Automated Classification of Dugong Calls and Tonal Noise by Combining Contour and MFCC Features

Abstract

To expand the spatial and temporal scales of passive acoustic monitoring of animals, automatically detecting target sounds among noises with similar acoustic properties is essential but challenging. In particular, the classification of tonal vocalisations and tonal noise remains a universal problem in bioacoustics research. The vocalisations of dugong, which is an endangered marine mammal that inhabits coastal seas, need to be monitored to enhance our understanding of its habitat use. However, detecting dugong tonal vocalisations is difficult due to the presence of tonal noise in the same frequency band. In this study, a classification method was developed for these signals to handle large acoustic data by reducing the labour required for manual inspection. Mel-frequency cepstral coefficients (MFCC) were extracted to characterise background sounds along with a few parameters of the signal contour, and a support vector machine was trained for binary classification. The classifier achieved an 84.4% recall and a 93.5% precision on the testing dataset even in a noisy shallow marine environment. This methodology enables the effective classification of dugong calls and similar tonal noises by combining contour and MFCC features and can extend the spatial and temporal scale of acoustic monitoring of the endangered dugong. This technique is potentially applicable to the monitoring of other endangered marine mammals that produce tonal vocalisations.