Unsupervised identification of Greater Caribbean manatees using Scattering Wavelet Transform and Hierarchical Density Clustering from underwater bioacoustics recordings

Introduction This work presents an unsupervised learning-based methodology to identify and count unique manatees using underwater vocalization recordings. Methods The proposed approach uses Scattering Wavelet Transform (SWT) to represent individual manatee vocalizations. A Manifold Learning approach, known as PacMAP, is employed for dimensionality reduction. A density-based algorithm, known as Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN), is used to count and identify clusters of individual manatee vocalizations. The proposed methodology is compared with a previous method developed by our group, based on classical clustering methods (K-Means and Hierarchical clustering) using Short-Time Fourier Transform (STFT)-based spectrograms for representing vocalizations. The performance of both approaches is contrasted by using a novel vocalization data set consisting of 23 temporally captured Greater Caribbean manatees from San San River, Bocas del Toro, in western Panama as input. Results The proposed methodology reaches a mean percentage of error of the number of individuals (i.e., number of clusters) estimation of 14.05% and success of correctly grouping a manatee in a cluster of 83.75%. Discussion Thus having a better performances than our previous analysis methodology, for the same data set. The value of this work lies in providing a way to estimate the manatee population while only relying on underwater bioacoustics.