Visual analytics of environmental data collected with passive acoustic monitoring

• Grant number: 21/08322-3
• Support Opportunities: Regular Research Grants
• Duration: March 01, 2022 - February 29, 2024
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Maria Cristina Ferreira de Oliveira
• Grantee: Maria Cristina Ferreira de Oliveira
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• Associated researchers: Milton Cezar Ribeiro ; Moacir Antonelli Ponti ; Rosane Minghim
• Associated scholarship(s): 22/10391-6 - Visualization and representation learning strategies in ecoacoustics considering semi-supervised and unsupervised scenarios, BP.TT

Abstract

Passive Acoustic Monitoring (PAM) has been recognized as a promising approach to biodiversity tracking by recording animal abundance and distribution, offering a non-invasive, standardized and efficient approach to collecting ecological data at large spatial and temporal scales. PAM provides unprecedented access to information about changes in the environment and in ecosystems, and can support data-intensive monitoring and conservation programs. The technology is particularly suitable for studying tropical landscapes, characterized by extreme biodiversity and visibility impaired by dense vegetation. However, its expansion is challenged by a classic big data scenario: whilst it allows collecting, with little effort, thousands of large audio files, data analysis by manual inspection quickly becomes unfeasible. The objective of this project is to develop robust computational solutions to meet some demands placed by scholars in ecology and biodiversity, related to the analysis of acoustic recordings collected through PAM. The processing of audio by machine learning techniques is inherently complex, data volumes are expressive, and the systematic processing of natural soundscapes is still relatively recent and a very challenging research topic. As recorders capture ambient sound from multiple sources, regardless of their nature, it is common for acoustic events of interest to be masked, captured faintly, or occur overlapped with other events that may or may not be of interest. As the spatial location of insects and animals varies, while the recorders remain in a fixed position, the acoustic recordings collected are naturally noisy, and occurrences of similar events show great variability. This scenario makes it extremely difficult to perform essential tasks of identifying and tagging occurrences of interest. Successful machine learning strategies in acoustic recordings obtained under more controlled conditions are not particularly effective in this scenario, which motivates the investigation proposed in this project. We will initially consider some critical tasks that are closely related, namely, the automation of the task of tagging acoustic recordings, and the development of machine learning models for the identification, retrieval and classification of events of interest in natural soundscapes. (AU)