Visual Analytics to Assist Event Labelling Verification in Soundscape Ecology

• Ano de defesa: 2025
• Autor(a) principal: Rodrigues, David Souza
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Acoustic data labelling; Acoustic features; Analítica visual; Características acústicas; Ecologia de paisagem acústica; Etiquetação de dados acústicos; Information visualisation; Soundscape ecology; Visual analytics; Visualização de informação
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-22072025-143024/

The study of soundscapes has benefited from the development of automated data collection methods, such as Passive Acoustic Monitoring (PAM). However, as collecting data became more manageable, it also resulted in expressive volumes of data to be inspected and processed. An important task consists of annotating the data regarding the presence of relevant acoustic events. Nevertheless, the sheer amount of data makes manually labelling all the data unfeasible. Furthermore, the manual process is prone to error, resulting in inconsistent and mistaken labels. As a result, there have been attempts to investigate supervised machine-learning algorithms for event classification. The quality of labels is thus essential for ecological research and supervised machine learning algorithms, which use them as training sets. This work developed a tool named \'VT-AA\' (Verification Tool for Acoustic Annotations), aimed to support the analysis and interpretation of acoustic labels provided for an acoustic database by assisting with the identification of possible errors and inconsistencies and the observation of inter- and intra-species call variability. The tool uses Visual Analytics techniques, such as Multidimensional projection and Parallel Coordinates plot. In order to validate the developed tool, a labelled database with bird species collected on the ecological Corridor of Cantareira-Mantiqueira was used. Features were extracted from the collected audio files and corresponding labels and used as input for the tool. By identifying different groups of labels, it was possible to identify outliers that may correspond to incorrectly classified labels or labels that, even if correctly classified, are not good representatives of their respective classes. The case studies indicate the tool has promise in contributing to optimise the analysis of large amounts of labelled acoustic data by specialists.