Feature Extraction Methods for the Detection of North Atlantic Right Whale Up-Calls

Abstract

A In this paper, a feature extraction method for detecting North Atlantic Right Whale (NARW) upcalls is proposed. The core of the scheme is an integration of Mel Frequency Cepstral Coefficients (MFCC) or Gammatone Frequency Cepstral Coefficients (GFCC) with Discrete Wavelet Transforms (DWT) for the extraction of features of NARW up-calls. Several types of wavelets are tested in terms of detection accuracy. Once the up-call features are extracted, popular classifiers such as Support Vector Machines and K-Nearest Neighbors are applied to classify call types. Detection results show that the upcall detection rate by using Mel Frequency Cepstral Coefficients is 73.82%, and False alarm rate is 2.48%. However, the upcall detection rate and false alarm rate are improved to 92.27% and 1.48% when these are integrated with Discrete Wavelet Transforms for feature extraction. GFCC works better than MFCC as a feature extractor if it is used alone. When both are integrated with DWT, the advantage of the former disappears. Furthermore, it is shown through experimental studies that the proposed detection scheme is superior to those obtained using spectrogram-based techniques in terms of both detection accuracy and speed. Due to its effectiveness and efficiency, the proposed algorithm can potentially be implemented in auto-detection buoys.