PorCC: a new high-accuracy click classifier to study harbour porpoises in the wild

Cosentino, Melania and Guarato, Francesco and Tougaard, Jakob and Nairn, David and Jackson, Joseph C. and Windmill, James F. C. (2019) PorCC: a new high-accuracy click classifier to study harbour porpoises in the wild. In: The 15th Danish Marine Mammal Symposium, 2019-01-23 - 2019-01-25.

[img]
Preview
Text (Cosentino-etal-DMMS2019-PorCC-a-new-high-accuracy-click-classifier-to-study-harbour-porpoises)
Cosentino_etal_DMMS2019_PorCC_a_new_high_accuracy_click_classifier_to_study_harbour_porpoises.pdf
Final Published Version

Download (851kB)| Preview

    Abstract

    Harbour porpoises (Phocoena phocoena) are difficult to observe at sea, even with good weather conditions, due to their small size and cryptic behaviour. However, they are highly vocal, producing narrow-band high-frequency (NBHF) echolocation clicks, which makes them well suited for passive acoustic monitoring (PAM). Such PAM systems must be coupled with a classification algorithm to identify the likely porpoise signals among other transient signals. We present a new harbour porpoise click classifier (PorCC) for full-waveform signals, with an improved performance over current classifiers. PorCC was developed in MATLAB and uses the coefficients of two logistic regression models in a decision-making pathway to assign each signal to one of three categories: high-quality click (HQ), low-quality click (LQ), or high-frequency noise (N). The first model uses click duration and QRMS (RMS-bandwidth / centroid frequency) to separate HQ from N. The second model uses click duration, QRMS, ratio between peak and centroid frequency, peak cross-correlation coefficient (against a model click), centroid frequency, and -3dB bandwidth to separate LQ from N. PorCC achieved hit rates > 90% for HQ clicks while keeping false alarm levels < 1%. The performance of PorCC was compared to that of PAMGuard’s Porpoise Classifier module (with standard settings) and the receiver operating characteristics curve was generated for both classifiers. The precision for HQ (HQ clips classified as HQ / total clips classified as HQ) was 31.2% for PAMGuard and 96.1% for PorCC, and the detectability index (d’) was 2.2 for PAMGuard versus 4.1 for PorCC. Results of this study show PorCC is a rapid, highly accurate method to classify NBHF clicks, which could be applied for real time monitoring, as well as to study harbour porpoises, and potentially other NBHF species, throughout their distribution range from data collected using towed hydrophones or static recorders.