Acoustic Machine Learning for Multi-Species Identification in the Pantanal Wetlands

Abstract

Trinh Quang Minh, Ngo Thi Lan , Nguyen Minh Hieu, Tran Minh Tan, Do Chi Tam, Lam Tan Phuong, Nguyen Chi Cuong and Bui Xuan Tung

Biodiversity monitoring in remote ecosystems, such as Brazil’s Pantanal wetlands, is often constrained by logistical barriers, seasonal fluctuations, and the vast scale of data acquisition. While Passive Acoustic Monitoring (PAM) offers a scalable alternative through continuous multi-species soundscape recording, high environmental noise and sparse labeled data remain significant hurdles for automated identification. This study proposes a robust machine learn-ing framework for multi-label acoustic recognition. By extracting high-dimensional features—including Mel-Frequency Cepstral Coefficients (MFCCs) and spectral centroids—from short audio segments, we developed an ensemble learning pipeline utilizing a Random Forest classi-fier. Our approach integrates both primary and secondary labels to account for background vocalizations, enhancing model resilience in complex acoustic environments. Evaluated on the BirdCLEF+ 2026 dataset using macro-averaged ROC-AUC, our results validate the synergy between advanced feature engineering and ensemble methods. This framework provides a scalable foundation for monitoring understudied taxa and informs conservation strategies in one of the world’s most vulnerable ecological frontiers.

PDF