Automatic detection and classification of marine mammal tonal calls

Marine mammal tonal vocalizations are studied to help marine biologists get a better knowledge of their behavioral context. The vocalizations are often analyzed by characterization of the frequency contours. An automated frequency contour extraction method extended by image-based technique is proposed. The spectrogram is first de-noised and then two binary images are derived with a low threshold and a high threshold, respectively. Candidate contours can be regionally searched on the low threshold intensity image. The final output is confirmed by choosing the one with the nearest distance to the seed points on the high thresholding image. After obtaining the frequency contour, time-frequency parameters (TFPs) can be extracted for classification. The performance of the detection method has been tested by a set of recordings with the results well resolved especially for contours with sharp slopes. Classification results with TFPs achieved a mean accuracy over 92% for four species.