ROCCA (Real-time Odontocete Call Classification Algorithm)
Researchers and software engineers from Bio-Waves, Inc. have developed a whistle classification program called ROCCA, which is currently available as a module in PAMGuard. ROCCA was developed for the identification of delphinid species based on their whistles. ROCCA can be used to extract whistles and tonal calls from spectrograms using either a semi-automated method (ROCCA’s ‘pick points’ function) or a fully automated method (PAMGuards ‘whistle and moan detector’ module). After whistle contours have been extracted, ROCCA can be used to automatically measure 50 acoustic features, including frequencies, slopes, duration, and the number and position of inflection points and steps. The measured features are used as inputs to one of several random forest classifiers developed for species identification. At present, Bio-Waves has developed three classifiers for dolphin whistles from two large geographic regions:
- A Tropical Pacific classifier: Trained using whistles recorded in the tropical Pacific Ocean. Includes eight species (bottlenose dolphins, Tursiops truncatus; striped dolphins, Stenella coeruleoalba; spinner dolphins, S. longirostris; pantropical spotted dolphins, S. attenuata; rough-toothed dolphins, Steno bredanensis; common dolphins, Delphinus spp.; short-finned pilot whales, Globicephala macrorhynchus;and false killer whales, Pseudorca crassidens.)
- An Atlantic whistle classifier (semi-automated): Trained using whistles recorded in the northwestern Atlantic Ocean. Includes five species (bottlenose dolphins; Atlantic spotted dolphins, Stenella frontalis; short-beaked common dolphins, Delphinus delphis; striped dolphins; and short-finned pilot whales). Should be used when processing northwestern Atlantic data using semi-automated methods.
- An Atlantic whistle classifier (automated). This classifier was trained using the same whistles as the semi-automated classifier, however the whistles in this training dataset were detected and extracted automatically using PAMGuard’s whistle and moan detector. This classifier should be used when processing northwestern Atlantic data using PAMGuard’s whistle and moan detector.