Few-Shot Learning and Learnable Frontends for Remote Monitoring of Bird Populations

Abstract

In response to changing ecological and environmental factors, automated monitoring of bird populations has become imperative for conservation efforts and as an indicator of change in its own right. This is particularly crucial in remote areas where manual observation is challenging. Bird vocalisations are highly suitable for population monitoring, particularly in environments where visual analysis is impractical. Automating the detection of bird activity would allow ornithologists and conservationists more time for in-depth analysis of the data.

Essential to this effort is the development of high-performing, efficient systems capable of operating on low-resource devices which are suitable for field deployment. However, the limited availability of extensive, temporally detailed, fully annotated datasets for bird audio poses a challenge for generalisation. Few-shot learning, especially in the context of sound event detection, emerges as a promising solution to address the scarcity of data. It offers a potentially lightweight approach to monitoring bird populations through their vocalisations.

Additionally, in the age of deep learning, the predominant input features to systems are time-frequency representations of the audio. Recent advancements in learnable frontends have enabled systems to learn from raw waveforms, learning new filterbanks for time-frequency representations or learnable compression which is applied to existing representations. The incorporation of learnable frontends into deep learning systems enables the learning of improved audio representations directly from the audio data.

This thesis seeks to advance automatic and remote bird population monitoring through the development of activity detection models and the integration of learnable frontends into bird audio analysis. Features and classifiers are explored to determine their suitability for bird audio monitoring, considering both performance and computational efficiency. This includes development of the AMPS feature set, derived from amplitude modulation, pitch and spectral features, which is suitable for use with low-resource classifiers. Additionally, a few-shot learning system for bioacoustic activity detection is developed as a system that can generalise from few labelled examples and is potentially suitable for field deployment. We incorporate learnable frontends into this system, yielding a relative 25% increase in F1-score over static time-frequency representations.

Furthermore, this thesis benchmarks traditional fixed-parameter frontends against a new generation of learnable frontends when applied to bird audio. We observe that Per-Channel Energy Normalisation is the best overall performer and that in general learnable frontends significantly outperform traditional methods. While the integration of learnable frontends enhances overall performance, those employing learnable filterbanks exhibit sensitivity to initialisation. We characterise the sensitivity of a learnable filterbank to its initialisation using several strategies on two audio tasks: voice activity detection and bird species identification. The limited movement of the filters from their initialisation suggests that alternative optimisation strategies may allow a learnable filterbank to reach better overall performance. To address this, we propose two mitigation strategies which modify the training strategy to encourage filter movement. While yielding inconclusive results, these attempts serve as a preliminary step for future research on the filterbank initialisation problem.