Fig 1.
Example network representing relationships between echolocation clicks A-E.
Circles represent nodes and lines are edges representing the similarity between nodes. Edge weight values indicate the similarity between each pair of connected nodes; where 1 indicates that the two nodes are identical, and 0 indicates that they are unrelated. This network is synthetic for illustrative purposes.
Table 1.
Gulf of Mexico delphinids listed in order of estimated abundance according to NOAA stock assessments [22, 23].
Table 2.
Gulf of Mexico HARP training and testing set deployment periods, locations, and automated click detector results.
Site designations are: MC = Mississippi Canyon, GC = Green Canyon, DT = Dry Tortugas, DC = DeSoto Canyon, and MP = Main Pass. Counts and durations of click detections and encounters were computed after false positive detections had been removed. Five minute bins containing > = 100 click detections were included in the classification analysis. Bins containing more than 5000 clicks were subsampled. Bold rows indicate deployments used for classifier testing; all other rows represent deployments used for classifier training.
Fig 2.
Effect of edge pruning threshold (pe).
The effects of pe choice on Phase 1 clusters (subplots A-C) were evaluated on a subset of bins from site MC. Effects on Phase 2 clusters (subplots D-F) were evaluated by running 20 iterations of the CW routine at each threshold. Horizontal bars represent the normalized distributions of each measured parameter, with the gray + indicating the parameter mean. In both phases, increasing pe increased the mean number of clusters identified and the mean percentage of nodes. Measures of cluster purity (Phase 1: within-cluster similarity, Phase 2: NMI) also generally increased, with the exception of pe = 0.70 (NMI = 1 with a single cluster across all partitions and no isolation). A mid-range threshold of pe = 0.95 was selected for both phases to allow identification of multiple clusters without over-pruning or over-training.
Fig 3.
Normalized sound pressure level (SPL) spectra, concatenated spectra, and ICI distributions of the seven dominant click types (A-G) identified in the training datasets across five sites by the automated clustering routine. Each row of plots represents a distinct cluster in the final partition. Left: Normalized mean cluster received sound pressure level spectra (solid line) with 25th and 75th percentiles (dashed lines). N indicates the number of summary nodes included in each cluster. Center: Concatenated mean spectra of the summary nodes in each cluster. Color scale indicates relative amplitude in dB (red is high, blue is low). Colored bar across the top indicates the site from which the spectra below were extracted. Site/color pairs are: MC/Blue, GC/Green, DT/Yellow, DC/Cyan, MP/Magenta. Right: Distribution of modal ICIs.
Table 3.
Frequency and ICI statistics by click type.
Cluster size indicates the number of five-minute bins included in each click type cluster out of a set of 5,000 bins used for training. The mean of the modal (most frequently observed) ICI is computed across all five-minute bins in the cluster.
Fig 4.
Classification agreement between automated and manual classifications as a function of classification certainty score.
Upper plot: Number of bins tested manually per site at each classification certainty level. Site/color pairs follow the legend in the lower right plot. Lower plot: Percentage of matching automated and manual classifications as a function of classification certainty score based on a comparison with a subset of 200 manually classified test bins per site. Black line represents the percentage of matching classifications between the two methods across all sites combined.
Table 4.
Rates of occurrence of each click type (A-G) in test sets by site.
Numbers indicate the percentage of test click types assigned to each training cluster by the automated classification algorithm for each site. Types that were predominantly restricted to shallow sites (MP and DC) in the training dataset were infrequently identified at deep sites during testing, and vice versa.
Table 5.
Classifier confusion across all sites based on manual evaluation.
Rows indicate automatic classifications and columns indicate manual classifications, for a manually verified subset of 1000 test summary nodes across five sites. Values represent the number of instances of each combination. Values on the diagonal indicate cases of agreement between the two classification methods. “Unk” labels represent test click types that were labeled as unknown because match confidence was low (automatic classification) or because they did not clearly match a template cluster (manual classification).
Fig 5.
Towed hydrophone array click types.
Clustered clicks from towed hydrophone array recorded in the presence of (A) Pantropical spotted dolphin (Stenella attenuata), (B) Atlantic spotted dolphin (Stenella frontalis), (C) Pilot whale (Globicephala sp.), and (D) Risso’s dolphin, (Grampus griseus). B, C and D were recorded in the Atlantic. Left subplots: Mean normalized spectra levels (solid line) with 25th and 75th percentiles (dashed lines); n indicates the number of clicks retained in the final cluster. Center subplots: Concatenated spectra of all clustered clicks. Color indicates normalized received sound pressure level (SPL). Right subplots: Inter-click interval (ICI) distributions.
Fig 6.
Map of recording site locations in the Gulf of Mexico with green/brown representing land masses, and white/blue representing water.
The five sites are named based on the federal lease block in which they are located: Mississippi Canyon (MC), Green Canyon (GC), Dry Tortugas (DT), DeSoto Canyon (DC) and Main Pass (MP). Contours are at 500 m, 1000 m and 1500 m. Topographical data are from [38].
Table 6.
Towed hydrophone array recording locations, encounter dates, and click detection counts by species.