Interpretable and Robust Machine Learning for Exploring and Classifying Soundscape Data

Abstract

The adoption of machine learning in Passive Acoustic Monitoring (PAM) has improved prediction accuracy for tasks like species-specific call detection and habitat quality estimation. However, these models often lack interpretability, and PAM generates vast amounts of non-informative data, as soundscapes are typically information sparse. Here, we developed ecologically interpretable methods that accurately predict land use from audio while filtering unwanted data. Audio from habitats in Southern India (evergreen forests, deciduous forests, scrublands, grasslands) was collected and categorised by land use (reference, disturbed, and agriculture). We used Gaussian Mixture Models (GMMs) on top of a Convolutional Neural Network (CNN)-based feature extractor to predict land use. Thresholding based on likelihood values from GMMs improved model accuracy by excluding uninformative data, enabling our method to outperform models such as Random Forests and Support Vector Machines. By analysing areas of acoustic feature space driving predictions, we identified "keystone" soundscape elements for each land use, including both biotic and anthropogenic sources. Our approach provides a novel method for ecologically meaningful interpretation and exploration of large acoustic datasets independent of specific feature extractors. Our study paves the way for soundscape monitoring to deliver robust and trustworthy habitat assessments on scales that would not otherwise be possible.