Applying machine learning methods to detect vocal sexual dimorphism in antbirds

• Grant number: 23/09512-6
• Support Opportunities: Scholarships abroad - Research Internship - Scientific Initiation
• Effective date (Start): December 20, 2023
• Effective date (End): April 19, 2024
• Field of knowledge: Biological Sciences - Zoology
• Principal Investigator: Reginaldo José Donatelli
• Grantee: Enrico Lopes Breviglieri
• Supervisor: Larissa Sayuri Moreira Sugai
• Host Institution: Faculdade de Ciências (FC). Universidade Estadual Paulista (UNESP). Campus de Bauru. Bauru , SP, Brazil
• Research place: Cornell University, United States
• Associated to the scholarship: 22/04384-7 - Vocal sexual dimorphism in three Antbird (Aves, Passeriformes) species in a Cerrado fragment in southeastern Brazil, BP.IC

Abstract

Communication is a crucial aspect of animal behavior, as social interactions rely on information exchange. Vocalizations play a significant role in conveying messages, and many studies have documented a wide range of vocal signals in different taxonomic levels. Vocal sexual dimorphism can help determine population balance, besides other essential factors for biodiversity conservation. Recently, new methods for automated detection and analysis of species vocalizations have been developed, in order to reduce time spent on this task. Therefore, this study aims to use supervised and unsupervised machine learning models to detect sexual dimorphism in the vocalizations of four species of Thamnophilidae in the Cerrado of the state of São Paulo. Birds of the family Thamnophilidae have a species-specific song, which is emitted by both sexes and may be involved in both territorial defense and social behavior through duets produced by the partners. Four species of the family whose members have external sexual dimorphism were selected: Herpsilochmus atricapillus, Herpsilochmus longirostris, Thamnophilus doliatus, and Thamnophilus pelzelni. In the first stage of the project, we found evidence that there is vocal sexual dimorphism in these species. With the data previously collected, the vocalizations will be characterized by multiple feature extractions, and then supervised and unsupervised analysis methods will be used, the first of which will help us understand how to apply automated analysis in new recordings, and the latter, which will allow a better understanding of what are the differences between the songs of males and females. As supervised methods, hidden Markov models, models developed using algorithms such as Koogu and BirdNET, and the support vector machine will be used. As unsupervised methods, Affinity Propagation Clustering, K-medoids, and Gaussian mixture model-based clustering will be used. At last, we will make the validation of the models through recall, precision and the Area Under the Curve (AUC) method. From this study, we expect to establish models for automated detection and analysis of vocal sexual dimorphism, contributing to biodiversity monitoring and conservation. (AU)