JMIR Publications

ABSTRACT

Background:

Bioacoustics classification plays a crucial role in wildlife monitoring, ecological assessment, and health diagnostics. However, the presence of environmental noise, signal variability, and limited annotated datasets often hinders model reliability and deployment. Feature extraction and denoising techniques have become critical for improving model robustness, enabling more accurate interpretation of acoustic events across diverse bioacoustics domains.

Objective:

This review aims to systematically examine advancements in noise-resilient feature extraction and denoising techniques used in bioacoustics classification models. Specifically, it explores methodological trends, model types, real-world deployment, and application areas across ecological and health-related domains.

Methods:

A systematic review was conducted by searching eight electronic databases, yielding a total of 5,462 records. Studies were screened for inclusion if they entailed audio-based classification models, applied experimental or computational methods, and reported empirical performance. A total of 132 studies that fit the eligibility criteria were selected for full review by two independent reviewers. Risk of bias was assessed using a customized tool, with 87.9% (n = 116) of studies rated as low risk, 7.6% (n =10) as moderate risk, and 4.5% (n = 6) as high risk. Reporting quality was evaluated using the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) checklist.

Results:

Out of the 132 included studies, the majority 84.8% (n=112) focused on developing novel classification models, with deep learning and hybrid approaches being the most dominant. Feature extraction played a critical role, with 96.2% (n=127) studies clearly demonstrating feature extraction. MFCCs, spectrograms, and filter bank-based representations were the most common feature representations. Nearly half 47% (n=62) of the studies incorporated noise-resilient methods, such as adaptive deep models, wavelet transforms, and spectral filtering. However, only 14.4% (n=19) demonstrated real-world deployment across healthcare, biodiversity monitoring, and environmental surveillance.

Conclusions:

The integration of advanced deep learning architectures, robust feature engineering techniques and denoising techniques has significantly improved classification accuracy in bioacoustics. Challenges are however present in real-world deployment and proper utilization of denoising strategies in various datasets. Future direction in bioacoustics should focus on deploying noise resilient models into real-world cross domain generalization modules.