Spectrogram of a typical dolphin whistle showing the effect of the different processing stages of whistle contour extraction.
Distributions of three parameters (mean frequency, frequency slope, and curvature) describing whistle fragments for four species of odontocete (bottlenose dolphin—BD, common dolphin—CD, killer whales—KW, and Rissos's dolphin—RD).
Parameters describing the distribution of parameters from accumulations of 50 whistle fragments. From distributions of the type shown in Fig. 2 , the mean, standard deviation, and skew of each distribution is measured. By repeating this process, these new distributions are built. Parameters from these nine distributions are used in the linear discriminant classifier.
Whistle detection precision and recall for minimum whistle signal to noise ratios of 8 and 10 B and varying detection thresholds. Precision is the percentage of detected calls which matched human detection. Recall is the percentage of human detected calls which were automatically detected.
Classification success rates for (A) varying fragment lengths (averaged over section lengths between 40 and 100 fragments); (B) varying section lengths (averaged over fragment lengths between 30 and 50 bins).
Odontocete species used for classifier training and testing. Each species has been assigned to one or more geographic regions in which that species is known to occur. Note however that recordings may not have actually been collected from all of those different regions.
Detection results for whistles with a SNR > 10 dB for one-third of their length using a detection threshold of 7 dB. Results are derived from a comparison with visually annotated whistles. Precision is the percentage of detected calls which matched human detection. Recall is the percentage of human detected calls which were automatically detected. muDev is the average deviation in hertz from the human detected contour. Coverage is the percentage of the human detection included in the automatic detection and fragmentation is the mean number of automatic detections corresponding to each human detection.
Classifier Confusion Matrix for four species of odontocete found in the Polar Atlantic Region. Numbers on the diagonals represent the percentage of correctly classified sections of data. Off diagonal entries represent misclassifications. For example, 92.3% of beluga (BEL) sections were correctly classified, with 6.5% being misclassified as pilot whales (LPLT), 1.2% as white beaked dolphin (WBD) and none as white-sided dolphin (WSD). Numbers in parentheses are the standard error on the estimates of correct classification rates.
Confusion matrix for eight species of odontocete found in the Atlantic Frontier Region. (See Table I for species code definitions).
Confusion matrix for 11 species of odontocete found in the Gulf of Mexico. (See Table I for species code definitions).
Confusion matrix for 12 species of odontocete found in the Tropical Atlantic. (See Table I for species code definitions).
Summary of mean correct classification rates by region and the number of species used in the classifier dataset.
Article metrics loading...
Full text loading...