Acoustic Signatures of the Cerrado: Machine Learning Reveals Unique Soundscapes Across Diverse Phytogeographies

Abstract

This article explores the application of machine learning techniques in acoustic ecology to classify the formations of the Brazilian Cerrado (Forest, Savanna, and Grassland) based on their soundscapes. Considering the Cerrados importance for biodiversity and hydrology, as well as the challenges faced by the biome in the face of agricultural expansion, the study seeks more efficient and economical methods for identifying its physiognomies. Five statistical models were developed and evaluated, using both traditional Machine Learning and Deep Learning, with the use of Mel-Frequency Cepstral Coefficients (MFCCs) and spectrogram images as input variables. The performance of these models was measured by accuracy, precision, and recall metrics, revealing a superiority of the Convolutional Neural Network (CNN), which, despite requiring greater computational cost and training time, provided high precision in the classifications and valuable insights through the application of the LIME explainability technique. Moreover, the study proposes a majority vote classification methodology for frequently observed events, enabling reliable classifications through models with moderate performance. It is concluded that the choice of the ideal model for the classification of soundscapes of the Cerrado should consider a balance between accuracy, operational complexity, and efficiency. The conclusions of this study offer relevant directions for future research and the application of monitoring technologies in conservation and recovery efforts of biomes.