An acoustic bio-metric for sperm whales

Acoustic identification of sperm whales, or sperm whale groups, has both a biological and practical use. The biological Interest is especially in the ability 10 separate recordings with mixed vocalisations to reconstruct the individual time series. This allows to follow single animals during their dive and to track their activity. Practically, for passive detection, localisation and monitoring applications it is otten useful to have an estimate on the number of animals In an area. Additionally, the ability to Identify which signal belongs to which animal can aid and speed-up both tracking and localisation. This thesis presents an approach for identification of sperm whales uSing acoustic cues wilh the reqUirement that the algOrithm can run on limited resources In real-time. First, an attempt was mada to identify sperm whale groups using the rhythmic structure in cod as A protocol was developed for objective coda classification and to compare results between studies. Due to a lack of data it is not yet clear if the coda rhythm is truly unique for a sperm whale group. To further investigate this a software tool has been developed and made available to the research community . If adopted , this could allow to draw better conclusions concerning the importance of the rhythm in cod as In the future After codas were discarded lor identification, attention was focussed on finding characteristiC information in the time-frequency domain 01 a sperm whale click. The Gabor funclion was first used as a model for a click, bul was found to be too limited in its choice for Characteristic features and too unreliable due to varlabitity in the click Usel!. To generate a large selection of possible features a local discriminant basis was created based on a wavelet packet table. This led to features Ihat outperformed the Gabor function with a simple linear classifier lA study of the 1eature space constructed from the wavelet coefficients suggested the use of a non-linear claS SIfier Ihat can model the features' clusters. USing Gaussian kernels to describe the clusters, a radial basis function network (RBF) was constructed for the classification. Additionally, beSides the use of the RBF network that focussed on the cluster centres, classification with support vector machines (SVM), which focus on the cluster boundaries, was also evaluated It was found that a model describing cluster centres with RBF outperformed SVM. To its advantage. the RBF network required much less Information and computed much faster To initialise the classifiers a Gaussian Mixture Model (GMM) clustering algorithm was evaluated. The task of thiS algorithm is 10 perform an initial separation , uSing the Gaussian distributions that showed good performance with RBF, that allows the classifier 10 be trained. The oplimisation of GMM makes use of an expectation maximisatlon routine. a much more expensive algorithm than that was used for RaF. Therefore, It is intended to only run at the start of a recording to obtain a training set, atter which RBF can take over. Clustering with GMM showed capacity to estimate the number of animals and to prOvide enough Information to train the RBF classifier, but this should be tested on more data sets In the future. The classification approach herein presented allowed accurate classification on the avaitable data set in real-time, and the approach IS considered to be a reliable method that can be applied In autonomous monitoring applications or e.9. from a laptop on a boat.