Automatic acoustic heterogeneity identification in transformed landscapes from Colombian tropical dry forests

Abstract

Tropical ecosystems with high levels of endemism are under threat due to climate change and deforestation. The conservation actions are urgent and must rely on a clear understanding of landscape heterogeneity from transformed landscapes. Currently, passive acoustic monitoring uses the soundscape to understand the dynamics of biological communities and physical components of the sites and thus complement the information about the structures of landscape. However, the link between the analysis and quantification of ecosystem transformation based on acoustic methods and acoustic heterogeneity is just beginning to be analyzed. This document proposes a new beta Acoustic Heterogeneity Index (AHI) that quantifies the acoustic heterogeneity related to landscape transformation. AHI estimates the acoustic dissimilarity between sites modeling membership degrees of mixture models in three transformation states: high, medium, and low. We hypothesized that if acoustic recordings of different habitats are analyzed looking for particular patterns, it is possible to quantify the landscape heterogeneity between sites using sound. To calculate the AHI we propose a methodology of five steps: (1) filtering out recordings with high noise levels, (2) estimating acoustics indices, (3) including temporal patterns, (4) using GMM classification models to recognize habitat transformation levels, and (5) calculating the proposed AHI. We tested the proposal with data collected from 2015 to 2017 for 22 tropical dry forests (TDF) sites in two watersheds of Colombian Caribbean region. The sites were labeled by the level of landscape transformation using forest degradation indicators with satellite imagery. We compared these labels with the predicted transformation of our method showing an F1 score of 92% and 90% in regions of La Guajira and Bolívar respectively. To use AHI interactively, we analized the soundscapes similarities on geographic maps in the study regions. We identified that AHI allows estimating the similarity of points with similar transformations, and where the soundscape provides information about the transition states. This proposal allows complementing landscape transformation studies with information on the acoustic heterogeneity between pairs of sites.

Keywords

• Acoustic heterogeneity
• Landscape transformation
• Machine learning
• Ecoacoustics
• Acoustic indices