Acoustic Target Classification. John Horne, University of Washington

Transcription

1 Acoustic Target Classification Fred Mabel John Horne, University of Washington

2 Acoustic Measurements Amplitude (volts) 0 Surface Target Target Bottom Time ( seconds) Measure: amplitude f(frequency), elapsed time (everything else derived) How do you Discriminate, Classify, and Identify targets? D (wanted from unwanted), C categorize, I label

3 Acoustic Target Classification Decisions and Approaches Decisions: 1. Empirical, Experimental, Modeling 2. Scattering Region: Resonance, Geometric 3. Discrete Single or Multi Frequency, Wideband 4. Single Target, Ensemble Backscatter Approaches: 1. Prior Knowledge and Direct Sampling 2. Statistical Comparisons 3. Matching Models to Measures

4 Acoustic Classification Org Chart Empirical (in situ) Experimental (ex situ) Modeling frequency amplitude cage tether theoretical numeric TS S v L F q L F q L F

5 Ex Situ Target Strength Measurements species, length, frequency, tilt Target Strength (db) khz -40 Nakken & Olsen KRM Edwards & Armstrong Tilt (degrees) Henderson & Horne 2007

6 Empirical (in situ) Frequency Discrete Wideband Single F Amplitude Multiple F Pixel/ Voxel Image Analysis Acoustic Camera Angle Multi beam Inverse Active Tag Passive Resonance Peak Alternate Source Spectra metrics Visual ID metrics

7 Image Analysis: Digital Echogram Metrics - Goal: classify patches relative to environment - tabulate patch metrics - linear discriminant functions Nero and Magnuson 1989

8 Image Analysis: Shapes algorithm - schools module in Echoview Barange 1994

9 Burwen et al DIDSON/ARIS: Visual ID, Metrics

10 Multibeam Target Strengths Empirical measures Backscatter models Cutter & Demer 2007

11 Inverse Approach Multifrequency Measures+Inverse Algorithm+Backscatter Models Holliday 1977 Inverse algorithm Species/Size Class Resonance Peaks where S i is backscatter vector at frequency i, s ij is backscatter at frequency i and species/size class j, and ni is number of organisms at frequency i Critical needs: accurate scattering model, unique s bs for each species/size class

12 Empirical (in situ) Frequency Discrete Wideband Passive Resonance Peak Spectra Inverse Active Tag Alternate Source

13 Passive Acoustics: Hydrophones 1. Crustaceans (e.g. snapping shrimp) 2. Teleost fish with swimbladders (e.g. weak, red drum) 3. Marine mammals (e.g. whales, dolphins)

14 Passive Acoustics: Invertebrates Snapping Shrimp Au & Banks 1998

15 Passive Acoustics: Whale Tracking Whale Spectrograms IUSS Whale Track C. Clark, Cornell Univ

16 Resonance Peak Measures Depth (m) match resonance peaks to backscatter models Nero et al Current contributions from Stanton, Lavery, Jech

17 Wideband Frequency Spectra cod saithe, haddock - neural network & discriminant analysis horse mackerel mackerel Simmonds et al. 1996

18 Amplitude Single Frequency Multiple Frequency Single Echo Multiple Echo Single Echo Multiple Echo Target DTS Strength TS + sample Echo Envelope Aggregation Echo Envelope + environ Aggregation DTS PDF DPDF Amplitude P(Cat) Image Analysis Echo Envelope S v DS v PDF DPDF P(Cat) Ordination Classification

19 Target Strength & Direct Sampling Combine biological knowledge, net samples, and echogram patterns Sund 1935

20 bottom Echo Envelope Metrics Ensemble Backscatter PP volts 2 PT - linear discriminant functions school depth metres off Rose & Leggett 1988

21 Single f, Echo Envelope, + Enviro Shoal Depth L H Depth Perimeter Altitude - discriminant functions & PCA Scalabrin et al discriminant functions & neural networks Haralabous & Georgakarakos 1996

22 Multifrequency Classification & Regression Trees - schools module to define aggregations - aggregation & environmental metrics as covariates Fernandes 2009

23 Multifrequency Classification Multifrequency data Un-Supervised Semi-Supervised Supervised Data Only Some Reference Samples Many Reference Samples

24 Multifrequency Classification Multifrequency data Un-Supervised Semi-Supervised Supervised S v distribution Frequency differencing S v max amplitude Jech & Michaels Relative S v strength Frequency differencing DeRobertis & Ressler Goss et al. Anderson et al Kloser et al Korneliussen & Ona Woilez et al Sato et al

25 Depth (m) Depth (m) Probabilistic Backscatter Identification 3 Frequency Backscatter: small fish 18 khz 120 khz 200 khz krill walleye pollock Ping Number Backscatter (Sv, db) 5 Cluster Identification: fish dense schools krill Ping Number Membership Probability background ring down Anderson et al. 2007

26 Synthetic Echogram S v maximum amplitude Kloser et al. 2002

27 Frequency Response krill mackerel herring Korneiliussen and Ona (2007) mackerel herring horse mackerel boarfish Frequency (khz)

28 Semi-Supervised Classification K-means Clustering Woilez et al. 2012

29 MVBS Frequency Differencing Sato et al. 2015

30 Supervised Classification Training sets of pure samples All data Analyst-assigned categories Pollock Squid Euphausiids

31 Current Trends... as seen in renewed interest in resonance peaks as classifier - renewed interest in wideband for resolution and classifier (EK-80), community standards being established but not there yet - potential for machine learning/big data tools to become more prevalent - alternate platforms increasing spatial and temporal ranges but acoustic classification is an ongoing challenge - increased scrutiny from marine mammal community may constrain frequency range