Manatee Vocalization Detection Method Based on the Autoregressive Model and Neural Networks

• Published in: 2021 IEEE Latin-American Conference on Communications (LATINCOM) pp. 1 - 6
• Main Authors: Rios, Edwin, Merchan, Fernando, Higuero, Rosa, Poveda, Hector, Sanchez-Galan, Javier E., Ferre, Guillaume, Guzman, Hector M.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 17-11-2021
• Subjects: Autoregressive model; backpropagation; bioacoustic classification; Conferences; Estimation; Feature extraction; feedforward; Feedforward neural networks; manatee; neural networks; Sociology; Supervised learning; Training
• DOI: 10.1109/LATINCOM53176.2021.9647815

This work presents a scheme for the detection of manatee vocalizations in underwater recordings to support efforts in monitoring and population estimation of this species in western Panama. The proposed automatic detection scheme uses the autoregressive model as a feature extraction stage to feed two-layer feedforward neural networks that classify the signal as vocalizations or background noise. The neural network was trained with the scaled conjugate gradient backpropagation algorithm using supervised learning. The proposed scheme provides an accuracy of 92.4% on the training set for both classes.