with airguns % days with airguns Jan-05 Jan-06 Jan-07 Jan-08 16N049W 16N043W Jan-99 Jan-00 Jan-01 Jan-02 Jan-03 Jan-04 Jan-05 Jan-06 Jan-07 Jan-08 Jan-99 Jan-00 Jan-01 Jan-02 Jan-03 Jan-04 % days with airguns % days with airguns Jan-05 Jan-06 Jan-07 Jan-08

54 Hawaii beaked whales (glider test) t) Beaked whale detections, Oct. Nov. 2009

55 Hawaii beaked whales (glider test) t) Sperm whale

56 Overview Visual and acoustic marine mammal surveys Techniques for acoustic surveys data collection acoustic analysis density estimation Results and applications mid-atlantic minke whales Pacific blue whales Gulf of Alaska sperm whales Iceland right whales Current and future work research at the OSU/NOAA lab

57 Locations of OSU/NOAA acoustic research 60 E 120 E W 60 W N N 30 N N 0 30 S S 60 S 60 S km E 120 E W 60 W 0 GMT 2007 May 14 23:50:14 OMC - Martin Weinelt

58 Where next? By species: blue fin humpback minke right bow. sperm beaked North Pacific SOSUS X X? X Eastern Tropical Pacific X X? Gulf of Alaska X X X? X X X X Mid-Atlantic Ridge X X X X Bering Sea X X X X X Beaufort Sea X Hawai i i X Australia X X X Other regions and species... Antarctic several balaenopterids Antarctic many delphinids S. California many delphinids

59 Where next? Technical development: Better recognition techniques - handle more variability (e.g., humpbacks) - recognize tonal sounds better for delphinids - develop a library of detectors Better software for real-time surveys - Ishmael: for acoustic localization, call detection, recording, analysis,... - long-term spectra - software for checking detections Density estimation - calibrate detection rates / SNRs - acoustic propagation modeling to estimate probability of detection

60 Where next? Study whale distributions with respect to physical and biological ecosystem components El Niño effects, PDO effects to what extent do eddies and rings affect whale movements? habitat studies Study whale distributions as affected by anthropogenic factors... ship noise, seismic air gun sounds, other sound sources fishing pollution

61 The End K. Mullin

62 Passive acoustic data collection: software Ishmael was used here for detecting, locating, and tracking sperm whales acoustically

63 Acoustic localization li in Ishmael Select a sound, click a button to locate it Can also use automatic click detection Phone-pair method (2 phones) produces a bearing to the animal for hydrophone arrays towed from ships, it s ambiguous whether the bearing is left or right of the ship Frequency beamforming method (>= 3 phones) also produces a bearing can work better for faint sounds

64 Acoustic tracking in Ishmael A track is just a series of localizations over time time

65 ACR: case study Problem: extract harbor seal (Phoca vitulina) roar sounds from a large archive of recorded data distinguish seal roars from everything else, extract them extracted roars were used for distribution study and behavioral analysis ~1 year x 2 hydrophones x samples/s 1.4 Tbyte

66 ACR: Sound examples Hz Hz s s Harbor seal roar Wave breaking

67 ACR approach: two-stage t system Stage 1: Fast but crude used on recording archive used simple time-frequency characteristics extracted detected sounds for Stage 1 Stage 2: Slow but accurate operated on short sound segments used more detailed analysis determined which segments were seal roars 2 2

68 Stage 1 : Detect t calls in large archive Pre-conditioned the spectrogram eliminated tonal noise - boat sounds - 60 Hz electrical hum normalized background noise level - also removed wind noise and most flow noise removed snapping shrimp noise Detected seal roars by measuring ratio of energies in khz band and khz band minimum duration of 0.5 s required for detection - eliminated most impulsive noises

69 Stage 1: Detection ti example s

70 Stage 1: Preliminary i results Ran detector on ~300 days of sound Extracted ~12 days of sound segments Speed: ~24x real time on P

71 Stage 2: Classification Operated on the short sound segments from Stage 1 each segment possibly has a roar goal is to distinguish roars from non-roars Operated on the spectrogram representation pre-conditioned the spectrogram measured acoustic features built classifier

72 Stage 2: Pre-conditioning i example

73 Stage 2 : Feature estimation Wanted estimators that were insensitive to fading slow decay of call in T or F insensitive to SNR continuous without yes/no decisions not dependent on time alignment insensitive to echoes and other multipath effects relevant to animal call types - frequency contours (with harmonics) and noise bursts preferably related to obvious features of the sounds Came up with 68 estimators some from AcouStat (Fristrup 1992)

74 Stage 2 : Classification Needed to separate the feature vector groups in 68 Used Fisher linear discriminator: calculates V, the projection that best separates the groups in the mean-squared-error error sense Training data: each sample was a segment from Stage roars, 1011 non-roars in training set roar/non-roar determination made by me selected from throughout the year, day, and tidal cycle in-band SNR: 5-20 db

75 Stage 2 : Classification v 0

76 Stage 2 : Classification: choosing v 0 Now have two groups, each one more pure Do another classification on each part Recurse to build a classifier tree each leaf of the tree is pure: all roars or all non-roars tree was then pruned to improve its generalization ability

77 Stage 2 : Classification