Automated detection and localization of bowhead whale sounds in the presence of seismic airgun surveys
(Color online) Stereographic projection of Directional Autonomous Seafloor Acoustic Recorders (DASARs) deployment geometries in the Beaufort Sea, 2008. The DASARs at each site are roughly 7 km apart. The deployments in other years are similar, except that the five unlabeled DASARS arranged as a trapezoid near site 1 are absent. GPS tracks of the seismic vessel Gilavar are shown whenever airguns are firing between September 8 and October 8, 2008.
(Color online) Spectrograms of bowhead and airgun sounds from the Beaufort Sea (256 pt FFT, 90% overlap): (a) two bowhead whale calls covering different frequency ranges during relatively high ambient noise conditions, DASAR 2g, 2008; (b) downswept whale call with harmonics, multipath, and reverberation, DASAR 2f, 2008; (c) “n-shaped” bowhead whale call with harmonics, DASAR 2f, 2008; (d) same call as (b), recorded on DASAR 2a, 2008;(e) distant airgun signal, DASAR 5g, 2008; (f) distant airgun signal, DASAR 2e, 2009; (g) strong airgun signal, DASAR 3g, 2008, generated at 21 km range. Intensity levels are in units of dB re 1 μPa2/Hz.
(Color online) (a) Airgun examples from three simultaneous seismic surveys in the Arctic, DASAR 5g, 2008, visible at 3, 7.5, and 13 s. (b) Examples of bearded seal and other biological signals, DASAR 2e, 2009. Intensity levels are in units of dB re 1 μPa2/Hz.
(Color online) Schematic of automated detection, classification, and localization algorithm for one site.
(Color online) Key steps in the image processing stage, demonstrated on the harmonic whale call in Fig. 2(c), along with an airgun pulse after 4 s: (a) Original spectrogram (256 pt FFT, 90% overlap); (b) “ridge image” of equalized spectrogram using 13 dB ridge threshold; (c) “contour image” of equalized spectrogram using 10 dB SNR contour threshold; (d) labeled ridge image after morphological opening, closing, and connected-component labeling; (e) contour image after morphological opening and closing; (f) final “transients” indicated by the large white boxes, comprised of ridge segments linked using the methods described in Secs. ??? and ???. Note how the harmonics of the bowhead whale call at 2 s have been successfully “linked.”
(Color online) Same as Fig. 5, but with steps performed on the strong airgun signal shown in Fig. 2(g). The small white box in (d) highlights a ridge component that would have been flagged as a bowhead call had it not been linked to the main pulse arrival through the shared contour segment in (e). The small white box in (f) indicates a ridge segment that was not linked to the airgun pulse via a common shared contour segment in (e).
(Color online) Distributions of the duration (s) and minimum frequency (Hz) of 141 796 whale call feature vectors obtained in 2008 from all sites, used to train the first neural network. (a) Marginal distribution for total minimum frequency; (b) 2D distribution of both parameters, with the intensity scaled to the log number of calls in a bin; (c) marginal distribution for total duration.
(Color online) Same as Fig. 7, but showing the distributions of 1 153 506 feature vectors not associated with whale calls, used to train the first neural network. The first bin in (c) has about 610 000 samples, but the y-axis is limited to 200 000 samples to improve the visibility of the rest of distribution.
(Color online) Progression of the bulk 2008 processing through each automated stage. Each subplot (a) through (e) represents sites 1–5, respectively. The output of each stage for all DASARs at a given site has been added together. The vertical extent of each color/shade indicates the number of events that have been removed on that date by the corresponding stage. Substantial local seismic survey activity took place near sites 3 and 4 between September 18 and 28. Starting from the top, the colors/shades show detections removed by interval filter (gray), image processing stage (orange), neural network stage (yellow), cross-DASAR matching (cyan), and localization (light blue). The dark blue area (bottom shaded area) shows final call counts.
(Color online) Comparison between manual analyses and fourth- and fifth-stage automated results, conducted on dates listed in Sec. V A. Data from 2007 (circles), 2008 (squares), 2009 (triangles), and 2010 (diamonds) are shown. Manual analyses from 2008 and 2009 were used to train portions of the automated classifier. The solid small blue symbols indicate the fourth-stage (neural network) automated performance, expressed in terms of excess call fraction (also known as “false discovery rate,” or one minus the “precision”) and missed call fraction (also known as one minus the “recall”), using default network thresholds discussed in Sec. V A. The large solid symbols indicate the corresponding performance of the fifth-stage (cross-DASAR matching) automated results, expressed in terms of excess call set fraction and missed call set fraction, using default network thresholds. The curves connecting hollow symbols show the neural network performance using optimized network thresholds, as discussed in the text. Subplots are comparisons between (a) all sites; (b) site 3 only (close to airgun surveys in 2007 and 2008); and (c) site 5 only (distant from airgun surveys). The legend indicates the number of individual call detections obtained by manual analyses for each year.
(Color online) Top: Spatial distribution of 19 125 whale call locations obtained by manual analysis of the last 12 h of 8 non-contiguous days in 2009. Bottom: distribution of 18 063 whale call locations computed by automated detection algorithm over the same time periods. No manual data from the top subplot were used to train the software.
Statistics of whale call locations for manual and analyzed positions shown in Fig. 11. Left column: Distribution of number of DASARS used to localize calls in (a) manual and (b) automated analysis. The “no sol” category indicates percentage of call sets that yield no localization solution (e.g., no crossed bearings, or failure to obtain bearings from enough DASARS). Right column: Distribution of area of 90% confidence ellipse for call locations, expressed in terms of radius of a circle of equivalent area, for (c) manual and (d) automated analysis. Italicized x-axis labels are in units of km; otherwise, units are meters. The right column only uses positions derived from two or more DASARs.
(Color online) Detailed comparison between manual and automated call set spatial distributions for site 2, August 28, 2008, between midnight and 4 a.m. (a) Spatial distribution of manual call sets that match an automated call set. A “match” between an automated and manual call set occurs when they share at least two DASARs in common. (b) Spatial distribution of automated call sets that match the call sets shown with (a). There are more locations plotted here than in (a) because a manual call set encompassing four or more DASARs may match two two-DASAR automated call sets. (c) Spatial distribution of “excess” automated call sets that do not match any manual call set; (d) manual call sets that do not match any automated call sets.
Significant dates discussed in paper.
Parameters and values used in the algorithm.
Feature inputs into neural networks.
Article metrics loading...
Full text loading...