The importance of acoustic background modelling in CNN-based detection of the neotropical White-lored Spinetail (Aves, Passeriformes, Furnaridae)

Machine learning tools are widely used in support of bioacoustics studies, and there are numerous publications on the applicability of convolutional neural networks (CNNs) to the automated presence-absence detection of species. However, the relation between the merit of acoustic background modelling...

Full description

Saved in:
Bibliographic Details
Published in:Bioacoustics (Berkhamsted) 2024-03, Vol.33 (2), p.103-121
Main Authors: Ventura, Thiago M., Ganchev, Todor D., Pérez-Granados, Cristian, de Oliveira, Allan G., de S. G. Pedroso, Gabriel, Marques, Marinez I., Schuchmann, Karl-L.
Format: Article
Language:English
Subjects:
Acoustic activity detection
bird sound recognition
computational bioacoustics
convolutional neural networks
Pantanal
transfer learning
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Machine learning tools are widely used in support of bioacoustics studies, and there are numerous publications on the applicability of convolutional neural networks (CNNs) to the automated presence-absence detection of species. However, the relation between the merit of acoustic background modelling and the recognition performance needs to be better understood. In this study, we investigated the influence of acoustic background substance on the performance of the acoustic detector of the White-lored Spinetail (Synallaxis albilora). Two detector designs were evaluated: the 152-layer ResNet with transfer learning and a purposely created CNN. We experimented with acoustic background representations trained with season-specific (dry, wet, and all-season) data and without explicit modelling to evaluate its influence on the detection performance. The detector permits monitoring of the diel behaviour and breeding time of White-lored Spinetail solely based on the changes in the vocal activity patterns. We report an advantageous performance when background modelling is used, precisely when trained with all-season data. The highest classification accuracy (84.5%) was observed for the purposely created CNN model. Our findings contribute to an improved understanding of the importance of acoustic background modelling, which is essential for increasing the performance of CNN-based species detectors.
ISSN:0952-4622
2165-0586
DOI:10.1080/09524622.2024.2309362