Detection and identification of manatee individual vocalizations in Panamanian wetlands using spectrogram clustering

This work presents a methodology to automatically detect and identify manatee vocalizations in continuous passive acoustic underwater recordings. Given that vocalizations of each manatee present a slightly different frequency content, it is possible to identify individuals using a non-invasive acoustic approach. The recordings are processed in four stages, including detection, denoising, classification, and manatee counting and identification by vocalization clustering. The main contribution of this work is considering the vocalization spectrogram as an image (i.e., two-dimensional pattern) and representing it in terms of principal component analysis coefficients that feed a clustering approach. A performance study is carried out for each stage of the scheme. The methodology is tested to analyze three years of recordings from two wetlands in Panama to support ongoing efforts to estimate the manatee population.