Vocalization based Individual Classification of Humpback Whales using Support Vector Machine

Visual surveys and DNA analysis hold primary importance in marine mammal observation. Especially, recognition of whales and their population density estimation rely heavily on these costly and tedious methods. Acoustic survey of vocalizing cetaceans, in contrast, is emerging as a promising technique for efficient, automatic, non-invasive and convenient observation and analysis of such vocally active organisms. However, it is, yet, too premature to replace the conventional methods altogether. In this paper, we have presented our results on recognition of individual humpback whales based on their vocalization data. Cepstral coefficients from song units extracted from audio records of seven humpback whales, re-sampled at 8KHz, were subjected to cepstral analysis. The extracted coefficients were used to develop multi-class support vector machine (SVM) classifier model. The test phase results indicated classification accuracy as high as 99% as compared to earlier best results of 88%, achieved by Gaussian mixture model (GMM) trained on cepstral coefficients. Furthermore, this improvement was attained with a highly reduced training dataset size. It was also verified that the song duration of approximately 128 seconds was sufficient for reliable identification in test phase for given dataset.