Home // IARIA Congress 2022, The 2022 IARIA Annual Congress on Frontiers in Science, Technology, Services, and Applications // View article
Authors:
Nestor Rendon
Susana RodrÃguez-Buritica
Claudia Isaza
Keywords: Machine learning, Ecoacoustics, Soundscape, Clustering
Abstract:
Passive Acoustic Monitoring (PAM) is one of the alternatives to monitoring endangered ecosystems. PAM uses acoustic recordings of monitored sites to understand the dynamics of communities, and landscape transformation, among other ecological indicators. PAM studies of landscape transformation have applied machine learning techniques using discrete labels for transformation states (i.e., high, medium, low). However, a site does not necessarily belong to a discrete label but can be between two transformation states. Thus, discretely labeling a degraded site while ignoring intermediate states is biased. Due to the natural variability of soundscape, multiple groups that describe different patterns are a requirement for clustering recordings that can belong to specific transformations. This paper proposes an unsupervised methodology based on clustering to identify the ecological transformation. Our proposal does not use transformation labels, either selecting the variables or training the models. This allows to find sites with intermediate states and associate different clusters to a specific level of ecological transformation. Similar groups of recordings were found and linked with ecological transformation using Gaussian Mixture Models (GMMs) in three periods of the day: morning (5-8), day (8-17), and night (17-5). We evaluated 13 Clustering Internal Validation Indices (CIVI) to know which one establishes the number of clusters associated with ecological transformation. Acoustic Indices (AIs) operated as variables to provide information on the acoustic complexity of the sites. We use the Dependence Guided Unsupervised Feature Selection (DGUFS) method to select the most relevant AIs. With data collected from 2015 to 2017, we tested the proposal in a Tropical Dry Forest ecosystem located in the Bolivar region of northern Colombia. Results showed that it is possible to identify the ecological transformation with an F1 score of 0.86 using the Scattering Distance (SD) index as CIVI. In the paper, we evidenced that it is possible to identify the ecological transformation not limited to known a-priori discrete labels.
Pages: 32 to 38
Copyright: Copyright (c) IARIA, 2022
Publication date: July 24, 2022
Published in: conference
ISBN: 978-1-68558-017-9
Location: Nice, France
Dates: from July 24, 2022 to July 28, 2022