Fig 1.
Overview of the workflow for a) training and b) applying a generalized and targeted species classification pipeline. Boxes indicate automated algorithms, in black are the standard steps between workflows and in gray is an additional step to delimitate data to a target group.
Fig 2.
Map showing the latitude and longitude locations of all passive acoustic monitoring sites. Sites used in the representative training dataset are shown in red; sites included in the case study dataset are shown in yellow. Site abbreviations in the western North Atlantic: Heezen Canyon–HZ, Oceanographer Canyon–OC, Bear Seamount–BR, Nantucket Canyon–NC, Babylon Canyon–BC, Wilmington Canyon–WC, Norfolk Canyon–NFC, Hatteras–HAT, Onslow Bay–ONB, Gulf Stream–GS, Blake Plateau–BP, Bermuda–BM, Blake Spur–BS, Jacksonville–JAX. Site abbreviations in the Gulf of Mexico: Mississippi Canyon–MC, Green Canyon–GC, Dry Tortugas–DT, Howell Hook–HH. Map land cover, shaded relief, and ocean-bottom relief were obtained from Natural Earth (http://naturalearthdata.com/).
Table 1.
List of labels assigned to classes representing signals from North Atlantic beaked whale species and other signal sources.
Class summary with abbreviation ID, description of the attributed class–species, taxonomic family, genus, or common descriptive features–and total number of bin-level clusters per class, median and interquartile range of detections per cluster, and sites where clusters were acquired. Site abbreviations: in the western North Atlantic, Heezen Canyon–HZ, Bear Seamount–BR, Nantucket Canyon–NC, Norfolk Canyon–NFC, Hatteras–HAT, Onslow Bay–ONB, Bermuda–BM, Blake Spur–BS, Jacksonville–JAX; in the Gulf of Mexico: Mississippi Canyon–MC, Green Canyon–GC, Dry Tortugas–DT, Howell Hook–HH.
Fig 3.
Signal classes formed using unsupervised clustering at a 5-minute bin level based on spectra, inter-click interval (ICI), and waveform envelope.
For each signal class, the top panels depict the mean and standard deviation of spectra among all clusters (top left) and concatenated mean cluster spectra (top right); the middle panels depict the mean and standard deviation of ICI distributions among all clusters (middle left) and concatenated cluster ICI distributions (middle right); and the bottom panels depict the mean and standard deviation of waveform envelops among all clusters (bottom left) and concatenated mean cluster waveform envelope (bottom right). Concatenated clusters have been sorted by peak frequency. Color map is normalized amplitudes on a scale from 0 (dark blue) to 1 (dark red). Refer to Table 1 for abbreviation IDs.
Fig 4.
Confusion matrix for the DNN classifier on the balanced test set.
Each class consists of 500 cluster examples formed at a 5-minute bin level. Values in the matrix show the total number of examples classified, on the right bar the recall rate, and on the bottom bar the precision rate. Refer to Table 1 for abbreviation IDs.
Table 2.
Confusion matrix for the targeted species classification pipeline with a hard negative filter on the case study dataset.
Values show the total number of 5-minute bins classified per class. Refer to Table 1 for abbreviation IDs. Manually labeled bins are the true class. Bins with no clear assignment are indicated with the class abbreviation in italic. The total number of false negative (FN) and false positives (FP), precision and recall rates are shown per species and in parentheses when only full-misclassified bins (full bin) are considered.
Fig 5.
Correlation between the total number of clicks per 5-minute bins manually labeled to a class (true class) and total number of clicks classified to predicted class by targeted species pipeline with a hard negative filter in the case study dataset.
Each point is a bin, and each subplot displays all bins manually assigned to a species, along with the class predicted by the DNN in color scale, with dark blue points representing true positives and other colored circles representing misclassified bins. The class abbreviation displayed in bold shows the true class of each subplot. Points along the diagonal show bins for which the number of manually labeled clicks were correctly predicted to a class. Above the diagonal, more clicks were predicted to a class than the true number of clicks per bin, whereas below the diagonal, less clicks were predicted than the total number of true clicks per bin.
Fig 6.
Case study dataset bins without a manual label classified by the targeted species classification pipeline with a hard negative filter.
For each predicted beaked whale class, concatenated mean cluster spectra (top panel), concatenated cluster ICI distributions (middle panel), and concatenated mean cluster waveform envelope (bottom panel) are shown. Concatenated clusters have been sorted by site and classification prediction probability score. Color map is normalized amplitudes on a scale from 0 (dark blue) to 1 (dark red). Many of the clusters counted as misclassifications appear to have been correctly classified by the network, but were not labeled in the manual dataset. Refer to Table 1 for abbreviation IDs.
Table 3.
Confusion matrix for all classification pipelines on the case study dataset at site WC.
Values show the total number of 5-minute bins classified per class organized by class and pipeline used: a targeted species classification pipeline with a hard negative filter (Th), or a moderate negative filter (Tm), and a generalized species classification pipeline (G). Manually labeled bins are the true class. Refer to Table 1 for abbreviation IDs. Bins with no clear assignment are indicated with the class abbreviation in italic. The total number of false negative (FN) and false positives (FP), precision and recall rates are shown per species and in parentheses when only full-misclassified bins are considered and those that are partially misclassified are excluded.
Fig 7.
Case study dataset bins at site WC without a manual label and classified by a targeted species pipeline with a moderate negative filter and a generalized species pipeline.
For each predicted beaked whale class, concatenated mean cluster spectra (top panel), concatenated cluster ICI distributions (middle panel), and concatenated mean cluster waveform envelope (bottom panel) are shown. Concatenated clusters have been sorted by site and classification prediction probability score. Color map is normalized amplitudes on a scale from 0 (dark blue) to 1 (dark red). Many of the clusters counted as misclassifications appear to have been correctly classified by the network, but were not labeled in the manual dataset. Refer to Table 1 for abbreviation IDs, and Fig 6 for bins without a manual label and classified by a targeted species pipeline with a hard negative filter.