Compensating class imbalance for acoustic chimpanzee detection with convolutional recurrent neural networks

TL;DR

This study improves automatic acoustic chimpanzee call detection by addressing class imbalance using spectrogram denoising, loss functions, and resampling within a CRNN framework, achieving significantly better detection rates.

Contribution

It introduces a comprehensive pipeline for compensating class imbalance in deep learning-based acoustic detection, with novel application of denoising and resampling techniques.

Findings

Spectrogram denoising significantly improved detection performance.

Binary cross entropy loss yielded the best results.

Resampling effects varied, sometimes decreasing performance.

Abstract

Automatic detection systems are important in passive acoustic monitoring (PAM) systems, as these record large amounts of audio data which are infeasible for humans to evaluate manually. In this paper we evaluated methods for compensating class imbalance for deep-learning based automatic detection of acoustic chimpanzee calls. The prevalence of chimpanzee calls in natural habitats is very rare, i.e. databases feature a heavy imbalance between background and target calls. Such imbalances can have negative effects on classifier performances. We employed a state-of-the-art detection approach based on convolutional recurrent neural networks (CRNNs). We extended the detection pipeline through various stages for compensating class imbalance. These included (1) spectrogram denoising, (2) alternative loss functions, and (3) resampling. Our key findings are: (1) spectrogram denoising operations significantly improved performance for both target classes, (2) standard binary cross entropy reached the highest performance, and (3) manipulating relative class imbalance through resampling either decreased or maintained performance depending on the target class. Finally, we reached detection performances of 33% for drumming and 5% for vocalization, which is a >7 fold increase compared to previously published results. We conclude that supporting the network to learn decoupling noise conditions from foreground classes is of primary importance for increasing performance.