An Optimised Grid Search Based Framework for Robust Large-Scale Natural Soundscape Classification

Abstract

Large-scale natural soundscapes are remarkably complex and offer invaluable insights into the biodiversity and health of ecosystems. Recent advances have shown promising results in automatically classifying the sounds captured using passive acoustic monitoring. However, the accuracy performance and lack of transferability across diverse environments remains a challenge. To rectify this, we propose a robust and flexible ecoacoustics sound classification grid search-based framework using optimised machine learning algorithms for the analysis of large-scale natural soundscapes. It consists of four steps: pre-processing including the application of spectral subtraction denoising to two distinct datasets extracted from the Australian Acoustic Observatory, feature extraction using Mel Frequency Cepstral Coefficients, feature reduction, and classification using a grid search approach for hyperparameter tuning across classifiers including Support Vector Machine, k-Nearest Neighbour, and Artificial Neural Networks. With 10-fold cross validation, our experimental results revealed that the best models obtained a classification accuracy of 96% and above in both datasets across the four major categories of sound (biophony, geophony, anthrophony, and silence). Furthermore, cross-dataset validation experiments using a pooled dataset highlight that our framework is rigorous and adaptable, despite the high variance in possible sounds at each site.